Till startsida
University of Gothenburg
To content Read more about how we use cookies on gu.se

Past seminars



Upcoming seminars are visible in the calendar.


April 29: Gemma Boleda "Distributional Semantics and Linguistic Theory"

Gemma Boleda's presentation slides
Distributional semantics provides multi-dimensional, graded, empirically induced word representations that successfully capture many aspects of meaning in natural languages, as shown in a large body of work in computational linguistics; yet, its impact in theoretical linguistics has so far been limited. In this talk, I will present methods and results in distributional semantics that are of relevance for theoretical linguistics, in three areas: semantic change, polysemy and composition, and the grammar-semantics interface (specifically, the interface of semantics with syntax, on the one hand, and with derivational morphology, on the other). Talk based on the following paper: Boleda, G. 2020. Distributional Semantics and Linguistic Theory. Annual Review of Linguistics, Vol. 6: 213-23. [Pre-print version]

April 22: Herbert Lange "Learning Domain-Specific Grammars from Example Sentences"

Herbert Lange's presentation slides
For domain-specific applications computational grammars can be a useful resources. One challenge is that the domain experts and the grammar engineers usually are two separate parties. To bridge between the two, we present a method to learn a domain-specific grammar from a wide-coverage grammar using natural language example sentences.

We model the learning process as a constraint optimization problem and show that we can learn subgrammars from positive examples. Furthermore we show how negative examples can be included to allow for an iterative learning process and how the quality of the grammar can be improved by merging grammar rules.

April 15: Asad Sayeed and Yuval Marton "Data requirements for thematic fit modeling"

Asad Sayeed and Yuval Marton's presentation slides
Thematic fit is the extent to which a given noun fits a given semantic role that is associated with a given verb. For example, "knife" fits better as an instrument for the verb "cut" than does "towel". Previous work (Hong, Sayeed, and Demberg, 2018; Sayeed, Shkadzko, and Demberg,2018) modeled human thematic fit judgements through a neural network approach that involved a very large automatically-annotated corpus.
Received wisdom in the machine learning era is that more data with poorer-quality annotation is cheaper and more effective than a smaller amount of higher-quality annotation (Petrov et al., 2010).

Calling into question the wide applicability of this received wisdom, we replace some automatic annotation layers in the Sayeed et al. dataset with annotations derived from a better SRL tool, parser, and morphological analyzer. We obtained higher correlations with human-collected thematic fit judgements by training our models on dramatically less data with higher annotation quality. We therefore call for more nuanced judgment in automatic annotation design decisions in the NLP community.

Joint work between Yuval Marton (University of Washington; Bloomberg LP) and Asad Sayeed (University of Gothenburg).

April 1: Chatrine Qwaider "Deep Learning for Arabic Computational Linguistics (Sentiment Analysis as case study)"

Chatrine Qwaider's presentation slides
The Seminar is part of Reading course on applying deep learning for Arabic Computational Linguistics. I present an overview of relevant literature on employing deep learning architectures for different Arabic NLP tasks like Language Identification , Sentiment Analysis, Entity Recognition and so on. However, Dialectal Arabic faces a number of challenges when it comes to NLP, resulting in weak performance systems and models. One of the reasons for this is that we try to first build models for Modern Standard Arabic and later use the models to predict the Dialectal Arabic, something which does not work well. In this talk, I pick Sentiment Analysis as a case study in order to show the power of deep learning for Dialectal Arabic. A mixed LSTM and CNN network is presented for SA, giving reasonable results. We also experiment with fine tuning the pertained language model BERT in order to build a classification model for Dialectal SA on it. Lastly, we compare both the traditional DL and BERT results.

March 11: Adam Ek "Enhanced dependency parsing"

Adam Ek's presentation slides
In this talk I introduce the framework “Enhanced Universal Dependencies”, an add-on to standard universal dependency annotations focusing on highlighting semantic relationships between words in a sentence. In the talk on-going work is presented, aimed at parsing enhanced universal dependencies in 16 languages using an encoder-decoder framework with attention.

March 4: CLASP/Linguistics seminar: Elin McCready "Dogwhistles: Ideology and Trust"

Elin McCready's presentation slides
In political speech, it is often strategically important to signal one's ideology to a subset of listeners, especially when that ideology may be controversial. The term "dogwhistle" refers to a kind of coded message sent which sends one message to all listeners and an additional message to a class of `savvy' interpreters; this kind of messaging is prevalent in political discourse. This talk describes an approach to dogwhistles which takes them to send coded messages in a way dependent on recognition of the speaker's political ideology. After laying out some criteria for an account of dogwhistles, a game-theoretic account is proposed and then extended to a general notion of communicative trust.

February 26: Vladislav Maraev and Bill Noble "The effect of laughter on dialogue act recognition"

Vladislav Maraev and Bill Noble's presentation slides
We investigate how useful BERT is for dialogue act recognition. We analyse benefit of BERT's pre-training procedure and the importance of fine-tuning in the dialogue setting. To confirm that the model learns to represent dialogical features, we look at how it uses laughter, a phenomenon specific to dialogue, and analyse where laughter is most helpful for dialogue act recognition.

February 21: Mehdi Ghanimifard - Final seminar "Why the pond is not outside the frog? Grounding in contextual representations by neural language models"

Mehdi Ghanimifard's presentation slides
In this thesis, to build a multi-modal system for language generation and understanding, we study grounded neural language models. Literature in psychology informs us that spatial cognition involves different aspects of knowledge that include visual perception and human interaction with the world. This makes spatial descriptions a compelling case for the study of how spatial language is grounded in different kinds of knowledge. In six studies, we investigate what and how neural language models (NLM) encode spatial knowledge.

In the first study, we ask if the language model has a systematic generalisation to learn the grounding on the unseen composition of representations. Then in the second study, we show the potentials in using uni-modal knowledge for detecting metaphors in adjective-nouns compositions. In the third study, we explore the traces of functional-geometric distinction of spatial relations in uni-modal NLM. This distinction is essential since the knowledge about object-specific relations are not grounded in the visible situation. Following that, in the fourth study, we inspect representations of spatial relations in a uni-modal NLM to understand how they capture the concept of space from the corpus. The predictability of grounding spatial relations from contextual embeddings is vital for the evaluation of grounding in multi-modal language models. In the fifth study, we try to evaluate the degree of grounding in language and vision with adaptive attentions. In the sixth study, we use adaptive attention to understand if and how additional bounding box geometric information could improve the generation of relational image descriptions.

The primary argument of the thesis is that spatial expressions in natural language are not always grounded in direct interpretations of the locations. In a joint model of vision and language, the neural language model provides spatial knowledge that is contextualising the knowledge from visual repre-sentations about locations. The knowledge in the language model comes from locative expressions in the dataset used for the training task and is also shaped by the aspects of the model design.

February 20: CLASP/CLT seminar: Desmond Elliott "Compositional Generalization in Image Captioning"

Desmond Elliott's presentation slides
Image captioning models are usually evaluated on their ability to describe a held-out set of images, not on their ability to generalize to unseen concepts. We study the problem of compositional generalization, which measures how well a model composes unseen combinations of concepts when describing images. State-of-the-art image captioning models show poor generalization performance on this task. We propose a multi-task model to address the poor performance, that combines caption generation and image--sentence ranking, and uses a decoding mechanism that re-ranks the captions according their similarity to the image. This model is substantially better at generalizing to unseen combinations of concepts compared to state-of-the-art captioning models.

February 12: Lasha Abzianidze "Natural theorem proving for natural language: theory and application" 

Lasha Abzianidze's presentation slides
If we assume Montague's belief that there is ¿no important theoretical difference between natural languages and the artificial languages of logicians¿, then there should exist a proof system for natural languages too, like it is for various logics. But is such a natural proof system sensible? In my talk, I will present a version of a natural proof system that is specially designed to account for natural language inference (NLI) in a systematic way. The proof system, called the Natural Tableau, is based on a semantic tableau method and operates on terms of higher-order logic, which represent a more natural way of modelling linguistic semantics. A Natural Tableau-based theorem prover is able to model both shallow and logical reasoning, demonstrated on standard NLI benchmarks. The prover can also tackle the problem of knowledge sparsity with supervised knowledge induction.

February 10: Lasha Abzianidze "The Parallel Meaning Bank: a corpus of translations annotated with formal meaning representations"

Lasha Abzianidze's presentation slides
(joint work with Johan Bos, Kilian Evang, Hessel Haagsma, and Rik van Noord)
The Parallel Meaning Bank (PMB) is a large collection of bitexts annotated with Discourse Representation Structures. Two main principles underlying the PMB are (i) a meaning-preserving property of translations and (ii) compositionality of meaning. The meaning-preserving principle drives formal meaning representations of translations to the same representation. At the same time, the compositionality principle reduces modelling of phrase semantics to modelling of lexical semantics. While the modelling of lexical semantics is still a challenge, the shift to the lexical level enables us to project formal meaning representations from a pivot language to other languages. This is a very attractive idea, but it faces enormous challenges. How did we deal with them? In the presentation, I will describe the annotation layers of the PMB, a pipeline of NLP tools, and its online annotation environment.

January 29: Shiri Lev-Ari "Language evolution and change from a social networks perspective"

Shiri Lev-Ari's presentation slides
Languages exhibit great variability in their structures. In this talk I will show that some of the cross-linguistic differences could be accounted for by languages¿ adaptation to their social environment. I will demonstrate how properties of the community structure, such as its size and interconnectivity, influence how information travels, and consequently the structure of the grammar and vocabulary that the community develops. I will further show how network structure dynamics interact with cognitive biases in a manner that affects linguistic stability and the likely agents of change.


December 4: Shalom Lappin "Modelling the Effect of Context on Sentence Acceptability"

Shalom Lappin's presentation slides
Joint work with Jey Han Lau, The University of Melbourne; Carlos Armendariz, Queen Mary University of London; Matthew Purver, Queen Mary University of London; and Chang Shu,University of Nottingham Ningbo China

We study the influence of context on sentence acceptability. First we compare the acceptability ratings of sentences judged in isolation, with a relevant context, and with an irrelevant context. Our results show that context induces a cognitive load for humans, which compresses the distribution of ratings. Moreover, in relevant contexts we observe a discourse coherence effect which uniformly raises acceptability. Next, we test unidirectional and bidirectional language models in their ability to predict acceptability ratings. The bidirectional models show very promising results, with the best model achieving a new state-of-the-art for unsupervised acceptability prediction. The two sets of experiments provide insights into the cognitive aspects of sentence processing, and central issues in the computational modelling of text and discourse.

October 30: Colin Zwanziger "Propositional Attitude Operators via Homotopy Type Theory"

Colin Zwanziger's presentation slides
Since it interprets propositions by sets of possible worlds, the intensional logic of Montague (1973) does not distinguish propositions which are true in the same possible worlds. Because of this, the system does not satisfactorily interpret propositional attitude verbs (believe, know, etc.), a fact which has motivated the development of 'hyperintensional' logics.

I will discuss a hyperintensional system which naturally incorporates the intensional logic of Montague with the usual notions of homotopy type theory (HoTT). This system is based on a fragment of Shulman (2018). From HoTT, we inherit two notions of equality, ¿ and =, which we think of as expressing intensional and extensional equalities, respectively. From Montague, we inherit a syntax for intensional operators, which for us will, however, mean operators which respect intensional but not (necessarily) extensional equality. These are used to interpret propositional attitude operators.

This approach is in the tradition of other linguistically-motivated systems with two notions of equality (Thomason 1980, Fox and Lappin 2008). Some advantages are that it allows a treatment of de re belief, and inherits a nice model theory from HoTT.


October 16: Jean-Philippe Bernardy "A Logic with Measurable Spaces for Natural Language Semantics"

Jean-Philippe Bernardy's presentation slides
Joint work with
Rasmus Blanck
Aleksandre Maskharashvili

The ability of humans to reason under uncertainty has reflections
within natural language where we find various lexico-syntactic
constructions which allow us to express uncertain information.
Moreover, we are able draw conclusions - make inferences under
uncertainty. To give an adequate account to this crucial aspect of natural language, it has been long argued for employing probabilistic tools in defining semantics of natural language. In this abstract we address this issue by proposing a Logic with Measurable Spaces (LMS). We argue that LMS is suitable to represent the semantics of a number of important natural language phenomena. LMS draws inspiration from several sources. It is aims at being decidable (like descriptive logics). It features Sigma spaces (like Martin-Löf type-theory). It internalises the notion of the cardinality (in fact, here, measures) of spaces and ratios
thereof, allowing to capture the notion of event probability.

September 9: Matthew Marge "Towards Natural Dialogue with Robots"

Matthew Marge's presentation slides
Robots can be more effective teammates with people if they can engage in natural language dialogue. In this talk, I will address one fundamental research problem to achieving this goal: understanding how people will talk to robots in collaborative tasks, and how robots could respond in natural language to maintain an effective dialogue that stays on track. The unique contribution of this research is the adoption of a multi-phased approach to building spoken dialogue systems that starts with exploratory data collection of human-robot dialogue with a human ¿wizard¿ standing in for the robot¿s language processing behind the scenes, and ends with training a dialogue system that automates away the wizard.
With the ultimate goal of an autonomous conversational robot in mind, I will focus on the initial experiments that aim to collect computationally tractable human-robot dialogue without sacrificing naturalness. I will show how this approach can efficiently collect dialogue in the navigation domain, and in a form suitable for training a conversational robot. I will also present a novel annotation scheme for dialogue semantics and structure that captures the types of instructions that people gave to the robot, showing that over time these can change as people better assess the robot's capabilities. Finally, I¿ll place this research effort in the broader context of enabling better teaming between people and robots.
This is joint work with colleagues at ARL and at the USC Institute for Creative Technologies.

May 29: Lauri Karttunen "Training a Neural Model to Reason with Implicatives"


Implicative constructions, such as manage to and waste a chance, possess an underlying semantic property that we call the signature of the construction. Implicatives are ubiquitous and the compositionality of their signatures make them an important object of study for Natural Language Understanding. To this end, we introduced the Stanford Corpus of Implicatives (SCI). Drawing inspiration from other Natural Language Inference (NLI) corpora, SCI contains a set of triplets premise, hypothesis, and label. The label indicates the semantic relation between the premise and the hypothesis: entailment, contradiction or neither. The mission of SCI is two-fold: first, to provide a systematic coverage of the large set of implicative constructions in English; and second, to allow for the exploration of a new family of meta-learner models that strive for modular and compositional learning by taking advantage of the existence of the signatures. In particular, we introduced a new meta-learning model, the recursive routing networks (RRN), that efficiently learn to specialize to the fine-grained inferential signatures from the SCI corpus. We review the ability to generalize from seen constructions to similar unseen constructions, with special attention to meta-level properties of the implicatives.

Lauri Karttunen This is joint work with Ignacio Cases, building on our NAACL - 2019 presentation.



May 29: Larry Moss "Monotonicity in Natural Language Inference: An Update on Theory and Practice"


This talk reports on results in the last two years related to monotonicity in NLI. The starting point of this line of work was the suggestion by Johan van Benthem in the 1980's that one could combine the syntactic approach of categorial grammar (CG)  with the semantic idea of monotonicity. Later, I set his ideas on a firmer footing and  adapted them from (plain) CG to Mark Steedman's Combinatory CG (CCG). The move to CCG enables us to try the ideas on datasets of current interest, such as FraCaS and
SICK, and to compare performance with tools coming from machine learning, such as BERT.

The talk will show that one can do a certain amount of automated NLI using parse trees and polarity algorithms. It is hard to precisely say what that 'certain amount' comes to, but we have some quantitative data on the matter. The talk details  comparisons both with other systems that use logic in some form or other,  and also with systems that use deep learning alone.

This is joint work with a number of people, especially Hai Hu.


May 28: Annie Zaenen "WOPIS: Remarks on Word Order, Prosody and Information Structure: the prefield in Swedish (and Dutch)"


The talk is part of a very new project with Elisabet Engdahl and Filippa Lindahl. I will present some preliminary data on what one can find in first position in Swedish declarative main clauses and compare the situation in Swedish with that in Dutch. I discuss how relevant the notions of topic and focus are in the light of these data.


May 9: Stergios Chatzikyriakidis, "To infer or not to infer: Natural Language Inference and Computational Semantics" (docent lecture)



April 17: Thomas Hörberg, "Expectation-based processing of grammatical functions in Swedish"


Much research indicate that language processing is expectation-based, drawing on statistical patterns in the input (e.g., MacDonald 2013). In this talk, I present evidence for this idea from experimental and corpus-based studies on the comprehension and production of grammatical functions (GFs) in Swedish transitive sentences. The preferred word order in such sentences is SVO. However, Swedish also allows for OVS word ordering, with the object placed sentence-initially and the subject post-verbally. Since the NP argument GFs of such sentences may not be correctly determined from the sentence constituent order (i.e., NPs and verb), they are potentially ambiguous. They can therefore be costly to comprehend when the initial NP lacks case marking. In such cases, comprehenders need to revise their initial sentence interpretation as subject-initial upon encountering the disambiguating post-verbal subject NP (Hörberg et al. 2013).

However, corpus-based and typological research shows that GFs correlate with prominence-based (e.g., animacy and definiteness) and verb-semantic (e.g., volitionality) information, both in the frequency distributions in language use within individual languages (e.g., Bouma 2008), and the grammatical encoding of GFs across languages (e.g., Aissen 2003), creating complex statistical regularities in the distribution of prominence-based, morphosyntactic and verb-semantic properties. These properties and their interplay may be utilized during encoding and decoding of GFs in production and comprehension in order to overcome potential ambiguity problems.

I will present results from a corpus study of written Swedish investigating the distribution of these properties in subject-initial, object-initial and passive sentences. I will argue that writers tend to balance their use of these properties in order to avoid GF ambiguities. In particular, writers less frequently use OVS sentences when other morphosyntactic or animacy-based information about GFs are unavailible (Hörberg 2018). In such cases, writers more frequently use the unambiguous passive construction.

I will then present an expectation-based model of processing difficulty during incremental GF assignment in Swedish transitive sentences, based upon the statistical regularities observed in the corpus data (Hörberg 2016). Processing difficulty is quantified as the on-line change in the expectation of a particular GF assignment (subject- or object-initial) upon encountering the properties of a constituent (e.g., NP2) with respect to the previously encountered properties (e.g., NP1 and verb(s)) in terms of Bayesian surprise.

I will finally provide empirical evidence for this expectation-based model on the basis of a self-paced reading experiment, testing some of the most prominent model predictions. Here, by-region reading times converged with the region-specific Bayesian surprise predicted by the model. For example, NP2 reading times in ambiguous OVS sentences were mitigated when NP1 animacy and its interaction with verb class bias towards an object-initial word order.

These findings provide evidence for the expectation-based account in that they indicate that language users are sensitive to statistical regularities in their language during both production and comprehension of GFs. During production, writers seem to balance their use of morphosyntactic and prominence-based cues to GFs in a manner that accommodates
comprehension. During comprehension, incremental GF assignment draws upon statistical regularities in the distribution of morphosyntactic, prominence-based and verb-semantic properties.


April 10: Massimo Poesio, "Disagreements in Anaphoric Interpretation"

The assumption that natural language expressions have a single, discrete and clearly identifiable meaning in a given context, successfully challenged in lexical semantics by the rise of distributional models, nevertheless still underlies much work in computational linguistics, including work based on distributed representations. In this talk, I will first of all present the evidence that convinced us that the assumption that a single interpretation can always be assigned to anaphoric expression is no more than a convenient idealization. I will then discuss recent work on the DALI project that aims to develop a new model of interpretation that abandons this assumption for the case of anaphoric interpretation/coreference. I will present the recently released Phrase Detectives 2.1 corpus, containing around 2 million crowdsourced judgements for more than 100,000 markable, an average of 20 judgements per markable; the Mention Pair Annotation (MPA) Bayesian inference model developed to aggregate these judgements; and the results of a preliminary analysis of disagreements in the corpus suggesting that between 10& % and 30% of marbles in the corpus appear to be genuinely ambiguous.

Joint work with Jon Chamberlain, Silviu Paun, Alexandra Uma, Juntago Yu, Derya Cokal, Janosch Haber, Richard Bartle and Udo Kruschwitz.


March 29: Jey Han Lau, "Early rumour detection"


Rumours can spread quickly through social media, and malicious ones can bring about significant economical and social impact. In this talk I'll present an on-going work on rumour detection; particularly, we are interested in understanding how *early* we can detect them. Although there are numerous studies on rumour detection, few are concerned with the timing of the detection. A successfully-detected malicious rumour can still cause significant damage if it isn't detected in a timely manner, and so timing is crucial. To address this, we present a novel methodology for early rumour detection. Our model treats social media posts (e.g. tweets) as a data stream and integrates reinforcement learning to learn the number minimum number of posts required before we classify an event as a rumour. Experiments on Twitter and Weibo demonstrate that our model identifies rumours earlier than state-of-the-art systems while maintaining a comparable accuracy.


March 27: Jey Han Lau, "Deep-speare: A joint neural model of poetic language, meter and rhyme"


In this talk, I will present a paper on poetry generation that was published in ACL2018. In the paper we propose a joint architecture that captures language, rhyme and meter for sonnet modelling. We assess the quality of generated poems using crowd and expert judgements. We found that the stress and rhyme models perform very well, as generated poems are largely indistinguishable from human-written poems. Expert evaluation, however, reveals that a vanilla language model captures meter implicitly, and that machine-generated poems still underperform in terms of readability and emotion. Our research shows the importance expert evaluation for poetry generation, and that future research should look beyond rhyme/meter and focus on poetic language.


March 20: Oliver Bott, "Incremental Interpretation of Relative Scope?"


This talk is about the incremental construction of the semantic representation. I will first briefly introduce an incremental semantic theory that can deal with incremental construal of scope readings in an event semantic framework. The main part of this talk, however, will be about processing evidence for or rather against such a fully incremental theory.

In a running eye-tracking study, we (joint work with Fabian Schlotterbeck) investigate the time course of linear scope construal in sentences with a quantifier and negation. Semantic complexity is manipulated by comparing monotone increasing(UE) and monotone decreasing (DE) quantifiers (cf. Deschamps et al. 2015) in interaction with the presence or absence of sentence negation. An offline pretest confirmed that the sentences were interpreted as intended and a first eye-tracking experiment established clear processing differences between negated and non-negated scope disambiguated sentences with DE vs. UE quantifiers. DE quantifiers incurred overall more processing costs than UE quantifiers, and these processing costs interacted with the presence or absence of negation in the expected direction.

(1-a) Mehr als die Hälfte der Studenten hat (nicht in der Mensa gegessen¿
More than half of the students has (not)in the mensa eaten¿
¿More than half of the students did (not) eat in the mensa¿¿

(1-b) Weniger als die Hälfte der Studenten hat (nicht) in der Mensa gegessen¿
Fewer than half of the students has (not) in the mensa eaten¿
¿Fewer than half of the students did (not) eat in the mensa¿ ¿

A second eye-tracking experiment tested sentences such as (1-a/b) with the main verb occurring only after the negation with sentences with the verb aß (ate) in verb second position before the negation. The verb position was manipulated to investigate whether effects of quantificational complexity could show up even before the verbal predicate was encountered (cf. Bott & Schlotterbeck, 2015 for the same logic). To our surprise, effects of semantic complexity only showed up at the very end of the sentences and during rereading.

A third running eye-tracking experiment tests our materials embedded in larger discourse contexts establishing the Question Under Discussion (QUD): "What proportion of the students did or did not eat in the mensa, respectively?" Tian et al. (2016) proposed that non-incremental effects observed for processing negation may be due to the timing of QUD accommodation. Besides this contextual embedding the sentences were changed into cleft constructions (lit. transl. from German: It were less/more than half of the students, that (not) in the mensa have eaten). First results indicate: Even though the effects occurred earlier than in our previous experiments, complexity effects due to monotonicity still seem to only emerge after having read a complete minimal sentence including the verb. Furthermore, finding qualitatively the same pattern of effects suggests that semantic complexity is clearly at issue in sentences with DE quantifiers and negation beyond effects related to QUD accommodation.

To summerize, the results of the present study on the time course of scope interpretation reveal essentially non-incremental effects. The processing of quantifier scope thus seems to depend on a larger domain than just the scopal operators themselves. Corroborating the conclusion drawn in Bott & Schlotterbeck (2015), quantifiers seem to be interpreted with respect to scope only after having encountered a complete minimal sentence. If time allows, I will contrast the non-incremental processing of scope information with results from experiments showing highly incremental, predictive processing of quantificational restriction. Taken together, our experiments suggest a qualitatively different time course of interpreting the scope and the restrictor argument during online semantic processing.


March 8: Aarne Talman, "Neural Network of NLI Fail to Capture the General Notion of Inference"

Natural language inference (NLI), the task of determining if a sentence is entailed by one or more given sentences, has been a very popular line of research in the NLP community. Due to the popularity and recent advances in neaural network, architectures, significant progress has been made in NLI research, especially with the introduction of various pre-trained contextual language models, like ELMo and BERT. However, there are number of concerns also raised about the current NLI research mostly due to the shortcomings of the current NLI datasets.

In my talk, I will introduce the neural network approaches used in NLI and describe our sentence representation architecture, Hierarchical BiLSTMs (HBMP), which has been successful in many NLI tasks. I will give an overview of some of the criticism and negative results in NLI and show how in our most recent experiments even the pre-trained language models fail to generalize across different NLI datasets.


February 27: Heather Burnett, "A Conceptual Spaces Model of Socially Conditioned Language Change"


Since the mid 1990s, the development of mathematical and computational models of language variation and change, such as (Clark and Roberts, 1993; Niyogi and Berwick, 1997; Yang, 2000; Yang, 2002; Kauhanen and Walkden, 2018) among others, has yielded enormous advances in our understanding of the cognitive processes that underly these phenomena. However, although it has been observed since at least (Labov, 1963) that many (if not most) linguistic changes are socially conditioned, formal models have been almost exclusively focused on the grammatical and/or psychological aspects of change, neglecting its social aspects. On the other hand, many non-mathematically oriented approaches in sociolinguistics and linguistic anthropology (see (Bucholtz and Hall, 2005; Bucholtz and Hall, 2008) for an overview) have stressed the role that social meaning, ideologies and identity construction play in language use, and they have developed articulated theories of how meaning and ideological structure mediate the relation between social change and language change. The goal of this paper is to outline a model which brings together insights from identity-oriented theories of language change and unites them with formal theories of language use and understanding. More specifically, we use (Gärdenfors, 2000; Gärdenfors, 2014)¿s Conceptual Spaces framework to formalize speaker/listener ideological change and use epistemic game theory, particularly signaling games with an iterated best response solution concept, such as the Rational Speech Act model (RSA) (Franke, 2009; Frank and Goodman, 2012; Burnett, 2017) to formalize the link between ideology, linguistic meaning and language use. We then show how this new framework can be used to shed light on the mechanisms underlying socially-motivated change in French grammatical gender.

Lecturer: Heather Burnett (LFF, CNRS Université Paris Diderot) (joint work with Oliver Bonami)


February 13: Kathrein Abu Kwaik, "A Lexical Distance Study of Arabic Dialects"

We conduct a computational cross dialectal lexical distance study to measure the similarities and differences between the Arabic dialects and the MSA. We exploit several methods from Natural Language Processing (NLP) and Information Retrieval (IR) like Vector Space Model (VSM), Latent Semantic Indexing (LSI) and Hellinger Distance (HD), and apply them on different Arabic dialectal corpora. We measure the overlap among all the dialects and compute the frequencies of the most frequent words in every dialect. The results are informative and indicate that Levantine dialects are very similar to each other and furthermore, that Palestinian appears to be the closest to MSA.

February 6: Vlad Maraev, "Predicting laughter relevance spaces in dialogue"

In this talk, we address the task of predicting spaces in interaction where laughter can occur. We introduce the new task of predicting actual laughs in dialogue and address it with various deep learning models, namely recurrent neural network (RNN), convolution neural network (CNN) and combinations of these. We also attempt to evaluate human performance for this task via an Amazon Mechanical Turk (AMT) experiment. The main finding of the present work is that deep learning models outperform untrained humans in this task.


December 19: Riza Batista "Detecting semantic similarity between human-articulated statements and Quranic verses"


As the book of religious teachings of Islam, the Quran is frequently cited on many platforms, e.g., on social media, in web sites, and in public speeches or interviews. As such, the Quran has become susceptible to misinterpretation ---intentional or otherwise---in human-articulated statements citing Quranic verses. This can also be attributed to the fact that Quranic verses have a high level of linguistic sophistication. For example, metaphors are often used in the Quran, especially as a means for persuasion. In our work, we seek to determine how close a human-articulated statement is to its cited Quranic verse. To this end, we are developing an approach based on the measurement of semantic similarity between any given pair of texts: a human-articulated statement and a verse from what is considered as a gold standard English version of the Quran. Specifically, we are automatically detecting metaphors in the statements, with the aim of assessing whether they are correct translations. As a supporting resource for development and evaluation, we are leveraging a manually annotated corpus consisting of 300 pairs of statements and verses.

December 12: Andy Luecking "Turning context into meaning"


Some expressions seem to be more context-sensitive than others, namely indexical and demonstrative ones. Both kinds of expressions have a conspicuous commonality, however: they tend to co-occur or even are replaced by a non-verbal act such as a hand-and-arm gesture. Indicating can be achieved by pointing, demonstrating by an iconic gesture. Taking gestures semantically serious, it is claimed, entails modifications in the linguistic theorising on how context can turn into meaning. Based on experimental results, it is -- contrary to standard Kaplanian claims -- argued that pointing gestures do not give rise to direct reference, but rather serve a descriptive, locative function. This reconsideration of reference has repercussions on the standard two-stage view on deferred reference, which are discussed and tentatively solved. A semantic account to iconic gestures is derived from event metaphysics, leading to "locomotor propositions" consisting of a situation-semantic judgment involving a gesture event and a semantic type. Combining co-speech gestures with plural noun phrases (NPs) also runs into difficulties: NPs standardly modelled as generalised quantifiers do not provide discourse referents (DRs) as requires for multimodal integration. Accordingly and finally, a theory of quantified noun phrases is presented, that provides the required DRs and in this sense is "referentially transparent".

December 5: Zhaohui Luo "Dependent Event Types"


This talk studies how dependent types can be employed for a refined treatment of event types, offering a nice improvement to Davidson's event semantics. We consider dependent event types indexed by thematic roles (DETs) and illustrate how, in the presence of refined event types, subtyping plays an essential role in semantic interpretations.

Two applications of DETs are studied. The first shows that DETs give a natural solution to an incompatibility problem (sometimes called event quantification problem) in combining event semantics with the traditional compositional semantics. The second concerns selectional restriction: it is shown that DETs offer flexible but nice treatments of selectional restriction in the MTT-semantic setting with events.

December 3: Zhaohui Luo "Universes in MTT-semantics"


In type theory, a universe is a type of types. Universes play important roles when modern type theories (MTTs) are employed as foundational languages for linguistic semantics. In this talk, I'll report work on two kinds of universes in the study of MTT-semantics. The first kind may be called linguistic universes which include CN, the universe of common nouns, and LType, the universe employed in the study of coordination. It is shown how they are introduced and used in semantic studies and, in particular, their usefulness is reflected in how they facilitate \Pi-polymorphism in various semantic formalisations.

I shall then study logical universes. In order to formulate MTT-semantics adequately, proof irrelevance needs to be enforced in the underlying type theory. For example, in type theory UTT, this is possible because there is the universe Prop of all logical propositions. However, in Martin-Löf's type theory, this is impossible because types and propositions are identified in MLTT. I propose that the extension of MLTT with h-logic, as developed in the HoTT project, can be used adequately as a foundational language for MTT-semantics, since there is a built-in notion of proof irrelevance in h-logic.

November 28: Devdatt Dubhashi, Mikael Kågebäck and Asad Sayeed "Learning (a language) to Communicate Efficiently"


Although languages vary enormously, there are nevertheless universal tendencies in word meanings, such that similar or identical meanings often appear in unrelated languages. A major question is how to account for such semantic universals and variation of the lexicon in a principled and unified way. An influential approach to this question proposes that word meanings may reflect adaptation to pressure for efficient communication -- this principle holds that languages are under pressure to be simultaneously informative (so as to support effective communication) and simple (so as to minimize cognitive load). We offer computational support for this principle in the domain of color words i.e, how languages partition the semantic space of colours by linguistic terms. Our framework uses reinforcement learning for automated agents to generate partitions that are efficient and consistent with those found in many languages in the World Colour Survey. We argue that our framework provides a flexible and powerful tool to address similar fundamental questions about universals in other domains as well.

November 21: Bartosz Wieckowski "Intuitionistic multi-agent subatomic natural deduction for belief and knowledge"


In this talk, we will consider a natural deduction system which aims at the proof-theoretic analysis of reasoning with complex multi-agent belief (resp. knowledge) constructions (involving, e.g., forms of reciprocating or universal belief, or intentional identity). Making use of a normalization result for the system, we shall propose a proof-theoretic semantics for the intensional operators for intuitionistic belief and knowledge which explains their meaning entirely by appeal to the structure of derivations. Since the system enjoys the subexpression property, a refinement of the subformula property, it is fully analytic. We will also compare this approach to the logic and semantics of belief and knowledge with other intuitionistic approaches.

November 20: Ielka van der Sluis "The PAT project: Annotation and Evaluation of Pictures and Text"


In this talk I will present the PAT project in which we investigate the use, effects and optimisation of documents that contain pictures and text (PAT). While the benefit of including pictures has been established, the design of pictures, text, and picture-text relation(s) has not been researched in a systematic manner. PAT aims to gain an in-depth understanding of their characteristics to augment existing theories on cognitive processing of multimodal presentations. Resulting models will be validated by implementing them in natural language generation algorithms and comparing their output to human-authored presentations.

The PAT project launches a methodical investigation of multimodal instructions (MIs) used in first-aid practices to help people in need. Currently, there are no guidelines for the design of MIs that effectively instruct people to operate an AED, place a victim in a recovery position, remove ticks etc. The huge variations in pictorial and verbal means employed in these instructions demonstrate the urgency to obtain validated guidelines based on empirical evidence collected from readers and users. Investigating multimodality in these MIs allows us to evaluate the effectiveness of combining pictures and text in a practical context focussing on e.g. attention, comprehension, recall, user judgements, and task performance.

The PAT project makes use of a annotated corpus of MIs and a workbench that has been developed for the annotation and retrieval of the MIs. The MIs are first-aid instructions that appear in Het Oranje Kruisboekje and variations of these instructions from other sources, like hospitals, health and safety organisations and the internet.

In the PAT project approaches from Information Design Research and Computational Linguistics employing corpus collection and analysis, (automatic) annotation, experimentation, and natural language generation are combined. The project will deliver theoretical results in terms of empirically validated models for effective MIs. Results of practical value include new annotated multimodal corpora, implemented taggers to automatically annotate potentially effective properties of MIs, algorithms to automatically generate effective text-picture combinations and authoring guidelines to produce good quality instructions.

Project website: https://www.rug.nl/let/pat

Lecturer: Ielka van der Sluis detailed CV: https://www.rug.nl/staff/i.f.van.der.sluis/

November 14: Patrick Blackburn "The clarification potential of instructions: Predicting clarification requests"


The hypothesis motivating this talk is that conversational implicatures are an important source of clarification requests, and in this talk I will do two main things. First, I will motivate the hypothesis in theoretical, practical and empirical terms and formulate it as a concrete Clarification Potential Principle: implicatures may become explicit as fourth-level clarification requests. Second, I will present a framework for generating the clarification potential of an instruction by inferring its conversational implicatures with respect to a particular context. I will discuss the evaluation of the framework, illustrate its performance using a human-human corpus of situated conversations, and argue that much of the inference required can be handled using classical planning.

This talk is based on joint work with Luciana Benotti of Logic, Interaction and Intelligent Systems Group, Universidad Nacional de Córdoba, Argentina.

Many of the main ideas can be found in the paper: Modeling the clarification potential of instructions: Predicting clarification requests and other reactions, by Luciana Benotti and Patrick Blackburn, Computer Speech & Language 45: 536-551 (2017)

October 31: Bill Noble "Measuring linguistic style alignment: Social and psychological perspectives"


In conversation, speakers tend to adapt their speech to be more similar to that of their interlocutor. Such alignment is observed across various linguistic phenomena. In this talk, we will consider linguistic style alignment and some ways to measure it. We will also explore whether lingistic style alignment is sensitive to social factors, such as social network centrality, or if it can be explained by automatic psychological priming alone.

October 24: Marco Baroni "Systematic compositionality in recurrent neural networks (and, if time allows, humans) (joint work with Brenden Lake, João Loula, Adam Liska, Germán Kruszewski, Tal Linzen)"


Recurrent neural networks (RNNs) are remarkably general learning systems that, given appropriate training examples, can handle complex sequential processing tasks, such as those frequently encountered in language and reasoning. However, RNNs are remarkably sample-heavy, typically requiring hundreds of thousands of examples to master tasks that humans can solve after just a few exposures. The first set of experiments I will present shows that modern RNNs, just like their nineties ancestors, have problems with systematic compositionality, that is, the ability to extract general rules from the training data, and combine them to process new examples. As systematic compositionality allows very fast generalization to unseen cases, lack of compositional learning might be one root of RNNs training data thirst. I will next present a study where RNNs must solve an apparently simple task where correct generalization relies on function composition. The results suggest that a large random search in RNN space finds a small portion of models that converged to a (limited) compositional solution. Finally, if time allows, I will present ongoing work in which we study the compositional abilities of human subjects, trying to uncover the priors that subtend their generalization skills.

October 22: Marco Baroni "Tabula nearly rasa: Probing the linguistic knowledge of character-level neural language models trained on unsegmented text (work in collaboration with Michael Hahn)"


As recurrent neural networks (RNNs) have recently reached striking performance levels in a variety of natural language processing tasks, there has been a revival of interest in whether these generic sequence processing devices are effectively capturing linguistic knowledge. Nearly all studies of this sort, however, initialize the RNNs with a vocabulaty of known words, and feed them tokenized input during training. We are instead running an extensive, multi-lingual (English/German/Italian) study of the linguistic knowledge induced by RNNs trained at the character level on input data with whitespace removed. Our networks, thus, face a tougher and more cognitively realistic task, having to discover all the levels of the linguistic hierarchy from scratch. Our current results show that these "near tabula rasa" RNNs are implicitly encoding a surprising amount of phonological, lexical, morphological, syntactic and semantic information, opening the doors to intriguing speculations about the degree of prior knowledge that is necessary for succesful language learning.

October 17: Vlad Maraev "Towards KoS/TTR-based proof-theoretic dialogue management (joint work with: Jonathan Ginzburg (Université Paris Diderot), Staffan Larsson, Ye Tian (Amazon Research), Jean-Philippe Bernardy)"


This paper presents the first attempt to implement a dialogue manager based on the KoS framework for dialogue context and interaction. We utilise our own proof-theoretic implementation of Type Theory with Records (TTR) and implement a basis dialogue that involves mutual greeting. We emphasize the importance of findings in dialogue theory for designing dialogue systems which we illustrate by sketching an account for question-answer relevance.

October 10: Mehdi Ghanimifard "Spatial Knowledge In Neural Language Models"


Understanding and generating spatial descriptions requires, among other things, knowledge about how objects are related geometrically. The wide usage of neural language models in different areas, including in generation of scene descriptions, motivates our study how spatial geometric knowledge is encoded in them. We first examine how spatial descriptions are attended by state of the art model of attention in CNNs. We argue that adaptive attention is good at predicting what the objects are but less good on how they relate geometrically. Then we explore different models of encoding explicit spatial information in an end-to-end scene description model. We summarize with the implications of this work for improving image captioning system.

October 3: Rasmus Blanck "A Compositional Bayesian Semantics for Natural Language"


We propose a compositional Bayesian semantics that interprets declarative sentences in a natural language by assigning them probability conditions. These are conditional probabilities that estimate the likelihood that a competent speaker would endorse an assertion, given certain hypotheses. Our semantics is implemented in a functional programming language. It estimates the marginal probability of a sentence through Markov Chain Monte Carlo (MCMC) sampling of objects in vector space models satisfying specified hypotheses. We apply our semantics to examples with several predicates and generalised quantifiers, including higher-order quantifiers. It captures the vagueness of predication (both gradable and non-gradable), without positing a precise boundary for classifier application. We present a basic account of semantic learning based on our semantic system. We compare our proposal to other current theories of probabilistic semantics, and we show that it offers several important advantages over these accounts.

May 16: Richard Sproat "Induction of Finite-State Covering Grammars for Text Normalization (joint work with Kyle Gorman)"


In this talk I will introduce our work on applying neural methods to the problem of text normalization. Though the performance of the system overall is good, it is prone to what we term "silly errors", where for example, "2mA" is read as "two million liters". We have found that finite-state covering grammars are useful for mitigating such errors, and I will discuss induction of such covering grammars from data. I start with presenting our work on inducing grammars for number names (123 verbalized as one hundred (and) twenty three). This work draws inspiration from the (small) linguistics literature on number names, and our method allows one to train finite-state transducers with small amounts of data (a few hundred examples). I will compare the performance with that of an RNN trained on orders of magnitude more data. I will then report on our ongoing work on inducing grammars for a wider range of text normalization problems.

May 14: Richard Sproat "A computational model of the discovery of writing"


This paper reports on a computational simulation of the evolution of early writing systems from pre-linguistic symbol systems, something for which there is poor evidence in the archaeological record. The simulation starts with a completely concept-based set of symbols, and then spreads those symbols and combinations of these to morphemes of artificially generated languages based on semantic and phonetic similarity.

While the simulation is crude, it is able to account for the observation that the development of writing systems ex nihilo seems to be facilitated in languages that have largely monosyllabic morphemes, or that have abundant ablauting processes. We are also able to model what appears to be two possible lines of development in early writing whereby symbols are associated to the sounds of all morphemes linked to a concept (as seems to have been the case in Sumerian), versus just one morpheme linked to a concept (as seems to have been the case in Chinese). Finally, the model is able to offer an account of the apparent rapid development of writing in Mesopotamia that obviates the need to posit a conscious invention of writing, as proposed by Jean-Jacques Glassner. The proposed model thus opens a new approach to thinking about the emergence of writing and its properties, something that, as noted above, has scant direct archaeological evidence.

The software is released open-source on GitHub.

March 15: Sam Bowman "Sentence Understanding with Neural Networks and Natural Language Inference"


Artificial neural networks now represent the state of the art in most large-scale applied language understanding tasks. This talk presents a few methods and results, organized around the task of recognizing textual entailment, which measure the degree to which these models can or do learn something resembling compositional semantics. I discuss experiments on artificial data and on a hand-built million-example corpus of natural data (SNLI/MultiNLI), and report encouraging results.

Bowman, Samuel R., Christopher Potts, and Christopher D. Manning. "Recursive neural networks can learn logical semantics." arXiv preprint arXiv:1406.1827 (2014).

March 12: Sam Bowman "Two Early Efforts toward Using Deep Learning in Syntax and Semantics"


This talk will present two ongoing projects that aim to lay the groundwork to use results from artificial neural networks research in NLP to inform research on core linguistic questions. The first project (based partially on WIlliams et al. 2017) concerns latent tree learning: efforts to discover the optimal tree structures for use in guiding semantic composition for applied language understanding tasks. The second concerns the evaluation of simple neural network models on the classic linguistic acceptability judgments task. This project (in progress, with Alex Warstadt) builds on Lau, Clark, and Lappin '16, and introduces a new dataset of expert acceptability judgments and a new suite of semi-supervised learning experiments with neural networks.

Williams, Adina, Andrew Drozdov, and Samuel R. Bowman. "Learning to parse from a semantic objective: It works. Is it syntax?." arXiv preprint arXiv:1709.01121 (2017).

Lau, Jey Han, Alexander Clark, and Shalom Lappin. "Grammaticality, acceptability, and probability: a probabilistic view of linguistic knowledge." Cognitive Science 41.5 (2017): 1202-1241.

March 8: Stephan Oepen "Holes in Meaning Construction with Minimal Recursion Semantics" (joint work with Dan Flickinger)


In joint work with Dan Flickinger, we provide a semi-formal review of
the meaning construction process in the English Resource Grammar
(ERG), which derives underspecified logical-form meaning
representations for a broad range of lexical classes and
syntactico-semantic constructions in English. We critically revise
and extend the proposal for an MRS algebra by Copestake et al. (2001;ACL) and speculate about the applicability of ERG-style meaning construction to the more coarse-grained syntactic analyses of enhanced Universal Dependencies.

March 8: Dag Haug "Glue semantics for Universal dependencies" (joint work with Matthew Gotham)


In this talk I explore the use of techniques from Glue semantics for
composing meaning representations based on Universal Dependencies (UD) syntactic structures. The UD tree is rewritten as meaning constructors consisting of a lambda term over DRSs and a linear logic formula that guides the semantic composition. Unlike many other frameworks, Glue semantics does not presuppose a one-to-one mapping from syntax to semantics, which is useful when dealing with highly underspecified syntactic representations of the UD kind.



December 13: Rasmus Blanck "Rough Sets and Degree Modifiers"


Rough sets were introduced by Pawlak in 1982, as a generalisation of classical set theory. A rough set is characterised by its upper and lower approximation, respectively, the objects that might belong to the set, and the objects that surely belong to the set. Although this approach has some similarities with fuzzy set theory, the perceived fuzziness of rough
sets does not come from an underlying fuzzy logic, making rough sets a little less fuzzy than fuzzy sets.

In this talk, I will entertain the possibility that rough sets can be used to model degree modifiers. After an introduction to  rough set theory, I will briefly discuss its relation to fuzzy set theory, and point out some possible advantages of rough sets. Finally, I will reintroduce some fuzziness by generalising to probabilistic rough sets.

November 17th: Maxime Amblard "A Formal Account of Disorders in Dialogues"


This talk will present the project SLAM (Schizophrenia and Language - Analyse and Modelling). Since 2011, we build and analyse a corpus of interviews of patient with schizophrenia, in french.
Schizophrenia is well-known among mental illnesses for the severity of the thought disorders it involves, and for their widespread and spectacular manifestations ranging from deviant social behavior to delusion, not to mention affective and sensory distortions.
The goal of the SLAM project is twofold: (i) to discuss how the concepts of rationality and logicality may apply to conversational contexts in which one of the speakers is a patient with schizophrenia, and (ii) to use logical framework to model specific manifestations, namely disorders in conversational speech.

Our data are taken from transcriptions of real conversations between a psychologist and a patient with schizophrenia. Data collection and selection relied on theoretical hypotheses from psychiatry and psychopathology. Confronted with such a pathological conversation, any "ordinary" speaker intuitively feels that there are some incoherencies or discontinuities. We use a DRT (Kamp and Reyle 1993) like semantics in order to propose an interpretation model for such incongruities.

On our recent works, we focus on the extension of compositional semantics based on TTDL (de Groote 2006), a lambda-calculus with continuations. One of our research project is to develop TTDL for Dialogue, in the same perspective as TTR (Cooper and Ginzburg 2002, Cooper 2004, Cooper and Ranta 2008).
(Another one is the french translation of the Fracas resource, but it is not directly rely to SLAM)

The talk will present the SLAM corpus and project, and then (briefly) sketch the on-going works.

October 25: Aleksandre Maskharashvili "An Abstract Categorial Grammar Approach to the Discourse Modeling"


Various theories have been proposed in order to analyze a discourse in terms of rhetorical (discourse) relations. The main assumption in those theories is that every meaningful piece of a felicitous discourse is related to some piece of that discourse with a rhetorical relation. This gives rise to a notion of a rhetorical (discourse) structure. In order to analyze a discourse, both from the parsing and structural points of view, formal grammars of discourse, D-STAG and D-LTAG, were proposed. They build their discourse grammars on top of sentence-level grammars. Discourse connectives are main lexical means for expressing rhetorical relations. They play a similar role in discourse grammars as words do in sentence-level grammars. A discourse connective may appear inside a clause (a clause-medial position) or in front of a clause (a clause-initial position). The grammars of D-STAG and D-LTAG are capable of modeling cases where discourse connectives occupy only clause-initial positions. To process discourses where a discourse connective appears at a clause-medial position, D-STAG and D-LTAG make use of preprocessing of a discourse, which involves moving connectives from clause-medial positions to clause-initial ones. Afterwards, the grammars of D-STAG and D-LTAG can be employed to parse the discourse and simultaneously construct its rhetorical structure. Thus, D-STAG and D-LTAG, each makes use of a two-step process to analyze a discourse. We develop a single-step, purely grammatical approach for analyzing a discourse. Our framework is Abstract Categorial Grammars (ACGs). Our encoding falls into the class of second-order ACGs, which guarantees that the tasks of discourse parsing and generation are of polynomial complexity. In addition, our encoding puts together the discourse-level and sentence-level grammars within a single grammar. This makes our approach beneficial for reducing problems related to ambiguity that arise in the case of treating the discourse-level and sentence-level grammars separately.

September 13: Sharid Loaiciga "What is it? Disambiguating the different readings of the pronoun 'it'


Machine translation of pronouns is problematic for different reasons. Languages differ in their pronoun systems, creating mismatches in features like gender, number, case, formality, animacy and overtness. Another reason is functional ambiguity. Some pronouns have the same surface form but different functions. In this talk, I will address the problem of predicting one of three functions of the English pronoun 'it': anaphoric (e.g., 'The party ended late. It was fun.'), event reference (e.g., 'He can't speak Finnish. It annoys me.') and pleonastic (e.g., 'It's been raining all day.').

I will present experiments using a maximum entropy classifier (MaxEnt) trained on gold-standard data and self-training experiments with a recurrent neural network classifier (RNN) trained on silver standard data, annotated using the MaxEnt classifier. I will show an analysis demonstrating that these models, rather than one being better than the other, have different strengths. I will also present an example of the integration of source pronoun function into an n-gram language model used for target pronoun translation prediction.

The it-disambiguation task is valuable for MT but also for the field of coreference resolution. Standard coreference resolution systems focus on identifying nominal-referential instances, de facto grouping together and discarding the event and pleonastic categories. Linguistically, however, event instances are also referential. I will finish the talk by brainstorming some ideas about how to integrate my work into this field.

May 29: Judith Holler "On the pragmatics of face-to-face communication: the role of the body in social cognition and social interaction"


Coordination is at the heart of human conversation. In order to interact with one another through talk, we must coordinate at many levels, first and foremost at the level of our mental states, intentions and conversational contributions. In this talk, I will present findings on the pragmatics of multi-modal communication from both production and comprehension studies. In terms of production, I will, firstly, throw light on how co-speech gestures are used in the coordination of meaning to allow interactants to arrive at a shared understanding of the things they talk about, and, secondly, on how gesture and gaze are employed in the coordination of speaking turns in spontaneous conversation, with special reference to the psycholinguistic and cognitive challenges that turn-taking poses. In terms of comprehension, I will focus on communicative intentions and the interplay of ostensive and semantic multi-modal signals in triadic communication contexts. My talk will bring these different findings together to make the argument for richer research paradigms that capture more of the complexities and sociality of face-to-face conversational interaction. Advancing the field of multi-modal communication in this way will allow us to more fully understand the psycholinguistic processes that underlie human language use and language comprehension.


May 10: David Schlangen "Learning and Maintaining a Lexicon for Situated Interaction"

If, when asked to "point at the mug", a physically unimpaired person
seems unable to identify a potential referent that is standing in front
of them, we might hesitate to ascribe knowledge of the meaning of the
word "mug" to them, whatever else they may be able to tell us about mugs
(e.g., "wooden mugs were produced probably from the oldest time, but
most of them have not survived intact.", or "mugs are similar to cups").
And yet computational models of word meaning are good at the latter
(e.g., by simply linking to knowledge repositories like wikipedia, where
the previous sentence about wooden mugs was taken from), and fail at the

In this talk, I will present our recent work at learning a lexicon for
referential interaction, where the referential aspects of word meaning
are modelled through perceptual classifiers taking real images as input.
I show that this representation complements other computational meaning
representations such as those derived from distributional patterns, as
well as decompositional or attribute-based representations. The lexicon
is learned through (observation of) interaction, and is maintained and
defended in interaction.

May 5th: Eve Clark: Language as (Graded) Expertise

Just as in the acquisition of other forms of expertise, learning a first language depends on three essential ingredients: exposure, practice, and feedback. Young children are exposed to the community language; they practice it in interaction with more expert speakers from around the age of 1, and they receive feedback on their practice. Fior example, adults check up on their errors with reformulations in the shape of side-sequences and embedded corrections. Adults also offer feedback on appropriate usage, ratifying the information being added to common ground. Finally, in L1 acquisition, children are learning just what one can and can't do with language, as they learn to understand and produce it. In L2 acquisition, learners typically receive much less exposure in interactive settings, receive less feedback timed to pinpoint specific errors, and have less opportunity for practice in truly interactive settings.

May 4th: Herbert Clark "Performing depictions in everyday discourse"

Depicting is a basic method of communication on a par with describing and pointing (or indicating). The idea is that people use their hands, arms, head, face, eyes, voice, and body, with and without props, to stage physical scenes for others, generally as composite parts of utterances along with describing and pointing. Performing depictions, I will show, is inherently interactive, and people choose depictions to communicate things they could not do with language or pointing.

May 3rd: Eve Clark "Why Interaction Promotes Language Acquisition"


Children acquire language as they interact with adults from infancy onwards. Adults-parents and caretakers-are 'expert speakers' and they guide children's earliest steps in interaction, from gaze, to smiles, to reaching, to attempting to communicate. Although very young children can communicate some things early on by pointing and reaching, or by pushing things away, the nonverbal options here are limited in scope. Language offers a lot more. But to acquire language, children need extensive exposure in interaction. In this talk, I will review some of the evidence for how children manage the complex feat of acquiring the basics of a language and how to use it, by around age four- but just the basics. The process of acquiring a language, for all the things we can learn to do with language, lasts a good deal longer.

May 2nd: Herbert H. Clark "On the rational basis of communication"

Communication is often said to be a rational behavior. As Grice (1975) put it, "Talking [is] a special case or variety of purposive, indeed rational behavior." But what does it mean for a behavior to be rational? I will contrast two notions of rationality as they have been applied to language use, one cooperative (à la Grice) and the other interactive, and argue that both are legitimate. I will show how the interactive model, based on one type of rationality, accounts for a wide range of phenomena that are complementary for those accounted for in a Gricean cooperative model. 

March 22: Eleni Gregoromichelaki "Ad hoc grammatical categorisation in Dynamic Syntax"


The view of NLs as codes mediating a mapping between "expressions" and the world is abandoned to give way to a view where utterances are seen as actions aimed to locally  and incrementally alter the affordances of the context. Such actions employ perceptual stimuli  composed not only of "words" and "syntax" but also elements like visual marks, gestures,  sounds, etc. Any such stimuli can participate in the domain-general processes that constitute the "grammar", whose function is the dynamic categorisation of various perceptual inputs and their integration in the process  of generating the next action steps. Given these assumptions, a challenge that arises is how to account for the reification of such processes as exemplified in apparent metarepresentational practices like quotation, reporting, citation etc. It is argued that even such phenomena can receive adequate and natural explanations through a grammar that allows for
the ad hoc creation of occasion-specific content through reflexive mechanisms.

March 20: Hannes Rieser "A Process Algebra Account of Speech-Gesture Interaction"


The talk is based on extensive corpus work dealing with the interaction of gesture and speech in natural route-description dialogues. The issue discussed is how non-regimented gesture and speech processes can be modelled in a formal system. The main argument in the talk is that this cannot be achieved in structural paradigms currently in use. The proposal is to turn instead to process algebras in the tradition of Milner's pi-calculus. The special algebra discussed in the talk is a newly developed hybrid lambda-psi calculus which can transport typed lambda-expressions over communicating input-output channels. Central for the account is the notion of agent: Agents encode speech information, gesture information or both. They can put information on channels and send it to other channels or take information from others and combine it with the information they have. Speech-gesture interaction is conceptualised via channel interactions of this sort. Interactions are allowed, postponed or blocked using a typing system. Successful communication among agents leads to a multi-modal meaning representation serving as logical form.

Rieser, H. (2014). Gesture and Speech as Autonomous Communicating Processes. Talk at the Stuttgart Workshop on "Embodied meaning goes public". Stuttgart University, December 2014
Rieser, H. (2015). When Hands Talk to Mouth. Gesture and Speech as Autonomous Communicating Processes. Proceedings of Semdial 2015, Gothenburg
Rieser, H. (2017). A Process Algebra Account of Speech-Gesture Interaction. Preliminary version. Ms, Bielefeld University

March 1: Kristina Liefke "Relating Theories of Formal Semantics: established methods and surprising results"


Formal semantics comprises a plethora of theories which interpret natural language through the use of di¿erent ontological primitives (e.g. possible worlds, situations, individual concepts, unanalyzable propositions). The ontological relations between these theories are, today, still largely unexplored. In particular, it remains unclear whether the basic objects of some of these theories can be coded in terms of objects from other theories (s.t. phenomena which are modeled by one theory can also be modeled by the other theories) or whether some of these theories can even be reduced to ontologically poor(er) theories (e.g. extensional semantics) which do not contain ¿special¿ objects like possible worlds.

This talk surveys my recent work on ontological reduction relations between formal semantic theories. This work shows that, more than preserving the modeling success of the reduced theory, some reductions even improve upon the theory's modeling adequacy or widen the theory's modeling scope. My talk illustrates this observation by two examples: (i) the relation between Montague-/possible world-style intensional semantics and extensional semantics (cf. Liefke and Sanders 2016), and (ii) the relation between intensional semantics and situationbased single-type semantics (cf. Liefke and Werning, in revision). The first relation established through the use of associates from higher-order recursion theory.

February 20: Asad Sayeed "Semantic representation and world knowledge"

SlidesWhile general knowledge of the world plays a role in language use,
language processing in humans is also guided by formal intuitions about
linguistic representation. In this talk, I discuss research results in finding the boundaries between world knowledge and formalism-driven intuitions and situate them in the context of a larger research program in computational psycholinguistics.

The first result focuses on the semantics of predicates and their arguments and how they are interpreted by the human processor. English-speaking human raters judge doctors as more appropriate givers of advice than recipients and lunches as much more appropriate objects of "eat" than subjects. One of my recent projects resulted in the development of vector-space and neural network models of predicate-argument relations that model that succeed in achieving high correlations with human ratings.

The second result is about the interaction of world knowledge with higher order semantics. English-speakers tend to judge that the sentence "every child climbed a tree" refers to more than one tree, while "every jeweller appraised a diamond" is comparatively more likely to refer to a single diamond, based on their knowledge of trees and diamonds. Recent experimental results in the literature are ambivalent on the extent to which formal structure affects the power of world knowledge to influence these judgements. In response to this, I describe a recent judgement study I conducted using German scrambling that suggests a significant effect of formal representation on the plural interpretation of an object argument given a universally-quantified subject.

Both of these research efforts reveal underlying questions about the
influence of world knowledge on linguistic representations and suggest
ways to answer them.

February 1: Mathieu Lafourcade "Games with a Purpose: The JeuxdeMots project"


Human-based computation is an approach where some steps of a computation is outsourced to humans. Games with a purpose (GWAPs) are games aiming at resolving puzzled or collecting data, where humans still outperform machines.

The JeuxDeMots (JDM) projects is a web-based associative GWAP where people are invited to play on various lexical and semantic relations between terms. The aim of this project is to build a large lexico-semantic network, with various relations types and word refinements (word usages).

Text semantic analysis is the main application for exploiting this resource, however the use as a tool for providing help in the case of the "tip of the tongue" phenomenon is also fruitful. This presentation will present the principles behind the JDM project, as well has the results achieved so far (around 1 million terms for 67 million relations). The following aspects will be discussed: the interaction between the various games of the JDM environment, some inference mechanisms of relations and rules, word polarity and sentiments, and some ethical aspects. Some specific aspects of the JDM lexical network are detailed, such as : refinements, aggregated terms, inhibitory relations and relation annotations.


M. Lafourcade, N. Le Brun, and A. Joubert (2015) Games with a Purpose (GWAPS) ISBN: 978-1-84821-803-1 July 2015, Wiley-ISTE, 158 p.

M. Lafourcade, A. Joubert (2015) TOTAKI: A Help for Lexical Access on the TOT Problem. In Gala, N., Rapp, R. et Bel-Enguix, G. éds. (2015), Language Production, Cognition, and the Lexicon. Festschrift in honor of Michael Zock. Series Text, Speech and Language Technology XI. Dordrecht, Springer. 586 p. 140 illus. ISBN: 978-3-319-08042-0. (pp. 95-112)



November 30: Peter Sutton "A probabilistic, mereological account of the mass/count distinction"


In this paper, we attempt to answer the vexing question why it should be the case that only certain types of noun meanings exhibit mass/count variation in the lexicalization of their semantic properties, while others do not. This question has so far remained unanswered, or been set aside. We will do so by focusing on the role of context-sensitivity (already highlighted in recent theories of the mass/count distinction), and argue that it gives rise to a conflict between two pressures that influence the encoding of noun meanings as mass or count, one stemming from learnability constraints (reliability) and the other from constraints on informativeness (individuation). This will also lead us to identifying four semantic classes of nouns, and to showing why variation in mass/count encoding is, on our account, to be expected to occur widely in just two of them. Context-sensitivity forces a choice between prioritizing individuation, which aligns with count lexicalization, and prioritizing consistency, which aligns with mass lexicalization.

November 14: Jean-Philippe Bernandy "Efficient Parallel and Incremental Parsing of Practical Context-Free Languages"


We present a divide-and-conquer algorithm for parsing
context-free languages efficiently. Our algorithm is an instance
of Valiant's (1975), who reduced the problem of parsing to matrix
multiplications. We show that, while the conquer step of
Valiant's is O(n³), it improves to O(log² n) under certain
conditions satisfied by many useful inputs.

One observes that inputs written by humans generally satisfy
those conditions. Thus, there appears to be a link between the
ability for a computer to efficiently parse an input in parallel
and the ability for a human to comprehend such an input.

November 2: Ruth Kempson "Language: The Tool for Interaction -- Surfing Uncertainty Together"


With established recognition of the endemic context-relativity of language, it is now generally accepted that both parsing and production involve incremental context-relative decisions, requiring the concepts of both evolving contents and evolving contexts. Researchers across semantics, pragmatics, psycholinguistics, and computational linguistics are duly turning to the challenge of modelling language in terms that are compatible with such incrementality. Yet formal models of language remain largely grounded in the static terms of licensing sentential string- interpretation pairings reflecting only concepts such as compositionality, with little or no reflection of a time-linear process of information growth.

In this talk, I start by showing why linguists cannot avoid the challenge of defining grammar formalisms to reflect the dynamics of conversational dialogue, and how in order to achieve this, every aspect of linguistic knowledge needs to be recast as procedures for on-line incremental and predictive word-by-word understanding/production. I shall then briefly sketch the action-based Dynamic Syntax (DS) system to demonstrate its explanatory potential, by modelling what have been taken as canonical exemplars of semantic-independent syntactic processes, which in DS are all expressed in terms of incremental parsing/generation actions. I will show in passing how the resulting system, despite the lack of any conventional notion of syntax, nonetheless has the power to express both universal structural constraints and yet cross-language variability. Part of this will include the Directed Acyclic Graph characterisation of context as developing in lockstep with the evolving yet revisable content, demonstrating the system-internal potential for self/other-correction. The dynamics of conversational dialogue interactions will then emerge as the immediate consequence of this perspective on language; and I will briefly illustrate how this potential for interaction underpins all types of language-internal licensing constraint: syntactic, semantic, morphosyntactic and phonological.

I shall then turn to setting this perspective within the Predictive Processing (PP) model of cognition (Clark 2016), whose architectural properties the DS concept of language matches almost point by point. Like perception in the PP model, the DS grammar is a "fundamentally action-oriented" set of procedures, grounded in predictive processing resources shared by speakers (action) and hearers (perception) alike and "executed using the same basic computational strategy" leading to effects of interactive coordination without any need to invoke mind-reading or propositional inference. The result is that linguistic processing, perception, action, and thought are predicted to be "continuously intermingled" yielding representational updates "tailored to good enough online controls rather than aiming for rich mirroring". Instead, such updates are accomplished due to a strong version of affordance competition since the brain ¿continuously computes multiple probabilistically inflected possibilities for action¿ in a cost-effect balancing dynamic, with possibilities progressively winnowed down, allowing for possible revision, to yield at least one output in any successful outcome. To this set of characteristics (Clark 2016 p. 251), we have only to add the potential for interaction which such a language system predicts as default, and a wholly different perspective on language evolution opens up. Language can now be seen as an emergent and evolving system with manifest potential for consolidating cross-individual interactions, hence group effects, without ever having to invoke high-level inferences as external, "designer"-imposed motivation for such consolidation, this a dynamic for which language change already provides robust motivation.

October 18: Matthew Stone "A Bayesian model of grounded color semantics"


Natural language meanings allow speakers to encode important real-world distinctions, but corpora of grounded language use also reveal that speakers categorize the world in different ways and describe situations with different terminology. To learn meanings from data, we therefore need to link underlying representations of meaning to models of speaker judgment and speaker choice. This paper describes a new approach to this problem: we model variability through uncertainty in categorization boundaries and distributions over preferred vocabulary. We apply the approach to a large data set of color descriptions, where statistical evaluation documents its accuracy. The results are available as a Lexicon of Uncertain Color Standards (LUX), which supports future efforts in grounded language understanding and generation by probabilistically mapping 829 English color descriptions to potentially context-sensitive regions in HSV color space.

joint work with Brian McMahan.

September 13: Ev Fedorenko "The internal architecture of the language network"


A set of brain regions on the lateral surfaces of left frontal, temporal, and parietal cortices robustly respond during language comprehension and production. Although we now have strong evidence that this language network is spatially and functionally distinct from brain networks that support other high-level cognitive functions, the internal structure of the language network remains poorly understood. Deciphering the language network's architecture includes i) identifying its component parts, and ii) understanding the division of labor among those components in space and time. I will first present evidence that all language regions closely track linguistic input. I will then argue that some of the traditional "cuts" that have been proposed in the literature (e.g., based on the size of the linguistic units, based on the distinction between storage and computation, or based on syntactic category) do not seem to be supported by the available evidence. Even aspects of language that have long been argued to preferentially, or selectively, rely on a specific region within the language network (e.g., syntactic processing being localized to parts of Broca¿s area) appear to be distributed across the network. Further, the very same regions that are sensitive to syntactic structure in language show sensitivity to lexical and phonological manipulations. This distributed nature of language processing is in line with much current theorizing in linguistics and the available behavioral psycholinguistic data that show sensitivity to contingencies spanning sound-, word- and phrase-level structure. Time permitting, I will talk about recent work on decoding single word meanings and more complex meanings from the neural activity in the language network, and speculate that the organizing principles of the language network may have to do with meaning.

Relevant readings:

September 12: Ted Gibson "Information processing and cross-linguistic universals"


Finding explanations for the observed variation in human languages is the primary goal of linguistics, and promises to shed light on the nature of human cognition. One particularly attractive set of explanations is functional in nature, holding that language universals are grounded in the known properties of human information processing. The idea is that grammars of languages have evolved so that language users can communicate using sentences that are relatively easy to produce and comprehend. In this talk, I summarize results from explorations in two linguistic domains, from an information-processing point of view.

First, I consider communication-based origins of lexicons of human languages. Chomsky has famously argued that this is a flawed hypothesis, because of the existence of such phenomena as ambiguity. Contrary to Chomsky, we show that ambiguity out of context is not only not a problem for an information-theoretic approach to language, it is a feature. Furthermore, word lengths are optimized on average according to predictability in context, as would be expected under an information theoretic analysis. We then apply this simple information-theoretic idea to a well-studied semantic domain: words for colors. And finally, I show that all the world's languages that we can currently analyze minimize syntactic dependency lengths to some degree, as would be expected under information processing considerations.


Piantadosi, S.T., Tily, H. & Gibson, E. (2012). The communicative function of ambiguity in language.Cognition 122: 280-291.

Piantadosi, S.T., Tily, H. & Gibson, E. (2011). Word lengths are optimized for efficient communication.Proceedings of the National Academy of Sciences 108(9): 3526-3529.

Futrell, R., Mahowald, K., & Gibson, E. (2015). Large-scale evidence of dependency length minimization in 37 languages. Proceedings of the National Academy of Sciences 112(33): 10336-10341. doi: 10.1073/pnas.1502134112.

September 8: Ev Fedorenko "The human language network within the broader architecture of the human mind and brain"

 Link to the recorded talk


Although many animal species have the ability to generate complex thoughts, only humans can share such thoughts with one another, via language. My research aims to understand i) the system that supports our linguistic abilities, including its neural implementation, and ii) the relationship between the language system and the rest of the human cognitive arsenal. I use behavioral, fMRI, and genotyping methods in healthy adults and children, intracranial recordings from the cortical surface in patients undergoing pre- or intra-surgical mapping (ECoG), and studies of individuals with developmental and acquired damage.

I will begin by introducing the "language network", a set of interconnected brain regions that support language comprehension and production. With a focus on the subset of this network dedicated to high-level linguistic processing, I will then consider the relationship between language and non-linguistic cognition. Based on data from fMRI studies and investigations of patients with severe aphasia, I will argue that the language network is functionally selective for language processing over a wide range of non-linguistic processes that have been previously argued to share computational demands with language, including arithmetic, executive functions, music, and action/gesture observation. This network plausibly stores our linguistic knowledge, which can be used for both interpreting and generating linguistic utterances. Time permitting, I will speculate on the relationship between the language network and other networks, including, critically, the domain-general executive system, and the system that supports social cognition.

Relevant readings:

September 6: Ted Gibson "Language processing over a noisy channel"

Link to the recorded talk


Traditional linguistic models of syntax and language processing have assumed an error-free process of language transmission. But we know that this is not the case: people often make errors in both language production and comprehension. This has important ramifications for both models of language processing and language evolution. I first show that language comprehension appears to function as a noisy channel process, in line with communication theory.  Given si, the intended sentence, and sp, the perceived sentence we propose that people maximize P(si | sp ), which is equivalent to maximizing the product of the prior P(si) and the likely noise processes P(si → sp ).  I show how this simple formulation can explain a wide range of language processing phenomena, such as people’s interpretations of simple sentences, some aphasic language comprehension effects, and the P600 in the ERP literature. Finally, I discuss how thinking of language as communication in this way can explain aspects of the origin of word order, most notably that most human languages are SOV with case-marking, or SVO without case-marking.


Gibson, E., Bergen, L. & Piantadosi, S. (2013). The rational integration of noisy evidence and prior semantic expectations in sentence interpretation. Proceedings of the National Academy of Sciences, 110(20): 8051-8056. doi: 10.1073/pnas.1216438110.

Gibson, E., Piantadosi, S., Brink, K., Bergen, L., Lim, E. & Saxe, R. (2013). A noisy-channel account of cross-linguistic word order variation. Psychological Science, 4(7): 1079-1088. doi: 10.1177.

Gibson, E., Sandberg, C., Fedorenko, E., Bergen, L., & Kiran, S. (2015). A rational inference approach to aphasic language comprehension. Aphasiology. doi: 10.1080/02687038.2015.1111994.

June 16: Carla Umbach "Ad-hoc Kind-formation by Similarity"


The talk focuses on demonstratives of manner, quality and/or degree, like German "so", Polish "tak", and English "such" (mqd demonstratives). These demonstratives modify (some or all of) verbal, nominal and degree expressions. They can be used deictically and anaphorically, and may also occur as correlatives in equative comparison constructions. The example in (1) shows German "so" used deictically.

(1) a. (speaker pointing to someone dancing): So tanzt Anna auch. 'Anna dances like this, too.' -- manner
b. (speaker pointing to a table): So einen Tisch hat Anna auch. 'Anna has such a table / a table like this, too.' -- quality
c. (speaker pointing to a person): So groß ist Anna auch. 'Anna is this tall, too.' -- degree

A semantic interpretation of mqd demonstratives will be proposed starting from the intuition that there is a deictic component and a similarity component involved ¿ in all of (1a-c), the meaning of "so" can be paraphrased by "like this". The basic idea is that mqd demonstratives generate a class of items similar to the target of the pointing gesture, e.g., in (1b) a class of tables similar to the table the speaker points at. This interpretation accounts for fact that mqd demonstratives are directly referential differing from regular demonstratives only in expressing similarity instead of identity. Moreover, it accounts for their cross-categorical distribution.

The suggested analysis is compatible with Carlson's (1980) interpretation of English "such" as a kind- referring expression. In the case of quality and of manner similarity classes will be shown to behave like kinds, although they need not be previously given but are instead ad-hoc generated. In the case of degree, however, it will be argued (contra Anderson and Morzycki 2015) that the resulting similarity class does not establish a kind. In (1c) for example, the class of persons similar in height to the one pointed at does not exhibit kind-like behavior.
The similarity interpretation of mqd demonstratives includes three major research topics:

(i) the implementation of the similarity relation, which is done with the help of multidimensional
attribute spaces
(ii) the ad-hoc generation of kinds by similarity, which is shown experimentally to be restricted to
particular features of comparison, and
(iii) the interpretation of equative comparison constructions based on similarity classes.

In the talk, the focus will be on the second topic.

Anderson, C., and M. Morzycki (2015) Degrees as kinds. Natural Language and Linguistic Theory.
Carlson, G. (1980) Reference to kinds in English. New York and London: Garland.
Gust, H. & C.Umbach (2015) Making use of similarity in referential semantics. In H. Christiansen, I. Stojanovic,
G. Papadopoulos (eds.) 9th Conference on Modeling and Using Context, Context 2015, LNCS Springer. Umbach, C., & H. Gust (2014) Similarity Demonstratives. Lingua 149, 74-93.

May 12: Simon Dobnik "A Model for Attention-Driven Judgements in Type Theory with Records"



Joint work with John D. Kelleher, Dublin Institute of Technology, Ireland

Type Theory with Records (TTR) has been proposed as a formal representational framework and a semantic model for embodied agents participating in situated dialogues (Dobnik et al., 2014). Although TTR has many potential advantages as a semantic model for embodied agents, one problem it faces is the combinatorial explosion of types that is implicit in the framework due to the fact that new types can be created dynamically by composing existing types.

A consequence of this combinatorial explosion is that the agent is left with an intractable problem of deciding which types to assign to perceptual data. The term judgement is the technical term used in TTR to describe the assignment of a type to perceptual data that in practice would be implemented as a sensory classification.

This paper makes 3 contributions to the discussion on the applicability of TTR to embodied agents. First, it highlights the problem of the combinatorial explosion of type assignment in TTR. Second, it presents a judgement control mechanism, based on the Load Theory of selective attention and cognitive control (Lavie et al., 2004), that addresses this problem. Third, it presents a computational framework, based on POMDPs (Kaelbling et al., 1998), that offers a basis for future practical experimentation on the feasibility of the proposed approach.


Simon Dobnik is Senior Lecturer at the University of Gothenburg

May 4: Shay Cohen "Latent-Variable Grammars and Natural Language Semantics"



Probabilistic grammars are an important model family in natural language processing. They are used in the modeling of many problems, mostly prominently in syntax and semantics. Latent-variable grammars are an extension of vanilla probabilistic grammars, introducing latent variables that inject additional information into the grammar by using learning algorithms in the incomplete data setting.

In this talk, I will discuss work aimed at the development of (four) theoretically-motivated algorithms for the estimation of latent-variable grammars. I will discuss how we applied them to syntactic parsing, and more semantically-oriented problems such as machine translation, conversation modeling in online forums and question answering.

Shay Cohen is a Chancellor's Fellow at the Institute for Language, Cognition and Computation, University of Edinburgh.

April 27: Zhaohui Luo "MTT-semantics Is Both Model-theoretic and Proof-theoretic"



In this talk, after briefly introducing the formal semantics in modern type theories (MTT-semantics), I shall argue that it is both model-theoretic and proof-theoretic. This is due to the unique features of MTTs: they contain rich type structures that provide powerful representational means (e.g., to represent collections as types) and, at the same time, are specified proof-theoretically as rule-based systems whose sentences (judgements) can be understood inferentially.

Considered in this way, MTTs arguably have promising advantages when employed as foundational languages for formal semantics, both theoretically and practically.


Zhaohui Luo is a Professor of Computer Science at Royal Holloway, University of London. 

April 7: Staffan Larsson "Bayesian nets in probabilistic TTR"



There is a fair amount of evidence indicating that language acquisition in general crucially relies on probabilistic learning. It is not clear how a reasonable account of semantic learning could be constructed on the basis of the categorical type systems that either classical or revised semantic theories assume. We present probabilistic TTR (Cooper et al 2014) that makes explicit the assumption, common to most probability theories used in AI, that probability is distributed over situation types, rather than over sets of worlds. Improving on and going beyond Cooper et al (2014), we formulate elementary Bayesian classifiers (which can be modelled as two-layer Bayesian networks) in probabilistic TTR and use these to illustrate how our type theory serves as an interface between perceptual judgement, semantic interpretation, and semantic leaning. We also show how this account can be extended to cover general Bayesian nets.


Staffan Larsson is a professor of computational linguistics at CLASP.

March 16: Graeme Hirst "Who decides what a text means? (And what the answer implies for computational linguistics)"



Writer-based and reader-based views of text-meaning are reflected by the respective questions "What is the author trying to tell me?" and "What does this text mean to me personally?" Contemporary computational linguistics, however, generally takes neither view; applications do not attempt to answer either question.

Instead, a text is regarded as an object that is independent of, or detached from, its author or provenance, and as an object that has the same meaning for all readers. This is not adequate, however, for the further development of sophisticated NLP applications for intelligence gathering and question answering, let alone interactive dialog systems.

I will review the history of text-meaning in computational linguistics, discuss different views of text-meaning from the perspective of the needs of computational text analysis, and then extend the analysis to include discourse as well -in particular, the collaborative or negotiated construction of meaning and repair of misunderstanding.

Graeme Hirst's research interests cover a range of topics in applied computational linguistics and natural language processing, including lexical semantics, the resolution of ambiguity in text, the analysis of authors' styles in literature and other text (including plagiarism detection and the detection of online sexual predators), identifying markers of Alzheimer's disease in language, and the automatic analysis of arguments and discourse (especially in political and parliamentary texts).

March 11: John Kelleher "Attention Models in Deep Learning for Machine Translation"


In the last number of years deep learning models have made a significant impact across a range of fields. Machine Translation is one such area of research. The development of the encoder-decoder architecture and its extension to include an attention mechanism has led to deep learning models achieving state of the art MT results for a number of langauge pairs.

However, an open question in deep learning for MT is what is the best attention mechanism to use. This talk will begin by reviewing the current state of the art in deep learning for MT. The second half of the talk will present a novel attention based encoder-decoder architecture for MT. This novel architecture is the result of collaborative research between John Kelleher, Giancarlo Salton, and Robert J. Ross.

John Kelleher is a lecturer in the School of Computing at the Dublin Institute of Technology and a researcher at the Adapt research center. He currently supervises research projects in a number of areas including machine translation, activity recognition and discovery, dialogue systems, computational models of spatial language, and music transcription.

For the last number of years the majority of his research has used a machine learning methodology, and in 2015 he published a textbook on machine learning with MIT Press. John's collaborators on this research are Giancarlo Salton, who is a PhD student at the Dublin Institute of Technology, and Robert Ross who is a senior lecturer in the School of Computing at the Dublin Institute of Technology.

March 9: Stergios Chatzikyriakidis "Modern Type Theoretical Semantics: Reasoning Using Proof-Assistants"



In this talk, I will discuss the use of Modern Type Theoretical Semantics  (MTTs) , i.e. type theories within the tradition of Martin Löf (1974, 1981), for reasoning about natural language semantics. I will first present a brief introduction of the features that make MTTs an attractive formal language to interpret NL semantics to. In particular, I will discuss a number of issues that have been successfully dealt with using MTTs like adjectival/adverbial modification, copredication and intensionality among other things.

Then, I will argue that the proof-theoretic nature of MTTs, i.e. the fact that they are proof-theoretically specified, in combination with their expresiveness makes them fit to perform reasoning tasks. This proof-theoretic aspect of MTTs has been the main reason that a number of proof-assistants implement variants of MTTs. One such proof-assistant, Coq, will be used as a way to show the applicability of MTTs in dealing with Natural Language Inference (NLI).

Firstly, I will show how NL semantics can be implemented in Coq and
then I will present how one can use Coq in order to reason with these
semantics. I will draw examples from the FraCas test suite platform in order to show the predictions the implemented semantics make as regards inference. I will then discuss issues like coverage and proof-automation and a number of ideas for future work, like extracting type ontologies from GWAP lexical networks and creating a parser/translator that will translate between English (or any other language) and the syntax of Coq.

I will end the talk by discussing the potential use of Coq implementing other semantic frameworks, like Montague Semantics, Davidsonian semantics and eventually a discussion on how Coq can be used with TTR (or even ProbTTR).


Stergios Chatzikyriakidis is a researcher and research coordinator of CLASP. 

February 22: Jan van Eijck "Modelling Legal Relations"


Jan van Eijck, CWI and ILLC, Amsterdam (http://homepages.cwi.nl/~jve/)

(joint work with Fengkui Ju, Beijing Normal University, Beijing, China)

We use propositional dynamic logic and ideas about propositional control
from the agency literature to construct a simple model of how legal
relations interact with actions that change the world, and with actions
that change the legal relations.

This work is relevant for attempts to construct restricted fragments of
natural language for legal reasoning that could be used in the creation of
(more) formal versions of legal documents suitable for `legal knowledge bases'.

February 18: Charalambos Themistocleous "Doing Type Theory in R"



In this talk, I will present R language (or simply R ), a dynamic, lazy, functional, programming language that was designed in 1993 by Ross Ihaka and Robert Gentleman. R adopts the underlying evaluation model of Scheme with the syntax of S, (a programming language, which was developed by John Chambers at Bell Laboratories).

R is an open-source programming language and the flexible statistical analysis toolkit implemented in R , made it the lingua franca for doing statistics. The R package repository (CRAN) features 7861 available packages, which extent the language. Also, there are guides on CRAN that group sets of R packages and functions by type of analysis, fields, or methodologies (e.g. Bayesian Inference, Probability Distributions, Machine Learning, Natural Language Processing).

The statistical capabilities of R along with its functional capabilities can transform R into a rich environment for doing Type Theory. Thus, I will conclude this talk by discussing possible extensions of R for A Probabilistic Rich Type Theory for Semantic Interpretation (Cooper, Dobnik, Lappin, and Larsson, 2015).


Charalambos Themistocleous is a post-doc at CLASP. 


Page Manager: Stergios Chatzikyriakidis|Last update: 4/30/2020

The University of Gothenburg uses cookies to provide you with the best possible user experience. By continuing on this website, you approve of our use of cookies.  What are cookies?