Multi-word expressions (MWEs) have attracted much attention in NLP over the last decade or so, and in general linguistics, the interest in phraseology - which includes the linguistic study of MWEs - goes back much further. However, the broad comparative approach characteristic of research in linguistic typology seems not to have played any role in any of this work so far. On the contrary, comparative studies of MWEs in NLP (or phraseology in linguistics) have generally been contrastive rather than typological in scope, i.e., they deal with (a convenience sample of) a few languages, rather than with a systematic typological sample representative of the world’s linguistic diversity, with the result that no unitary cross-linguistically valid notion of MWE can be found in the literature. Approaching MWEs from a broad cross-linguistic perspective raises a number of intriguing theoretical and methodological questions, for linguistics and NLP alike. In fact, closer connections between NLP work on MWEs and linguistic research on lexical and semantic typology could have an important role to play for developing new research directions in both fields.
In this talk I report on-going work on the use of Modern Type Theories (MTTs), i.e. type theories within the tradition of Martin Löf (1974, 1981), in the study of linguistic semantics. In particular, I exemplify the use of Luo’s type theory with coercive sub-typing and show its applicability for a wide range of semantic phenomena, including adjectival/adverbial modification, co-predication and belief intensionality among others. I will then argue that the proof-theoretic nature of MTTs has the further advantage that these can be further implemented into reasoning engines in order to perform reasoning tasks. In particular, this proof-theoretic aspect 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). Lastly, I will discuss the issue of introducing probabilities into type theory. In particular, I want to focus on the problems that such endeavour might face and discuss possibilities on how such an extension of MTTs can be at least initiated.
The origin of syntactic structure has been a difficult problem for theoretical linguistics for many decades. One standard assumption is that it must be bootstrapped in some way from some external source of information, typically some hierarchically structured semantic representation. We will explore a radical alternative to this view: that distributional patterns in the raw data may suffice. Under some circumstances formal languages will have a unique canonical grammar which can form the basis for a learning algorithm which recovers a reasonable notion of syntactic structure: this approach relies on identifying irreducible elements of an algebraic structure - the syntactic concept lattice - canonically associated with every formal language.
I will give an overview of our work on building computational models of meaning of spatial descriptions in dialogue interaction such as "the chair is to the left of the table" or "turn right at the next crossroad" which include both linguistic and perceptual representations, for example those used in computer vision and robotics. As such models interface perceptional and conceptual domains they invariably require an application of statistical models and machine learning. Physical sciences have developed ways in which space can be described with high degree of accuracy, for example by measuring distances and angles. Such measures can be represented on a continuous scale of real numbers. However, humans refer to space quite differently: they use reference to discrete units such as points, regions and volumes and they also take into account what they know about the world and the objects, for example the dynamic kinematic routines between them. Spatial descriptions are also notoriously underspecified and vague and they have to be interpreted against appropriate perceptual and discourse contexts. In my ongoing work with Robin Cooper, Shalom Lappin and Staffan Larsson on Type Theory with Records (TTR) I have tried to give this practical experience theoretical foundations by exploring how such models relate to linguistic theory, in particular to formal semantics, and use the models as a test-bed for theory development.
Distributed representations such as Google's word2vec and Stanford's GloVe which have emerged out of the Deep Learning research community, have been shown to be capture deep semantic information and thus constitute powerful and highly scalable data driven frameworks for NLP. We show examples from work of our research group of how they can be used for word sense induction and automatic document summarisation and how they can be extended to capture time dynamics of language change.
Many phenomena in lexical semantics seem to involve gradedness. Synonymy is a case at hand: Instead of absolute synonymy, we find near-synonymy of words that are often substitutable but still differ in nuances of meaning. Polysemy also seems to come in degrees, with different uses of a word differing in their perceived similarity. We use distributional models to describe degrees of similarity of word instances, and combine them with logical form representations of sentence meaning. In this talk, we show how to use Markov Logic Networks (MLNs) to perform probabilistic inference over logical form with weighted distributional inference rules for the task of Recognising Textual Entailment (RTE). We also speculate how a human agent could make use of distributional information and integrate it with everything else they know through a probabilistic framework. We argue that if semantics is a heterogeneous mess (which seems likely), it is important to find the right probabilistic framework for reasoning over it.
In this talk, we will outline our grant proposal 'IncReD' (Incremental Reasoning in Dialogue). This project aims to extend insights on incrementality in language processing beyond the utterance level. Reasoning - which often plays the role of providing coherence and structure in larger chunks of language - is also incremental in the sense that we tend to form hypotheses regarding the arguments of our conversational partners before these arguments are fully explicit. In this sense incrementality in reasoning is analogous to syntactic and semantic incrementality. We aim to combine insights from a variety of fields (e.g. Artificial Intelligence, Formal Linguistics, Psycholinguistics, Philosophy) and use corpus methods, state of the art experimental techniques (e.g. the Dialogue Experimental Toolkit (DiET)) and formal models from syntactic, semantic and pragmatic domains (e.g. Dynamic Syntax (DS) and Type Theory with Records (TTR)) to develop a model of dialogue that accounts for a range of dialogue phenomena including reasoning. Specifically we intend to investigate: (1) What types of reasoning do people use in dialogue? What resources does this reasoning rely on and how are these resources accessed incrementally? (2) What happens in a dialogue (linguistically and interactionally) when there is a mismatch in the resources for reasoning between participants? What factors influence the arguments a person uses when conflicting resources are available? (3) How can this incremental human reasoning ability be formally modelled?
The question of whether grammatical competence should be represented by a formal grammar that provides a binary membership condition for the set of well-formed sentences (and their associated structures) in a language, or as a probabilistic system for determining relative values of grammatical acceptability has been a central issue in computational linguistics and cognitive science over the past two decades. In this talk I will present experimental evidence that speakers' judgements of grammatical acceptability are intrinsically gradient. I will show that unsupervised language models, augmented with grammatical scoring functions, can predict these judgements with an encouraging degree of accuracy over distinct domains and different languages. These results provide motivation for the view that grammatical competence is a probabilistic system. They also raise interesting questions about the nature of the language acquisition process. One of the main concerns of the talk will be to clarify the relationship between grammatical acceptability and probability. This work was done within the framework of my Economic and Social Research Council of the UK project Statistical Models of Grammaticality (SMOG).
We present a formal account of the meaning of vague scalar adjectives such as "tall" formulated in probabilistic Type Theory with Records. Our approach makes precise how perceptual information can be integrated into the meaning representation of these predicates; how an agent evaluates whether an entity counts as tall; and how the proposed semantics can be learned and dynamically updated through experience.
Universal Dependencies is a recent initiative to develop cross-linguistically consistent tree-bank annotation for many languages, with the goal of facilitating multilingual parser development, cross-lingual learning, and parsing research from a language typology perspective. In this talk, I outline the motivation behind the initiative and explain how the basic design principles follow from these requirements. I then discuss the different components of the annotation standard, including principles for word segmentation, morphological annotation, and syntactic annotation. I conclude with some thoughts on the challenges that lie ahead.
In formal language theory, a language is a set of strings and a grammar is an inductive definition of this set. It generates all and only the valid strings of the language. In linguistics, generative grammars have similarly aimed at defining all and only the valid sentences of natural languages - in other words, the notion of grammaticality. One objection to this idea is that grammaticality in natural language is not a binary notion but a gradient one. Does this imply that grammars are useless and should be replaced by probabilistic models of language?
In our view, not. We should just abandon the idea that the purpose of grammars is to define grammaticality. Grammars should rather be seen as ways of structuring data. Even statistical language models need grammars, often just very simple ones, For instance, a grammar might have the sequences of strings as its only structure. But we want to show how richer grammatical structures - often together with statistics - are a useful model that can for instance compensate for sparse data. We will also discuss the wide-spread beliefs that hand-written grammars cannot be robust and that they require too much work to be useful in practice.
Linguistic information is hard-coded in speech signal. By analysing specific acoustic properties, such as vowel formants and fundamental frequency, acoustic models of speech production aim to elicit this information. I summarise evidence from my research on speech processing and argue that these acoustic models provide only incomplete spectral description and under-represent interactions between acoustic properties. Consequently, they do not do justice to the complex linguistic information encoded in speech. I then propose a model for speech processing based on parameterised resonant signal elements and an algorithm that analyses vowel samples based on the proposed model. The algorithm provides a rich description of any given segmented vowel sample by using a large number of resonant elements with parameters that are chosen to accurately capture the time-frequency structure of the vowel. The parameters are then used to calculate probabilities. An application of the model successfully classifies vowels, stress, and speech variety. This model is an improvement over methods that only use a small number of formants to describe vowels, has the potential to be used in automatic speech recognition, and is promising for use in applications of forensic linguistics, and speech pathology. Finally, I discuss an ongoing work that aims to extend the model for the analysis of prosody.