Recent years have seen increasing interest in service robots endowed with communicative capabilities. These robots must often operate in open-ended environments and interact with humans using natural language to perform a variety of service-oriented tasks. Yet, developing dialogue systems able to robustly and accurately interpret spoken utterances remains a significant challenge. This is notably due to the presence of numerous speech disfluencies (pauses, repetitions, repairs, etc.) and to the globally poor performance of speech recognition, especially when operating in noisy, real-world environments. For my master thesis, I'm developing techniques and algorithms to improve the robustness of spoken dialogue comprehension at various processing stages, from speech recognition to the semantic interpretation. In this talk, I will focus on one particular technique: a model for speech recognition in natural environments which relies on contextual information about salient entities to prime utterance recognition. The hypothesis underlying our approach is that, in situated human-robot interactions, the speech recognition performance can be significantly enhanced by exploiting knowledge about the immediate physical environment and the dialogue history. To this end, visual salience (objects perceived in the physical scene) and linguistic salience (previously referred expressions within the current dialogue) are integrated into a single cross-modal salience model. The model is dynamically updated as the environment evolves, and is used to establish expectations about uttered words which are most likely to be heard given the context. The update is realised by continuously adapting the word-class probabilities specified in the statistical language model. Our work draws upon insights from cognitive science, as we know that humans systematically exploit dialogue and situated context to guide attentional mechanisms and help disambiguate and refine linguistic input by filtering out unlikely interpretations. In this talk, I will discuss the motivations behind our approach, describe our implementation as part of a distributed, cognitive architecture for mobile robots, and report the evaluation results on a test suite, which show a statistically significant improvement on the recognition rate.

In this research we want to use the Interaction Grammar to represent the construction of Persian noun phrases in trees. XMG is utilized for representing the constructions by using factorization and inherited hierarchy relations. Then with the help of XMG, we define IG by taking the advantages of polarities on the features, and tree descriptions for the various constructions that are introduced. Then we use Leopar for the graphical representations of the trees and parsing them. Finally we apply our test suites to the parser in order to check whether we have the correct parsing and representation of the phrases. The experimental result showed that we could parse the phrases successfully, even the most complex one which has various constructions in it.

In this talk, I will present my Master's thesis, which is aimed at developing a core computational grammar for English within a framework specially designed for factorising tree-based grammar and grammars equipped with a syntax/semantic interface. I will begin by outlining the context for the specification of the grammar, namely the linguistic decisions that were taken and the language of description. This includes familiarising the audience with large-scale tree-adjoing grammars (our starting point for the current grammar development) and the language of description (which provides a high level of abstraction and allows for meta-grammar specification). Further on, I will report on the adopted methodology, which was scaled up by implementing an actual grammar. I will conclude by reporting on the results and the phenomena covered. I will also outline some possible directions for future work.

I will show how a simple error mining algorithm running on corpus parsing results and coupled with an adapted interface may help to discover errors that are always present in large linguistic resources such as lexica and grammars. This work has since been completed by preliminary experiments to suggest corrections for potential lexical errors. The error mining algorithm can also be extended to deal with more complex cases of errors, for instance on form, lemma or category sequences that may suggest grammatical errors rather than lexical ones. Finally I will say a few words about the adaptation of the algorithm for knowledge acquisition tasks and the relationships of the algorithm with Expectation Maximization (EM) algorithm.

In this talk, I will present the applications of error-mining based techniques in two linguistic knowledge acquisition tasks with a large-scale HPSG for English. First, in the context of general purpose lexical acquisition, I will report the practical application of error-mining for the detection of missing lexical entries on the British National Corpus. As an extension to the standard error-mining technique, I will further report our recent chart-mining experiments for verb-particle acquisition, which extract indications of linguistic regularities from intermediate parsing results.

In the first part of the talk I focus on the linguistic properties of Multiword Expressions (MWEs), taking a closer look at their lexical, syntactic, as well as semantic characteristics. In the second part of the talk I focus on methods for the automatic acquisition of MWEs for robust grammar engineering. First I investigate the hypothesis that MWEs can be detected by the distinct statistical properties of their component words, regardless of their type, comparing various statistical measures, a procedure which leads to extremely interesting conclusions. I then investigate the influence of the size and quality of different corpora, using the BNC and the Web search engines Google and Yahoo. I conclude that, in terms of language usage, web generated corpora are fairly similar to more carefully built corpora, like the BNC, indicating that the lack of control and balance of these corpora are probably compensated by their size. Finally, I show a qualitative evaluation of the results of automatically adding extracted MWEs to existing linguistic resources. I argue that the process of the automatic addition of extracted MWEs to existing linguistic resources improves qualitatively, if a more compositional approach to the grammar/lexicon automated extension is adopted.

Nessie is a tool that computes semantic representations for complx syntactic units. It does so thanks to a lexicon that stores representations of words (described by lambda-terms) and an abstract syntax tree that guides the semantic construction process. Abstract Categorial Grammars (ACGs), on the other hand, are an abstracon of the syntax-semantics interface. They allow one to map a syntactic structure to a semantic representation. We will introduce these two notions and list some properties of ACGs explaining why they are interesting. The heart of the presentation will consist in showing that the languages Nessie can recognize are exactly those recognized by an ACG of order 2. We will conclude by deducing some interesting properties this result confers to Nessie.

When interlocutors are engaged in situated dialogue, their informational states evolve not only through dialogue acts but also through physical and sensing acts. All these acts can be performed either explicitly or tacitly during the interaction. However, even when acts are performed tacitly, often their execution can be inferred from subsequent acts. I present a model for the inference of tacit sensing acts in a non-traditional conversational system (a text-adventure game) using a non-traditional automated planner (an Artificial Intelligence planning system that is able to find plans when knowledge is incomplete).

GerTT is a TAG-based grammar for German, that is currently being implemented using an extension of multicomponent TAG (MCTAG), called TT-MCTAG (MCTAG with shared nodes and tree tuples). In our talk, we will motivate this formalism of choice and sketch the analyses of several aspects of German syntax. In particular, we will concentrate on coherent constructions and scrambling and on complementation. We will also point out some limitations of these analyses. Furthermore, the integration of lexical ressources will also be considered. Finally, we will mention the undespecified semantics for semantic processing, that is included in GerTT.

In this presentation, we introduce a formalization of various elliptical coordination structures within the Multi-Component Tree Adjunct Grammar framework. Numerous authors describe elliptic coordination as parallel constructions where symetric derivations can be observed from a desired predicate-argument structure analysis. We show that most famous coordinate structures, including zeugma constructions and coordination of unlike category, can be analyzed simply with the addition of a simple synchronous mechanism to the MCTAG framework . This mechanism consists of two steps : First we associate an unrealized elementary tree to a realized one inside a tree set, then we add shared links between derivation nodes. Thus using the extended domain of locality provided by MCTAGs most of coordinate structures which cannot be analyzed by Lexicalized Tree adjunct Gramar are given a proper MC-TAG analysis. We will discuss possible implemenation using tree factorization and discuss some possible extensions to this analysis to deal with control-verb phenomenon.

In this talk, we will give an overview of the IG (Interaction Grammars) formalism. The specificities of the formalism (polarization and underspecified structures) will be illustrated through examples. The parser based on IG and developped by the Calligramme team will also be presented.

In this talk, we will present a simple account of coordination in Interaction Grammars that can be extended to non-constituent coordinations and cluster-argument coordinations. Furthermore, we will show how this account can give a simple method that improves polarity filtering for supertagging.