Announcement for Workshop 5

"A Task-based Context Model for Answering Follow-up Questions" - Raffaella Bernardi and Manuel Kirschner, KRDB, Free University of Bozen-Bolzano

BoB is the multilingual chatter-bot developed at the Free University of Bozen-Bolzano (FUB) to help users obtain information about the Library's services (borrowing, ordering, reserving books, as well as finding out about Library opening hours, reading rooms etc.). Currently it is based on a shallow pattern matching approach which relies on hand-built regular expressions associated with pre-canned answers. For each language (German, English and Italian) it consists of around 700 pairs which have been written by FUB librarians starting from those developed in a ten year project at Hamburg University Library. In our present work, we address the limitations of this approach, focusing on follow up questions (FUs) which are tied to the dialogue context. The work we present has both theoretical and applied aspects. On the theoretical side, we propose a work-hypothesis about information seeking interaction and present the results obtained from a pilot study on FUs. We will show that users most often tend to ask FU questions about the actions or entities introduced in the previous few interactions. Based on these empirical results, on the applied side, we propose a task-oriented representation of the question-answer pairs that could help overcome the limitations mentioned above. This representation is based on the one hand on users' common actions, and on the other hand on entities from the library domain that are typically asked about. In our approach, we try to use this structure to model and predict the context of the interaction and hence improve the system's ability to answer FUs. We will describe a new search algorithm that is closely tied to the task (i.e. actions/entities) structure and hopefully give the results of a preliminary evaluation of the new framework, using the old system as a baseline.

Johan Bos

"Challenges in Open-Domain Question Answering" - Johan Bos, Università di Roma "La Sapienza"

Approaches to QA range from simple pattern-matching techniques to complex systems with deep linguistic analyses of question and potential answers. The simple techniques are sometimes interesting but do not scale. Complex systems are usually divided into manageable sub-components such as question analysis, document retrieval, answer candicate extraction, and answer selection and reranking. Using the Pronto QA system, I will illustrate what kind of problems an QA researcher encounters on each of these levels and what kind of solutions are implemented in state-of-the-art systems.

Robin Cooper

"Scaling up the formal language view to deal with innovation"- Robin Cooper, Göteborg University

We propose a shift in perspective away from the view of natural languages as formal languages (encapsulated in the title of one of Richard Montague's papers: "English as a Formal Language"). We propose that natural languages are instead to be seen as a collection of resources for constructing local languages for use in particular situations. This is suggested by the experience of constructing natural language grammars for particular applications using the Grammatical Framework developed by Aarne Ranta.

It points to a research programme investigating how such resources play a role in linguistic innovation by agents constructing situation-specific local languages and how they can be made dynamic, modified by the linguistic agent's exposure to innovative linguistic data.

As an illustration of this we look at some cases of lexical innovation using coercion which have been discussed in the Generative Lexicon literature by Asher and Pustejovsky. Our approach uses type theory with records which appears to offer some conceptual advantages in the treatment of innovation.

The material for this talk will be drawn from two papers available on: http://www.ling.gu.se/~cooper/records/nlres1.pdf and http://www.ling.gu.se/~cooper/records/copredinnov.pdf .

Arash Eshghi and Pat Healey

"Collective states of understanding" - Arash Eshghi and Pat Healey, QMUL

This paper uses an analysis of ellipsis in multi-party interaction to investigate the relative accessibility of dialogue context/common ground to direct addressees and side participants. The results show that side-participants frequently make direct use of the common ground established between a speaker and addressee despite the fact that, by definition, they did not directly collaborate with the speaker on constructing it. Moreover speakers seem to assume understanding on the part of the side-participants despite this lack of feedback. Different individuals can thus reach the same level of grounding through different levels of feedback. We conclude that multiparty dialogue leads to distinct collective states of understanding that are not reducible to the component dyadic interactions.

Tim Fernando

"Propositions-as-types in a natural language context" - Tim Fernando, Dublin

Constructive type theories have in recent years gained popularity as an approach to dialogue interpretation (e.g. Cooper, Ginzburg). A characteristic feature of constructive (as opposed to classical) type theories is the conception of a proposition as the type of its proofs (rather than say, the set of possible worlds in which the proposition is true). I will argue for a modification of this conception that decouples meaning (given by possible situations) from contextually supplied facts of the matter (for evaluating truth). This move is motivated by vagueness in natural language, and somewhat more speculatively, dialogue.

Raquel Fernandez and Stanley Peters

"Detecting local dialogue structure in open-domain, multi-party dialogue" - Raquel Fernandez and Stanley Peters, Stanford

While humans are able to listen to a meeting and recognize ``nuggets'' that have special importance without much effort, this is a difficult task for dialogue processing systems. In this talk we will examine how dialogue structure can be exploited to automatically detect interesting sparse conversational events (such as decisions, reports, and action item commitments) in open-domain spoken meetings.

Josef Genabith

"Treebank-Based Support Tools for Dialogue Systems" - Josef Genabith, Dublin City

This is joint work with Aoife Cahill, John Judge, Grzegorz Chrupala, Conor Cafferkey and Deirdre Hogan.

Jonathan Ginzburg

"A Deeper view of Grounding and CRification" - Jonathan Ginzburg, KCL

In this talk I will consider how to extend the issue-based approach to dialogue management to deal with conversational implicature and indirect speech acts. Although I will only offer a brief sketch, I will try to show that implicature fits in fairly straightforwardly into IBDM, exploiting its use of locutionary propositions in context. An altogether more challenging task are the diverse class of moves that, following Searle, came to be known as indirect speech acts (ISAs). Although there is a considerable literature on these, particularly under the rubric of recognizing the plan underlying an utterance, there has been little attempt to consider the cases where such recognition fails and some sort of repairing interaction takes place. I will sketch an account of several classes of ISAs, relating it to the basic issue of whether utterance plan recognition is a routine feature of utterance grounding. I will suggest that it is not, though a significantly easier task is --- genre recognition.

Pat Healey and Greg Mills

"New developments of the DiET chat tool: An experimental toolkit for the investigation of dialogue (incrementality and multi-party interaction)" - Pat Healey and Greg Mills, QMUL

Dialogue's inherently contextualised and interactive nature makes controlled experimentation problematic. Current techniques involve relatively coarse interventions, such as manipulations of task difficulty, changes in medium or the use of confederates who make scripted contributions.

In this talk we present a novel text-based experimental platform that addresses these problems by exploiting the fact that text-based dialogue is now a ubiquitous mode of communication and also by exploiting the increasing availability of real-time natural language processing techniques. We will demonstrate how the chat-tool goes beyond existing approaches by allowing fine-grained experimental manipulations which are sensitive to the syntax and semantics of the unfolding dialogue.

Fredrik Kronlid

"Scaling up Issue Based Dialogue Management" - Fredrik Kronlid, Göteborg

The problems of scaling up a two-party dialogue manager to handle also multi-party dialogue can be summarised as turn-taking problems, problems with obligations and rights to address issues, and grounding problems. We present an implementation of Issue Based Dialogue Management, MultiIBiS, scaled up to handle more than two parties. MultiIBiS makes a distinction beween "individual", "collective" and "distributive" questions in order to correctly handle obligations to answer questions, and features an implementation of the turn management model by Sacks, Schegloff and Jefferson to handle turn-taking, and an extended QUD-structure to handle obligations and rights to address issues. The implementation is not a complete dialogue system, but shows that it is indeed possible to scale up the IBDM model.

Udo Kruschwitz

"Towards Adaptive Dialogue-Driven Intranet Search" - Udo Kruschwitz, Essex

Until now most search engines have focused on individual queries and their answers. However, it is increasingly recognised that the key to retrieval is interaction. We are focusing on a search paradigm where automatically extracted domain knowledge is incorporated in a simple dialogue system in order to assist users in the search process. Instead of aiming at the Web in general we are primarily interested in smaller document collections with a more limited range of topics, e.g. academic and commercial Web sites and intranets. For such a setup our controlled task-based evaluations have shown that users strongly prefer a system that makes suggestions for query modifications over a standard search engine. The talk will go beyond controlled experiments and report on a longitudinal study of our system that has been in place as a search engine on a university intranet for more than a year. The talk will focus on our ongoing research and present some analysis of the log files collected so far.

Obviously, the term "dialogue" is used in a rather generic sense in this context (in fact, some workshop participants will probably argue that it it has been misused). Nevertheless, the rather pragmatic approach that we take in addressing a very practical problem will hopefully result in an interesting discussion.

Shalom Lappin

"Expressivity and Complexity in Underspecified Semantic Representation" - Shalom Lappin, KCL

Most systems for underspecified scope representation are not expressively complete in that they do not permit the enumeration of all possible subsets of scope interpretations. Ebert (2005) shows that a framework for underspecified scope representation is expressively complete iff it does not provide a tractable procedure for enumerating the set of possible scope interpretations. It is possible to reduce complexity in generating scope readings by using dynamic filters that are added incrementally in dialogue. However, such filters will only achieve limited savings in search space for certain non-worst case scenarios. The general complexity problem for underspecified scope representations, encoded as partial orders, turns out to be even more extreme that Ebert's result. I consider three possible ways of dealing with this complexity problem when generating scope interpretations from underspecified representations in the course of a dialogue.

Staffan Larsson

"Semantic Coordination in Dialogue: Outline of a Research Project " - Staffan Larsson, Göteborg University

SCID is a three-year project funded by The Swedish Bank Tercentenary Foundation, starting in January 2008. I will present the starting points for the project, and give an outline of our research agenda.

This project focuses on how linguistic meanings change, and how agents can coordinate on the meaning of words (and linguistic constructions in general). There is a growing perception that the mechanisms of change lie in communication in general and dialogue in particular, and that dialogue research has now got to a stage of maturity where it may have something interesting to say about semantic coordination and linguistic coordination in general.

An important practical application of the basic research proposed here is to make it easier for humans to interact with computers using natural language. Language is essentially a device for achieving informational coordination, that is, a device for making sure that we share information sufficiently to be able to get things done together. However, language is also something that needs to be coordinated in itself - we must share a language in order to share information. Sharing a language is a matter of linguistic coordination, that is, coordination with respect to how we talk: which words we use, what we mean by them, how we pronounce them, etc. To enable dialogue systems to coordinate their language in a similar way to humans, we need a formal theory of linguistic coordination. This project focuses more specifically on semantic coordination, i.e. coordination of meaning.

From a more theoretical perspective, the project is motivated by recent developments in international research in psycholinguistics and computational linguistics, and the observation that an important theory is lacking. There are interesting similarities and affinities between several disparate strands of research on coordination in dialogue, which open up the possibility of developing a theoretical framework where innovative research can be carried out.

In relation to the "scaling up challenge", this project can be regarded as an attempt to scale up the ideas behind the ISU approach to cover linguistic coordination. Also, we believe that a dialogue system capable of linguistic coordination would be better equipped to deal with dialogue in more complex domains such as open-domain question answering.

Oliver Lemon

"Statistical Learning and Probabilistic Planning for Dialogue Optimality" - Oliver Lemon, Edinburgh

I'll describe recent advances in decision-theoretic/statistical dialogue planning (e.g. TALK project results) which deal with the issues of noise, uncertainty, and optimization. To illustrate, I'll demonstrate a system (REALL) using reinforcement learning to optimize its dialogue plans under different noise conditions and time constraints, for different types of user, in the areas of Interactive Search, Self-Help/TroubleShooting, and Call Centre automation. I'll then describe the main research questions and directions arising from these techniques, including the new EC FP7 project "CLASSiC", with a focus on statistical planning for natural language generation in dialogue.

Paul Piwek

"From monologue to dialogue: mapping discourse to dialogue structure" - Paul Piwek, Open University

We discuss the problems and challenges that are posed by the task of automatically mapping monological text to dialogue structures (e.g., question-answer pairs). An automatically extracted collection of question-answer pairs is a potentially useful resource for question-answering systems. We focus on the specification of information-preserving mappings from monologue to dialogue, based on the discourse structure of the monologue, and discuss some initial results (from the domain of patient information leaflets).

Reference: Piwek, P., H. Hernault, H. Prendinger, M. Ishizuka (2007). T2D: Generating Dialogues between Virtual Agents Automatically from Text. In: Intelligent Virtual Agents, LNAI 4722, Springer Verlag, Berlin, pp.161-174.

Massimo Poesio

"Using lexical and commonsense knowledge for anaphora resolution: the next steps (with some ideas about dialogue)" - Massimo Poesio, Trento/Essex

Amanda Stent

"The impact of initiative on various measures of dialog quality" - Amanda Stent, Stony Brook

The SBU Rate-a-Course system is a spoken dialog system for collecting surveys. Students can call the system to rate a course or to obtain a summary of existing ratings for a course. In this talk, I will first describe the system and the types of variation that it contains. Then, I will present the results of an experiment we conducted with the system exploring the impact of different degrees of user dialog initiative on various measures of dialog quality. Although these results are particularly applicable to survey systems, they may also obtain to other tasks, e.g. tutoring, opinion mining.

If time permits, I will also talk about some work I am currently doing on grounding-based models of dialog in a classification framework.

Matthew Stone

"Discourse Coherence and Gesture Interpretation" - Matthew Stone, Rutgers

Joint work with Alex Lascarides

Ipke Wachsmuth

"Embodied Cooperative Systems: From Tool to Partnership" - Ipke Wachsmuth, Bielefeld

One of the most basic mental skills is inferring intentions - the ability to see others as intentional agents and to understand what someone is doing - which, in conjunction with other factors, brings about shared intentionality. Both, understanding others' intentions and representing them as being able to understand intentions, are relevant factors in coordinating actions, as is the ability to represent multiple goals and to update them dynamically. The incorporation of such faculties will also help to endow technical systems with collaborative functionality. "Embodied Cooperative Systems" is a new area in which we implement and evaluate cognitive interaction technology. Such systems are often embodied as robotic agents or as humanoid agents projected in virtual reality. In these contexts, the view that humans are users of a certain "tool" has shifted to that of a "partnership" with artificial agents, which can be considered as being able to take initiative as autonomous entities in cooperative settings. A central research question is how the processes involved interact and how their interplay can be modeled. For example, inter-agent cooperation relies very much on common ground, i.e. the mutually shared knowledge of the interlocutors. Mutual coordination is greatly facilitated by multimodality and especially by nonverbal behaviors. Gaze as well as pointing gestures are important means of coordinating attention between interlocutors and therefore related to both inferring and coordinating actions. We will outline these ideas taking the virtual humanoid agent "Max" as an example.

-- EleniGreg - 11 Jan 2008