ITRI Seminars - Summer 2001


		ITRI seminars - Summer 2001 ITRI seminars generally take place 12 noon on Thursdays in room W107 on the first floor of the Watts Building, University of Brighton (Moulsecoomb site). Occasional deviations from this pattern are indicated below. Information on how to find W107 is available on our contact page.
18 May abstract		Harry Bunt Linguistics and Computer Science, Tilburg University Transparency and Politeness in Human-Computer Dialogue NB: Friday, 12.00, in ITRI second floor meeting room
24 May abstract		Tijn Borghuis Eindhoven University of Technology Belief Revision with Explicit Justifications
7 Jun abstract		Marilyn Walker ATT Labs SPOT: A Trainable Sentence Planner NB 11.30 am start
14 Jun abstract		Nadjet Bouayad-Agha ITRI, University of Brighton Automatic generation of document structures from a discourse representation
21 Jun abstract		Carole Tiberius University of Surrey How to build a multilingual inheritance-based lexicon

		Previous ITRI seminars
		See also NLP seminars at COGS, University of Sussex

Abstracts

Harry Bunt
Transparency and Politeness in Human-Computer Dialogue

Is politeness a relevant issue in human-computer dialogue? One might think not, for what would be the point of being polite to a machine, and how could a machine possibly be thankful, respectful, apologetic, etc.? Still, in their book "The Media Equation" (1996), Reeves and Nass claim that politeness is indeed important in human-computer interaction, since people treat interactive machines fundamentally as social agents.

We will look at some corpus data and see that politeness expressions do indeed occur in human-computer dialogue, but also that they tend to take a different form than in human-human dialogue. We analyse these findings in terms of our context-change theory of dialogue, "Dynamic Interpretation Theory", which interprets dialogue utterances in terms of intended context-changing effects. We will suggest that the use of politeness expressions has a two-part explanation. One part is what might be called the "Principle of Conversational Transparency" (which relates to principles of transparency and observability in HCI), which says that cooperative dialogue participants act so as to make the state of the interaction transparent to each other; as such, politeness expressions form a source of information that is useful for successful communication. The other part is that natural human interaction is a form of social behaviour, and thus constrained by social conventions. We will finally discuss the consequences of this view for the design of dialogue systems.

Tijn Borghuis
Belief Revision with Explicit Justifications

An agent who updates his belief state with new information can reach a situation where his belief state becomes inconsistent and has to be revised. This talk explores belief revision for belief states in which an agent's beliefs as well as his justifications for these beliefs are explicitly represented. Treating justifications as first-class citizens allows for a deductive perspective on belief revision. We study the belief change operations emerging from this perspective in the setting of typed lambda calculus, and situate these operations with respect to standard approaches in the literature.

Marilyn Walker
SPOT: A Trainable Sentence Planner

Sentence planning is a set of inter-related but distinct tasks, one of which is sentence scoping, i.e. the choice of syntactic structure for elementary speech acts and the decision of how to combine them into one or more sentences. In this talk, I present a new sentence planner SPoT, and a methodology for automatically training it using on feedback provided by human judges. We reconceptualize the sentence planning task into two distinct phases. First, a very simple, randomiz sentence-plan-generator (SPG) generates a potentially large list of possible sentence plans for a given text-plan input. Second, the sentence-plan-ranker ( SPR) ranks the list of output sentence plans, and then selects the top-ranked plan. The SPR uses ranking rules automatically learned from training data. Then I discuss our evaluation of SPoT. We conducted an experiment to compare the output of SPoT with a currently operational hand-crafted template-based generator, two rule-based generators and two baseline generators. We show that SPoT performs as well as the hand-crafted generator, while requiring less development time, and better than the rule-based and baseline generators.

Nadjet Bouayad-Agha
Automatic generation of document structures from a discourse representation

The aim of this work is to allow a generation system to produce of variety of formatted texts, that is, texts structured in sections, paragraphs, lists, sentences, etc. The talk will be logically organised in three parts. First, I will motivate the use of the output of the document planning process, the "document structure", as a new level of data representation in the generation architecture. Next, I will discuss the requirements of the rhetorical input. Finally, I will present a method for the flexible mapping of the discourse representation onto several document structures. This method implements constraints for the partial ranking of the document structures according to their adequacy to the rhetorical input.

Carole Tiberius
How to build a multilingual inheritance-based lexicon

In this talk, I explore the use of inheritance networks to construct multilingual lexicons in which information can be shared at different levels of linguistic description (e.g. syntax, semantics, morphology, phonology), similar in spirit to the lexicons constructed in the PolyLex project (Cahill and Gazdar 1999) and the GREG project (Kilgarriff, Cahill and Evans 1999).

The multilingual nature of such lexicons raises methodological and theoretical issues for their design and development, such as:

The regulation of the inter- and intralanguage inheritance relations.
That is, how does the hierarchical structure of one language interact with the hierarchical structure between the languages.
Multilingual information sharing.
Which information can be shared in a multilingual inheritance network and how is it shared? For this talk, I focus on the sharing of morphological, phonological, and morphophonological information.
Development strategies.
How does one go about constructing a multilingual inheritance lexicon? Should the monolingual and multilingual hierarchical lexicons be developed in parallel and linked immediately upon construction or should a non-parallel development strategy be adopted, where the monolingual lexicons are first fully developed separately and only linked together at the end?

In this talk, I explore these issues by comparing different architectures for multilingual inheritance lexicons following Evans' (1996) proposals. In order to compare the different architectures, sample fragments have been implemented in DATR (Evans and Gazdar 1996), covering a small set of nouns and adjectives in Dutch, English, Danish, and Icelandic.

Cahill, L. and G. Gazdar. 1999. The PolyLex architecture: multilingual lexicons for related languages, In Traitement Automatique des Langues, 40:2, pp.5-23.
Evans, R. 1996. Exploiting inheritance in multilingual lexicons. Unpublished manuscript presented at the AISB-96 Workshop on Multilinguality in the Lexicon, Brighton.
Evans, R. and G. Gazdar. 1996. DATR: A Language for Lexical Knowledge Representation. In Computational Linguistics, 22(2), pp.167-216.
Kilgarriff, A., L. Cahill, and R. Evans. 1999. The GREG Framework for Multilingual Valency Lexicons. GREG Deliverable 2.1, ITRI.

Maintained by Adam Kilgarriff (Adam.Kilgarriff@itri.brighton.ac.uk).
Last updated Friday June 22 2001

©Information Technology Research Inst itute