Discourse structure

A methodology for motivating a set of coherence relations

My PhD work (Knott, 1996) looked at the question of how to decide in a principled way on a set of coherence relations to use in analysing and generating text. Although the general idea of coherence relations is widely accepted in computational treatments of discourse structure, there is considerable disagreement amongst researchers as to the nature of relations themselves: how many are needed, how they should be defined, and what exactly they model. No two researchers use the same set of relations, and new relations are constantly being created---the resulting proliferation makes for a great deal of confusion. In my thesis I propose a methodology for determining a standard, well-motivated set of relations. The methodology is founded on a conception of relations as modelling cognitive constructs, used by readers and writers when they process text. I argue that evidence for such psychological constructs can be sought in a study of the linguistic resources for signalling relations in surface text, and in particular in a study of the set of connective cue phrases in a language (Knott and Dale, 1992). On the basis of this argument, a three-stage method for motivating relations is proposed (Knott and Dale, 1996; Knott, 1993b). First, a very large corpus of cue phrases is gathered from naturally-occurring texts, using a simple pre-theoretical test. Second, this corpus is organised into a taxonomy of synonyms and hyponyms, using a second pre-theoretical test to determine the substitutability of one phrase by another in a range of contexts. The taxonomy motivates a feature-theoretic conception of relations, whereby cue phrases signal combinations of features of coherence relations, rather than whole relations (Knott, 93a). The final stage in determining relation definitions is to use the taxonomy to define a set of independent features, representing orthogonal dimensions of variation within the set of cue phrases (Knott and Mellish, 96).

Interdisciplinary studies of cue phrases and relations

The feature-theoretic conception of cue phrases and relations developed in my thesis has served as the basis for subsequent research in several areas. One group of studies examines the cross-linguistic validity of the proposed set of features. Studies have been carried out on English and Dutch (Knott and Sanders, 96) English and German (Stede, 94), and French (Rossari and Jayez, 98). The feature-based conception of cue phrases and relations has also found application in computational treatments of discourse structure and lexical semantics. The conception meshes well with emerging accounts of discourse structure in terms of lexicalised tree-adjoining grammars (Cristea and Webber, 97; Webber and Joshi, 98; Webber, Knott and Joshi, 99; Webber et al, 99), and has also formed the basis for an analysis of subordinating conjunctions in a lexical knowledge base (Litkowski, 98). Another group of studies focus on the issue of cue phrase ambiguity. The feature-based account of cue phrases sheds interesting light on the question of whether very general cue phrases such as ``and'' and ``but'' should be thought of as polysemous or underspecified, from a Gricean standpoint (Oberlander and Knott, 96). It has also proved useful in interpreting the results of psychological studies in which cue phrases are used as an experimental window on subjects' discourse processing strategies. A recent study (Stevenson et al, in preparation) notes the problems posed by ambiguous cue phrases and reports new experiments using maximally specific phrases. Another psychological study finds independent evidence for the feature-based account of relations from cluster analyses of disagreements between text analysts (Knott and Sanders, 96). A final strand of research emerging from the study of cue phrases is corpus-based. The large collection of cue phrases gathered during the study has been used in studies of the distribution of cue phrases in large corpora (Marcu, 97; Cristea and Webber, 97).

Natural language generation and text planning

A final research interest is in natural language generation. The focus of this work to date has been the ILEX project, on which I worked from 1995 to 1998, along with Mick O'Donnell, Jon Oberlander, and Chris Mellish.

The ILEX system

ILEX (the Intelligent Labelling Explorer) is a text generation system which operates in a museum gallery, producing descriptions of objects encountered during a personalised guided tour. The current version of the system runs over the web, delivering pictures and text for a collection of objects in the Modern Jewellery gallery of the Royal Museum of Scotland. Descriptions are generated at run-time, from a knowledge base of facts. They are individually tailored to the communicative context in which they are generated, featuring comparisons to objects already seen, relevant examples and interesting background information, and avoiding repetition of facts already presented: see Oberlander et al, 98 for an overview. ILEX is a Dynamic Hypertext system; one of a number of recent text generation systems investigating a new and potentially very interesting paradigm in human-computer interaction (Knott et al, 96, Dale et al, 98). Interaction with the user can be thought of as a form of mixed-initiative dialogue: the user is free to browse through the collection of objects in any order; the system decides how to describe each selected object, in such a way as the {\em sequence} of object descriptions forms a coherent whole.

Text planning in ILEX

One novel aspect of ILEX's domain is that the system is not able to plan far into the future, as it cannot know which objects the user is going to choose. Moreover, for any given object description, the system's communicative goal is very underspecified: it must simply present as much interesting and educationally important information to the user as is possible within a locally and globally coherent text. This means that conventional text-planning paradigms, which rely on the decomposition of high-level communicative goals and the construction of large hierarchical plans, are not applicable. The system therefore makes use of a notion of {\em opportunistic} generation (Mellish et al, 98b), in which a network of interrelated goals is provided, along with a set of rules specifying when the satisfaction of one goal places another related goal on the agenda. This framework provides a good platform for experimenting with different bottom-up text planning algorithms. We have so far considered a range of stochastic search techniques (Mellish et al, 98a), and a new method for integrating constraints due to coherence relations with constraints due to focussing mechanisms (Oberlander et al).

Natural language generation and computational semantics

A second interesting line of research in ILEX relates to theories of natural language semantics; in particular to accounts of generic propositions and non-standard quantifiers. Often, the most important information to be conveyed by a museum guide is not about the particular artefacts in the gallery, but about general classes of objects of which they are representatives. Integrating generic propositions appropriately into descriptions of individual objects is a difficult problem, and one which raises many active research issues in formal semantics (Knott et al, 97). However, I believe that addressing some of these issues from the perspective of natural language generation (rather than from the traditionally-adopted perspective of interpretation) could yield some interesting insights; and this is a direction I would like to pursue.