Staffan Larsson, University of Gothenburg

2008-2012

The starting point of this project is a view of linguistic meaning as something fundamentally dynamic and social. This view has become more widespread during the last decade or so; however, formal semantics has not adapted to this development. The purpose of the present project was to develop a formal theory of how linguistic meaning is negotiated in dialogue, as well as a formal pragmatic model of the strategies that are used to coordinate linguistic meanings among dialogue participants.

The project plan was based on three parallel tracks: theory development, empirical analysis of semantic coordination in dialogue, and implementation of computer models based on the theory. The focus of the project has been on theory development, which is motivated by the fact that the TTR formalism has proven very useful in this context and has generated substantial research results relevant to linguistics, language technology and philosophy of language.

The main difference between the project as it was carried out in relation to the plan is therefore that the empirical and implementation tracks have received less attention than the theory development track. The empirical part has been limited to some initial corpus studies and a compilation of examples from the literature on first language acquisition, which have then been analysed in detail in connection to theory development and associated publications. Some implementation work has been done with regard to certain syntactic aspects of corrective feedback (se below for explanation of this term). A more ambitious implementation project has also been started independently, with the goal of creating a test-bed for future implementations of semantic coordination in dialogue systems.

The three most important results of the project

1. A formal semantics for the coordination of meaning
The idea of using type theory (TTR) to describe structured and dynamic meanings of linguistic expressions has been successful and the theory has developed substantially during the project.

In formal semantics, the connection between language and the world has traditionally been analysed using model theoretic semantics, where meaning is analysed as essentially extensional (Larsson 2011a). For example, the meaning of "blue" is the set of all blue things. The problem with this view is that it ignores several essential aspects of linguistic meaning, e.g. that linguistic meaning is (1) context-dependent and connected to perception, and (2) "open" and negotiable. It is not always possible to say what the extension of a linguistic expression is in an abstract and context-independent way. The theory developed in the present project instead focuses on the intensional aspects of linguistic meanings, i.e. the "methods" used by speakers of a language to relate linguistic expressions to the world (and other linguistic expressions).

To represent meanings of linguistic expressions, we have used TTR (Type Theory with Records). TTR integrates logical techniques suvh as binding and lambda calculus with objects similar to the feature structures used in formal syntax. This provides more structure than is available e.g. in unification-based grammatical systems. These properties are important for defining semantic updates. The logical aspects are important to relate our semantics to the model- and proof-theoretical tradition associated with formal semantics.

Cooper (2012) presents the latest published version of the TTR and the relationship to the theory of semantic coordination which has been developed in this project. In particular, it addressed the relationship to "frame semantics" (ramsemantik) as developed by Fillmore that has led to the development of FrameNet (https://framenet.icsi.berkeley.edu). It is suggested that record types are an appropriate formalization of frames and play an important role in an account of how meaning can be created or modified during the course of a dialogue. The main features of this argument was presented in a conference paper, Cooper (2010).

2. Dialogue moves for semantic coordination in first language acquisition
In the case of pragmatic theory, a preliminary taxonomy of conversational strategies for meaning coordination in first language acquisition has been developed and formalized (Larsson & Cooper 2009). This distinguishes a number of recurring patterns of interaction such as corrective feedback:

A: Nice bear
B: Yes, it's a nice panda

Here, B provides positive feedback to part of A's utterance, but also corrects A's use of the word "bear" and proposes "panda" as the correct expression. Based on such patterns of interaction, and relationships between the utterances involved, dialogue participants can generate hypotheses about the proper use and meaning of linguistic expressions.

In addition to corrective feedback, other dialogue strategies have been studied: explicit verbal definitions (Cooper & Larsson 2009), ostensive definitions (e.g., pointing to a colour sample and saying "blue") (Larsson, 2011b), and accommodation, i.e. silent adaptation to an observed innovative use of an expression (Larsson 2010).

3. Multidimensional meaning representations
In formal semantics, the analysis and representation of linguistic meaning has previously been restricted to compositional and extensional (see above) aspects of meaning. In the present project, a more complex view of linguistic meaning has been incorporated in a formal semantic theory, where not only compositional but also ontological and perceptual meaning is included. In addition, the theory can be used to provide detailed descriptions of gradual linguistic change.

Cooper & Larsson (2009) presents a number of example dialogues from the literature on first language acquisition, where children appear to learn word meanings from corrective feedback. A distinction is made between compositional and ontological meaning. Compositional meaning representations can be used to derive the meaning of a complex expression (such as sentence) from the meanings of its constituents. Ontological meaning relates an expression to other expressions of the language; for example, that a glove is a kind of garment. The article also suggests a way to make this distinction between compositional and ontological meaning in TTR.

Larsson (2011b) extends the theory to also deal with perceptual meaning; see below.

New research questions generated by the project

• How do compositional, ontological and perceptual aspects of meaning interact, for example with respect to compositionality?
• How can the theory be extended to handle connotations and functional aspects of meaning?
• How can the TTR theory of semantic coordination be implemented in a dialogue system?

The project's two most important publications

Cooper & Larsson (2009): Lays the foundation for the TTR-based theory of coordination of multi-dimensional meaning, partly through a description of a number of key dialogue strategies and partly through a formal semantics that allows gradual meaning changes, and finally, by linking together dialogue strategies and formal semantics.

Larsson (2011b): Extends the theory of multi-dimensional meaning to also handle perceptual meaning. The basic idea is that linguistic meaning relates to perception in the sense that (some) words have a perceptual meaning component that consists of a so-called classifier that takes perceptual data (signals from sensory organs) as input and determines whether the data should be classified as an instance of the meaning of the expression or not. This is illustrated with a simple "game" in which an "teacher" teaches a "pupil" the meaning of the word "right" by placing objects within a frame and either say "right" or "not right". By classifying the object's position as "right" or not, and comparing this to what the teacher says, the student can detect situations in which the classifier gives the "wrong" results. In such situations, the student can modify its classifier, which is an example of semantic coordination of perceptual meaning of dialogue. Important progress reported in this article includes:

- Language and perception are related in that compositional, ontological and perceptual aspects of meaning are represented in a common framework

- Perceptual meaning is seen as classifiers of sensory input, which allows for the linking of formal semantics and the extensive body of research on statistical classification

- The theory fits well into an established theoretical framework for implementation of dialogue systems based on information states and how they are updated

Other types of communication of project results

Invited lectures and seminars:

- Robin Cooper, Staffan Larsson: Towards a formal view of Corrective Feedback. Dialogue Matters Final Workshop, June 2, 2008,

- Robin Cooper, Staffan Larsson: Towards a formal view of ordination and learning in Dialogue. Coordination of Agents Workshop, November 4th-5th 2008, London.

- Staffan Larsson: Do Dialogues Have the Content? Workshop on Dialogue and Inference: Queen Mary University of London on Wednesday 27th October.

- Staffan Larsson and Robin Cooper: Semantic Coordination in Dialogue. Nordling PhD course on topics in Language & Cognition. Hosted by the Faculty of Humanities, NTNU - Trondheim, Norway September 6-9th, 2010. http://www.hf.ntnu.no/sprakprosesseringslab/ResearchSchool/NordLing.html

Project Home Page (also available in English): http://www.flov.gu.se/forskning/forskningsprojekt/allman-sprakvetenskap/semantisk-koordinering/