Integrating pragmatic insights with HPSG: an exploration of theoretical and methodological issues

A one-year project, funded by AHRC, which started January 1, 2006.
Marina Terkourafi is a full-time researcher on the project, Ann Copestake is the Principal Investigator. We are also collaborating with Mayumi Masuko and Valia Kordoni.


Marina Terkourafi. What use is ‘what is said’? First Workshop on Utterance Interpretation and Cognitive Models, Université Libre de Bruxelles,  2006.

Ann Copestake and Marina Terkourafi. Conventional speech act formulae: from corpus findings to formalization. Constraints in Discourse, NUI Maynooth, Ireland. 2006.

Ann Copestake and Marina Terkourafi. Conventional speech act formulae in HPSG. International Conference on Head-Driven Phrase Structure Grammar, Varna, Bulgaria. 2006.

Project summary

This project contributes to linguistic research on language convention. Previous work has shown that a variety of forms are used by speakers to make offers, requests etc, and that the preferred form depends on various aspects of the context. We will provide an account of this which is integrated with formal approaches to syntax and semantics. To do this, we will look at a range of languages (English, Greek and Japanese) and investigate methods for acquiring enough natural data to test our hypotheses. In the long-term, this research will benefit computer systems which generate and interpret dialogue.

Project Background: Modelling politeness in a Greek HPSG

The origins of the current work lie in a small project which was funded by a British Academy Small Research Grant `Modelling politeness in a Greek HPSG'. Funding for this project enabled Marina Terkourafi to make several visits to Cambridge to collaborate with Ann Copestake and Aline Villavicencio and to Saarbrücken where she collaborated with Valia Kordoni (see Terkourafi and Villavicencio, 2003).

The BA project was envisaged as a preliminary study that might lead to a larger scale project on modelling politeness. However, during its course, it became apparent that there were many aspects of the work that we had not foreseen and that it had much wider ramifications than we had envisaged. Our initial starting point for collaboration was our shared interest in conventionality in language, but establishing common ground was not straightforward and there was very little prior research to draw on. Working on politeness in HPSG required us to consider how current pragmatics research in general could interact with formal syntax and semantics, which in turn has led to questions of differences in methodology and formalism which we believe are highly significant. The two research areas make such different assumptions that creating a real bridge between them is a significant challenge.

While the new project is not in itself computational, we think that computational linguistics may help provide the bridge between HPSG and pragmatics. Computational linguists often use a stochastic approach to determining dialogue acts / conversational moves. Stochastic/corpus-driven methods are now also incorporated within large linguistically-motivated HPSGs to handle ambiguity, and though the objectives for doing this are normally stated in engineering terms, we speculate that there is theoretical relevance too.

Terkourafi, Marina & Villavicencio, Aline (2003) 'Toward a formalisation of speech act functions of questions in conversation.' In: Bernardi, Rafaella and Michael Moortgat (ed.) Questions and Answers: Theoretical and applied perspectives. Utrecht Institute of Linguistics OTS. 108-119. Available online at:

Detailed project description

Research questions

  1. How can contextual (extra-linguistic) information be integrated with formal syntactic and semantic analyses? Specifically, can hypotheses regarding dimensions of pragmatic variation in Cypriot-Greek (Terkourafi, 2002) be formalised within Head-driven Phrase Structure Grammar (HPSG: Pollard and Sag 1994) and if so, what are the theoretical implications for the framework?
  2. Can the formalised accounts be extended to other languages, specifically English and Japanese?
  3. What is the role of corpus studies as opposed to linguistic intuition in this enterprise? Can existing corpora substitute for corpora gathered specifically for the purpose of addressing context-dependent variation? Can alternative data gathering techniques be used?
We do not expect to fully answer all these questions during a one-year project. But since they are very closely interrelated (see 'research context'), it would be inappropriate to investigate one aspect in detail while neglecting the others. Rather, we envisage a three-pronged approach, tackling formalisation, cross-linguistic aspects and methodology in parallel, with the aim of this project being primarily to clarify the questions and to provide a firmer basis for further research.

Aims and objectives

We will explore the potential for cross-fertilisation of ideas, both theoretically and methodologically, between the disciplines of pragmatics and formal grammatical frameworks.

Recent years have witnessed a steady growth in research on linguistic pragmatics, both from an intra-cultural and a cross-cultural perspective. Several small-scale studies have attempted to tap into native speakers' or learners' intuitions regarding the impact of gender, age, ethnicity, setting, and so on, on language use, producing a wealth of data from majority and minority languages, and from various social groupings within broader communities (see, e.g. recent issues of the International Journal of the Sociology of Language, and the Journal of Pragmatics). For the most part, these findings are descriptive and have had little impact on research conducted within various grammatical frameworks. Sign-based approaches, such as HPSG, aim to integrate pragmatics with other levels of analysis, but although some previous research (e.g., Green 1987, Paolillo 2000, Pollard and Sag 1994, Bender 2001, forthcoming) has noted points of contact between the two disciplines, it has not explored the precise nature of this interaction, or the full implications of addressing pragmatics within a lexicalist framework.

Pragmatic findings pose new questions for lexicalist formalisms. Accommodating these findings will enhance the scope of the analyses but may challenge their formal basis. Pragmatic analyses, on the other hand, need to move beyond description, and advance testable hypotheses. At the same time, they must develop rigorous methodologies facilitating cross-linguistic and cross-situational comparison of their findings, such that the large amounts of data collected empirically can be fully exploited.

By highlighting common concerns between the two disciplines, we aim to focus the questions they ask, as well as help formulate their results, in ways that enhance their cross-disciplinary applicability/relevance. We will refine the research questions of the previous section to clarify their interactions and specify possible alternative approaches.

Research context

Frequency data from a corpus of over 2,000 recorded spontaneous Cypriot-Greek exchanges (Terkourafi, 2002) shows that particular verbal formulae (verb-forms inflected for mood, number/person, and rendered with a particular accent and intonation) preferentially realise particular speech acts (offers or requests). However, these preferences are not stable across contexts; rather, extra-linguistic features (interlocutors' gender, age and social class, their relationship, setting, sequential placement of the utterance), affect the choice of formulae. In prior research (conducted under British Academy grant, 'Modelling Politeness in a Greek HPSG') we related compositionally constructed semantics incorporating message types (Ginzburg and Sag, 2000) to structures describing illocutionary force. Two approaches were explored. Firstly, lexical defaults were drawn upon as a way of building the observed preferences into the grammar as automatic yet defeasible inferences (Terkourafi and Villavicencio 2003). However, this approach only considers interpretation and does not adequately model variation. In subsequent unpublished work, we started to explore the potential of a probability-based account for modelling degrees of conventionalisation of speech-act formulae. This is a radical step to take, opening up a new set of research issues.

In principle, quantitatively-motivated empirical pragmatic analyses and formally-oriented work within HPSG share several theoretical and methodological concerns. Encoding contextual features within the sign raises the question of the number and precise identity of the features needed to capture effects such as conventionalisation. However, we now believe that this cannot be settled independently of the general question of formalising pragmatic constraints in HPSG. Developing the pragmatic component of the framework involves first clarifying the nature of pragmatic features in HPSG, where they may be specified, what form their values might assume and how such values may be specified, how 'soft' the corresponding constraints may be, when/how they may be overridden, and how they interact with other dimensions (phonological, syntactic, semantic) of the sign during interpretation/production. Languages which grammaticise the consequences of pragmatic processes, such as Korean and Japanese, have constituted prime points of departure for investigating related issues (Pollard and Sag 1994, Siegel 2000, Engdahl, to appear). A theoretically-compelling account of the Cypriot-Greek data must also address the wider cross-linguistic questions.

These theoretical questions have to be related to methodology. What constitutes appropriate data for addressing pragmatic questions remains an open issue. Pragmatic intuitions are often more graded than syntactic and semantic ones, though not less real, as shown by native speakers' spontaneous judgements about (and, reactions to) the appropriateness of various utterances in different situations. While linguistic introspection is a resource whose value we acknowledge, we aim to use this only complementarily, i.e. to check the viability of the hypotheses we formulate based on the observable data.

Exploring the question of which methodological tools are appropriate for investigating speakers' communicative competence contributes to the ongoing debate about the nature and role of data in theoretical linguistics (e.g., Studies in Language 28:3; forthcoming special issue of Lingua). Lexicalist approaches, including HPSG, prioritise data from introspection, as this ensures comprehensive coverage of the (syntactico-semantic) phenomena investigated and of the range of grammatical alternatives. While comprehensiveness is certainly a concern for corpus-based analyses (including Terkourafi's (2002) data), using data from introspection is open to scrutiny as to its representativeness.

There is limited research on the interface of lexicalist grammars with pragmatics. Besides the HPSG work cited above, Asher and Lascarides (2003: 304ff) is relevant to our approach but is not empirically grounded. From a different perspective, some computational work (e.g., Jurafsky 2004; Carletta et al. 1997) has addressed the question of illocutionary act recognition (though different terminology is used) but does not relate illocutionary acts to general grammars. Stochastic work within HPSG exists, but is viewed as an engineering approach to cut down lexical and syntactic ambiguity and not as a formally interesting technique.

In the medium-term, probably as part of a follow-on project, we hope to provide results that might be incorporated into attempts to provide a common framework for implementing HPSG (the Grammar Matrix: Bender et al, 2002), which is distributed as Open Source. In the long-term, our research could have important implications for the development of computational dialogue systems, particularly in generating more natural dialogue.

Research Methods

Although our primary approach to formalisation is theoretical, we will also use existing computational implementations of constraint-based grammars, specifically the LKB system (Copestake, 2002) and investigate whether it is possible to adapt the existing stochastic techniques to our current purposes, to allow us to run simulations to check the probabilistic approach. This would enable large amounts of data to be checked automatically and enforce a level of precision. As part of our investigation of cross-linguistic aspects, we will also exploit an existing computational grammar of Greek running on the LKB system (Kordoni and Neu 2003). Professor Masuko (Waseda University) has agreed to advise us on Japanese.

To investigate to what extent existing resources can be used, we will experiment with a variety of English corpora (the Computer Laboratory has licences to the main ones, including those available via the Linguistic Data Consortium). To gain information on the full range of grammatical alternatives, we will conduct pilot studies on combining corpus-data with elicited experimental and introspective data via purpose-designed experiments enabling targeted data acquisition. In this way, we aim to contribute to developing alternative methodologies for collecting data about communicative competence.


Asher, Nicholas and Lascarides, Alex (2003) Logics of conversation. Cambridge: Cambridge University Press.
Bender, Emily (2001) Syntactic variation and linguistic competence. Ph.D. thesis, Stanford University.
Bender, Emily (forthcoming) Variation and formal theories of grammar: HPSG. Elsevier Encyclopedia of Language and Linguistics. Oxford: Elsevier Science.
Bender, Emily M., Dan Flickinger and Stephan Oepen. 2002. The Grammar Matrix: An Open-Source Starter-Kit for the Rapid Development of Cross-Linguistically Consistent Broad-Coverage Precision Grammars. In Carroll, John, Nelleke Oostdijk, and Richard Sutcliffe, eds. Proceedings of the Workshop on Grammar Engineering and Evaluation at the 19th International Conference on Computational Linguistics. Taipei, Taiwan. pp. 8-14.
Borsley, Robert (in press) Introduction to special issue of Lingua.
Carletta, Jean et al. (1997) The reliability of a dialogue structure coding scheme. Computational Linguistics 3: 13-31.
Copestake, Ann (2002) Implementing typed feature structure grammars. Stanford: CSLI Publications.
Engdahl, Elisabet (to appear) Integrating pragmatics into the grammar. In: Mereu, Lunella (ed.) Boundaries of morphology and syntax. Amsterdam: John Benjamins. Ginzburg, Jonathan, and Sag, Ivan (2000) Interrogative investigations. Stanford: CSLI Publications.
Green, Georgia (1996) The Structure of CONTEXT. In: Yoon, J. (ed.) Proceedings of the 5th Annual Meeting of the Formal Linguistics Society of the Midwest. Jurafsky, Daniel (2004) Pragmatics and computational linguistics. In: Horn, Laurence and Ward, Gregory (eds.) The Handbook of Pragmatics. Oxford: Blackwell. 578-604.
Kordoni, Valia and Julia Neu (2003) Deep grammar development for Modern Greek. In: Bender, Emily, et al. (eds.) Proceedings of the Workshop on ideas and strategies for multilingual grammar engineering. ESSLLI 2003.
Masuko, Mayumi (1992). Referential and honorific expressions in Japanese: Towards a formal approach. Unpublished PhD dissertation, University of Cambridge.
Paolillo, John (2000) Formalizing formality. Journal of Linguistics 36: 215-259.
Pollard, Carl and Sag, Ivan (1994) Head-driven Phrase Structure Grammar. Chicago University Press.
Siegel, Melanie (2000) Japanese Honorification in an HPSG framework. In: Ikeya, A. and Kawamori, M. (eds.) Proceedings of the 14th Pacific Asia Conference on Language, Information and Computation. Waseda University International Conference Center, Tokyo. Logico-Linguistic Society of Japan, 289-300.
Terkourafi, Marina (2002) Politeness and formulaicity. Journal of Greek Linguistics 3: 179-201.
Terkourafi, Marina and Villavicencio, Aline (2003) 'Toward a formalisation of speech act functions of questions in conversation'. In: Bernardi, Rafaella and Michael Moortgat (ed.) Questions and Answers: Theoretical and applied perspectives. Utrecht Institute of Linguistics OTS. 108-119.