Intentions in dialogue modelling

Recent psycholinguistic evidence appears to show that interlocutors tend to exhibit “egocentric” behaviour as a default strategy in communication (see work by Keysar, Barr et al, Horton & Gerrig, but cf. Clark and Schaefer 1989; Brennan & Clark 1996; Hanna et al 2003). On the other hand, there is also good evidence for effective coordination between interlocutors in joint tasks mediated by language (e.g. Clark 1996 a.o.). Achievement of such coordination has been assumed to be based on some notion of strategic common ground or mutual knowledge computation on the part of the participants in dialogue. Models adopting this point of view standardly rely on some notion of intention-recognition as the operative mechanism effecting coordination. In turn, this process can only be modelled explicitly via the metarepresentation of the interlocutor’s mental states in terms of their respective beliefs, intentions, plans and goals. However, the egocentrism displayed by subjects in perspective-adjustment/audience-design experiments (Keysar, Barr, Horton, Gerrig et al) mentioned above seems to imply that metarepresentation of the interlocutor’s mental state cannot be taken as a default strategy: “common ground is a functionally distinct process that belongs to an ‘adjustment’ stage of processing, but that imposes no constraint on production or comprehension per se” (Barr & Keysar 2006: 904) Language acquisition evidence also casts doubt on intention-based metarepresentational accounts of communication since children are able to communicate effectively before 'theory of mind' abilities have developed. Moreover, given the computational complexity of modelling recursive metarepresentations of other parties’ mental states in order to explain communication, mechanistic psycholinguistic models of coordination and dialogue attempt to base apparent common ground computation on more basic memory mechanisms not specifically dedicated to computing intentions (Pickering & Garrod 2004, Horton & Gerrig 2005). In pragmatics, models based on Gricean intentions have difficulty with accounting for egocentric effects as they generally presuppose some notion of “mutual knowledge” and intention-recognition as the basis for working out “speaker meaning”.

Two directions though diverge from these assumptions:
(a) empirical accounts of conversational mechanisms within the broad Conversation Analysis framework;
(b) Wittgensteinian philosophical accounts opposing the Gricean individualistic communication model
Conversation Analysis has long eschewed the role of intentions as the basis for effective communication. Here coordination in dialogue is seen as emergent without necessary calculation of common ground/speaker’s intention. Instead participants rely on the rich structure that the dialogue environment supplies and the feedback mechanisms integrated in conversational structure: “speakers recraft their utterances mid-stream, taking into account the responses, or more often the lack of them, from recipients . . . As a result, what is produced is actually a joint production, which can hardly correspond to the speaker’s own initial intention or goal.” (Goodwin 1979; 1981)
From this point of view, communication is accomplished via a trial-and-error process (see also Arundale 2008; Kempson et al 2009) and the content of “intentions” may emerge as a result of communication, instead of guiding it. Hence, in conversation, inference based on mutual knowledge/common ground is not a prerequisite for communication and speech act recognition in dialogue is shown to be derivative (Levinson 1983)

In branches of analytic philosophy influenced by the late Wittgenstein's ideas on language, current implementations of the Gricean programme in pragmatics are undermined from two surprisingly contrary directions:
(a) Millikan’s teleosemantic approach to language content (Millikan 1984)
(b) Brandom’s social-inferential account of communication (Brandom 1994).
Millikan argues against Gricean approaches to communication from a naturalistic point of view. In her account, language and communication are examined on the basis of phenomena studied by evolutionary biology. Linguistic understanding is seen on the analogue of direct perception. An account of meaning is given in terms of natural meaning and the function that linguistic devices have been selected to perform (their “survival value”). These functions are defined through what linguistic entities are supposed to do (not what they normally do or are disposed to do) so that “function”, in this sense, is a normative notion. Hence norms of language, ‘conventions’, are “uses” that had survival value. And “meaning simply is function”. But, from this point of view, function, i.e. meaning, does not depend upon speaker intentions. Speakers can behave purposefully in producing tokens of linguistic devices (as hearts and kidneys behave purposefully) but without representing hearers’ mental states or having intentions about hearers’ mental states. Similarly, hearers understand speech through direct perception of what the speech is about without necessary reflection on speaker intentions. Gricean mechanisms can be invoked but only as derivative or in cases of failure of the normal functioning of the mechanisms primarily involved in communication. In pragmatics, an intermediate position regarding the issue of Gricean intentions is advocated by Recanati (2004) who adopts Millikan’s account of understanding as direct perception for what he terms “primary pragmatic processes”, i.e., the pragmatic processes that are involved in the determination of the truth-conditional content of an underspecified linguistic signal.

Besides Millikan’s naturalistic viewpoint, Brandom’s rationalist programme in philosophy, and semantics/pragmatics in particular, also eschews the individualistic character of accounts of meaning espoused by the Griceans. Instead, he analyses meaning/intentionality in terms of linguistic social practices. Here again we find the Wittgensteinian concern to explain meaning in terms of use. Brandom’s general agenda involves explaining knowledge-that in terms of knowledge-how and his entire semantic theory is grounded in pragmatics. So meaning/content, but also beliefs and intentions, are to be accounted for in terms of the linguistic game of giving and asking for reasons. Brandom adopts an inferentialist account of meaning inspired by the Gentzen-style semantics of logical connectives in terms of introduction and elimination rules. Thus 'reasons' are cashed out in terms of the primitives (a) entitlements to claims, and (b) commitments following by claims. So the conceptual content of a sentence is given by its inferential role as a premise or conclusion within an exchange of reasons. Meaning then arises from the inferential articulation of the set of sentences instead of their representational properties. The inferential role of a sentence is determined by both other sentences (entitlements) that may be used as premises in inferences with the sentence in question as conclusion and by other sentences that may be inferred from it (commitments). Importantly, the inferential significance of a sentence also includes the perceptual circumstances that support its application as a 'true' reflection of the world and the practical consequences for action that emanate from its being adopted or claimed as 'true'. Thus, the rules/norms by which speakers use language (pragmatics) have explanatory primacy over representational content (semantics) which is seen as a kind of epiphenomenon. Brandom’s model of communication is then defined as a kind of scorekeeping. Participants in conversation are engaged in a discursive practice in which they keep score on one another by undertaking and attributing commitments and entitlements.

In the domain of computational semantics, Brandom’s ideas have been adopted by Kibble (2006a,b) among others (e.g. Matheson et al 2000; Walton & Krabbe 1995; Singh 1999). The reason for this innovation are the problems associated with private mental states such as intention and belief. Firstly, from a practical point of view, such states, according to Kibble, cannot be straightforwardly attributed to entities like software agents, corporate entities, institutions etc. (cf. Cohen & Levesque 1990a,b). Secondly, there are problems associated with the formal modelling of states like belief: (a) Omnidoxasticity: do rational agents believe all the logical consequences of their beliefs? (b) Inconsistency: should the fact that sometimes rational agents hold inconsistent beliefs lead to the conclusion that they believe any arbitrary proposition? (c) Many accounts ignore the inferential articulation of beliefs which becomes important in cases of belief revision, e.g., where a belief has to be retracted: should all consequential beliefs be retracted too and if not which ones should remain? (d) It is well-known that the notion of mutual belief/knowledge, which, in many accounts, is part of the definition of common ground, threatens an infinite regress: “John believes that Mary believes that John believes . . . p”. Thirdly, beliefs, goals and intentions underdetermine what rational agents will do in conversation: social obligations or conversational rules may in fact either displace beliefs or intentions as the motivation for agents’ behaviour or enter as an additional explanatory factor (Traum & Allen 1994). In fact, an account of the conversational mechanisms might be possible without necessary reference to guessing speaker intentions (as e.g. in Conversational Analysis accounts). Finally, another important problem mentioned by Kibble and others concerns the intersubjectivity of beliefs and intentions, i.e. the fact that such private mental states are not directly observable and available to the interlocutors. In contrast, Brandom’s account presents an inferentialist account of communication which seeks to replace mentalist notions such as belief with public, observable practical and propositional commitments. Under this view, commitment does not imply ‘belief’ in the usual sense. A speaker may publicly commit to something which she does not believe. And ‘intention’ can be seen as the undertaking of a practical commitment or a reliable disposition to respond differentially to the acknowledging of certain commitments. Kibble’s work, following Brandom’s “scorekeeping” model, proposes the outline of a model of dialogue involving updates of participants’ commitment stores which are records of each interlocutor’s propositional and practical commitments. These commitment stores are perspectival (as in Brandom’s account) in that there is a separate store for each participant according to their roles as speaker, addressee or hearer. Here is a summary of the basic features of Kibble’s model:
1. Each agent in a dialogue keeps a score of social commitments for all participants, including itself. Commitments can be classified into practical (commitments to act, corresponding to intentions in mentalistic accounts) and propositional or doxastic (commitments to justify an assertion, corresponding to beliefs).
2. Agents play one of three dynamically assigned roles at any given point in a dialogue: Speaker, Addressee, or Hearer who is not directly addressed.
3. For an agent to assert φ is to “acknowledge commitment” to φ; other agents may also attribute consequential commitments to the assertor.
4. A dialogue act constitutes an attempt to commit Addressee(s) to a proposition or a course of action.
5. Addressee’s options include accepting the proffered commitment, challenging it or requesting clarification. As we saw the intersubjectivity issue does not arise in such an account since commitments, unlike beliefs and intentions, are public and observable. The problem of mutual knowledge is partially resolved by distinguishing (a) ‘explicit’ overtly acknowledged commitments and (b) commitments following as the consequences of those (‘consequential’ commitments) which though do not have to be computed by the agent. Kibble also shows how some of the formal issues associated with the modelling of belief can be overcome in such an account. For example, the omnidoxasticity and inconsistency problems are resolved by distinguishing what ‘explicit’ commitments agents may undertake and what other agents would rationally attribute to them. Agents do not have to be committed to (or “believe”) all the consequential commitments of their claims even though they can be held accountable for those by interlocutors. Moreover, agents can adopt incompatible commitments even though they cannot be seen as entitled to them. The issue of ascribing mental states to non-human agents is also partially resolved: non-human interlocutors, like corporate entities or certain software agents, can be construed as able to undertake commitments and be held accountable for them.

But certain problems remain. As we just saw it is not absolutely clear what kind of responsibility as regards their commitments and the following of norms can be attributed to autonomous software agents. Also the issue of mutual knowledge/belief is not completely resolved as an infinite regress can arise at the level of ‘implicit’ commitment. One other more general problem with psychological interpretations of Brandom’s work is the holistic account of meaning that it explicitly necessitates. Brandom characterises his account of meaning as 'top down’, starting with the propositions expressed by whole sentences. It then seeks to understand the meanings of subsentential expressions through the contributions they make to the specification of such sentential contents. The reason for this stance is the priority of ‘judgment’ as the minimum unit in the inferential articulation of meaning. This is because, according to Brandom, only sentences/propositions can serve as premises and conclusions in inferences or be assigned illocutionary force. Besides the well-known problems of such holism, more acutely, the fact that it undermines the possibility of linguistic communication, there are also the issues of the incrementality of linguistic processing and compositionality of interpretation that need to be taken into account. It is a well-established psycholinguistic fact that language comprehension and production are incremental and the compositional articulation of meaning is an important factor for modelling online (subsentential) incrementality in the most natural way. And there is also evidence that such incrementality is essential for dialogue processing and the explanation of dialogue phenomena like split-utterances (Poesio & Rieser 2009; Gregoromichelaki et al 2009). In addition, ontogenenetic and phylogenetic accounts of language development and evolution seem to show that acquisition of individual concepts is prior to any notion of ‘judgment’ in the usual sense (e.g. children first learn individual words rather than being capable of formulating whole sentences). So any psychological model of language or communication will have to face these issues before adopting Brandom’s proposal as a replacement for an account of communication that rests on mentalistic notions like beliefs and intentions.

Poesio and Rieser in Dialogue and Discourse journal

The first article of the free access Dialogue and Discourse journal has appeared:

"Completions, coordination, and alignment in dialogue" by Massimo Poesio and Hannes Rieser

They propose a treatment of collaborative completions (a species of split utterances in dialogue) within the PTT framework, incorporating a theory of incremental utterance interpretation and an account of grounding. The analysis employs Lexical Tree Adjoining Grammar (LTAG), (Compositional) DRT and Situation Semantics to analyse the structure and semantics of these constructions. In addition, the generation and processing of collaborative completions is grounded on an intentional account, as well as a preliminary formulation within Pickering and Garrod’s alignment theory.

SemDial satellite workshop

SEMDIAL2010 will be the 14th workshop in the SEMDIAL series and it will be organized at the Institute of Psychology (Chair of Logic and Cognitive Science), Adam Mickiewicz University (AMU), Poznań.
Dates: 16-18 June, 2010.

Participants may also be interested in the "Intentions in Dialogue" workshop, sponsored by DYNDIAL, being held in Poznan on Tuesday 15th June (at the same location).

Modelling the achievement of effective coordination between interlocutors is often assumed to rely on some notion of strategic common ground or mutual knowledge computation. However, some recent psycholinguistic and language acquisition data seem to undermine this assumption, and computational models of dialogue often exclude high-order meta-representations of other parties' beliefs or intentions. This workshop will explore various points of views pertaining to this issue and its resolution.