Experimental toolkit for the manipulation of text-based dialogue
Interaction, Media and Communication group
Dialogue is the primary site of language use -- languages are first learned and first deployed in interaction. There is also a growing body of experimental evidence which shows that phonological, lexical, semantic and syntactic choices can be more strongly influenced by conversational context than by individual cognitive factors and that dialogue-specific mechanisms drive language change.
Despite its importance, our understanding of language processing in dialogue is much less advanced that our understanding of intra-individual processing. One key reason for this is practical; dialogue is inherently interactive and contextualised and this makes controlled experiments difficult. Current techniques involve relatively coarse interventions such as manipulations of task difficulty, changes in communication medium (e.g. audio +/- video) and control of participant's ability to interact -- e.g. by limiting access to an interaction, by using confederates who make scripted contributions or by controlling patterns of interaction (who talks to whom).
These techniques lack the kind of fine-grained experimental control found in studies of intra-individual language processing. This has held back progress in the analysis of dialogue. In this paper we present a new text-chat based experimental platform that addresses these problems (anon). This work exploits two developments. First the fact that text-based dialogue --chat rooms, instant messaging-- is now a ubiquitous mode of communication. Second the increasing availability of natural language processing techniques that can produce semantic and syntactic analyses of text at speeds that make real-time interventions feasible.
| Turn no. | Participant A's view | Participant B's view | |
|---|---|---|---|
| 1 | A: So we agree? | Target turn | A: So we agree? |
| 2 | B: agree? | Artificial probe | |
| 3 | A: yeah, to chuck out Susie Derkins | Response blocked by server | |
| 4 | B: oh ok | "Spoof" acknowledgement | |
| 5 | B: so, umm how about it? | Dialogue resumes | B: so, umm how about it? |
The basic strategy involves participants communicating using a synchronous text chat tool. Instead of passing turns directly to the appropriate chat clients, each turn is routed via a server. This is configured to detect, for example, target words or phrases and uses the Stanford parser to produce a syntactic analysis of each turn in real-time. This information can then be used to trigger specific experimental interventions. For example, a `spoof' clarification request that appears to originate from another participant. The recipient responds to the clarificiation, the server produces an acknowledgement and then subsequent turns are transmitted as normal. The `spoof' sub-dialogue is not seen by the other participant and we have shown this can be done without disruption to the dialogue or detection by the participants.
- The platform, which is distributed free under a GPL license, provides support for:
- systematic substitution of non-words, synonyms, hyponyms, hypernyms to investigate conceptual co-ordination,
- manipulation of turn origin, turn sequencing and dialogue history to investigate the procedural organisation of conversation,
- manipulations of semantic and syntactic constitutents to investigate co-ordination of linguistic structure.
This platform provides a powerful, task-independent, fine-grained technique for experimenting with dialogue. Interventions can be made in ways that are sensitive to the unfolding conversational context. Unlike confederate techniques the interventions can be made `blind' and inserted directly into an unscripted exchange. The new levels of experimental control this approach provides enable us to inspect dialogue mechanisms at a much higher resolution than has previously been possible.