A distinguishing component of STAC's research methodology for advancing our understanding of strategic conversation is to interleave theoretical work and analysis with empirical evaluation and validation using a dialogue manager of a working dialogue system. Accordingly, we have developed dialogue systems to provide corpora for studying strategic conversation, and we have annotated data from these corpora to build models of strategic conversational agents based on the data.
STAC is divided into four sub-projects, managed by three different research teams at: IRIT, CNRS/University of Toulouse in France, the School of Informatics at the University of Edinburgh in Scotland, and the School of Mathematical & Computer Sciences at Heriot-Watt University in Scotland. Two sub-projects relate to the linguistic study of strategic conversations while the remaining two concentrate on the development of agents that can test our linguistic hypotheses within a particular strategic domain.
STAC is divided into four sub-projects, managed by three different research teams at: IRIT, CNRS/University of Toulouse in France, the School of Informatics at the University of Edinburgh in Scotland, and the School of Mathematical & Computer Sciences at Heriot-Watt University in Scotland. Two sub-projects relate to the linguistic study of strategic conversations while the remaining two concentrate on the development of agents that can test our linguistic hypotheses within a particular strategic domain.
Theoretical linguistic studies on strategic communication (Toulouse)
This subproject of STAC begins with the observation that conversations often involve an element of planning and calculation of how best one can achieve one’s interests in light of what one knows about the interests of other conversational participants. To make our model of conversation fully general, we are interested in how conversations proceed in a setting in which the interests of the dialogue agents are opposed. Because we are interested in conversational situations where we cannot assume mutual underlying interests, standard game theoretic analyses do not apply and we have developed a new model of conversation. Our model exploits extensive games that resemble mathematical games like Banach Mazur games, involving two conversational agents and also a "jury" that serves as an evaluation of whether discourse moves satisfy or do not satisfy the conversationalist's goals. Conversations are analyzed as sequences of discourse moves, which are defined by the underlying linguistic theory of discourse structure and content, SDRT. Although our model grows out of a study of noncooperative conversations, it also applies to cooperative contexts so that Grice's cooperative understanding of conversation, for example, is treated as a special case of STAC's more general view of conversation as a strategic activity. We have also investigated notions of consistency, coherence and credibility in this setting, as well as the notion of rationality in large games and notions of discourse politeness.
This subproject of STAC begins with the observation that conversations often involve an element of planning and calculation of how best one can achieve one’s interests in light of what one knows about the interests of other conversational participants. To make our model of conversation fully general, we are interested in how conversations proceed in a setting in which the interests of the dialogue agents are opposed. Because we are interested in conversational situations where we cannot assume mutual underlying interests, standard game theoretic analyses do not apply and we have developed a new model of conversation. Our model exploits extensive games that resemble mathematical games like Banach Mazur games, involving two conversational agents and also a "jury" that serves as an evaluation of whether discourse moves satisfy or do not satisfy the conversationalist's goals. Conversations are analyzed as sequences of discourse moves, which are defined by the underlying linguistic theory of discourse structure and content, SDRT. Although our model grows out of a study of noncooperative conversations, it also applies to cooperative contexts so that Grice's cooperative understanding of conversation, for example, is treated as a special case of STAC's more general view of conversation as a strategic activity. We have also investigated notions of consistency, coherence and credibility in this setting, as well as the notion of rationality in large games and notions of discourse politeness.
Empirical linguistic studies on conversation (Toulouse)
STAC involves detailed investigations and experiments involving the automatic acquisition of discourse structure of both texts and dialogues. The experiments involved machine learning from conversations and texts annotated with discourse structure. STAC has undertaken an extensive annotation campaign and has constructed an annotated corpus of over 930 negotiation dialogues. The dialogues are in fact chat sessions from games of Settlers of Catan played on-line. The corpus of games was collected by the STAC members at The University of Edinburgh (http://settlers.inf.ed.ac.uk/socl/soclinfo.php). The annotation model is that given by SDRT, and STAC has served to verify SDRT principles and coverage for multi party conversation. (Click here for the annotation manual.) Additionally, STAC has proposed translation principles for different annotation models for discourse structure, and thus we also are investigating the automatic acquisition of discourse structure on other corpora, such as the RST Tree Bank, as well as the acquisition of features relevant to discourse parsing from resources like the Penn Discourse Tree Bank. |
Symbolic conversational agents (Edinburgh)
STAC's corpus of negotiation dialogues comes from sessions played on-line of the Settlers of Catan game. Edinburgh not only has collected these on-line games but also has worked on an automatic agent that plays the game itself. This agent uses symbolic rules and heuristics to calculate moves it deems in its own interest. The goal of this part of the project is to endow this agent with natural language capacities so that it can interact with human players. We can then evaluate various conversational strategies (e.g., with respect to how effective the conversational agent is at bargaining or at making trades). We also are investigating how various non-linguistic strategies for playing the game interact with conversational strategies. Thus this project will exploit the NLP tools for multi-party dialogue developed at Toulouse.
STAC's corpus of negotiation dialogues comes from sessions played on-line of the Settlers of Catan game. Edinburgh not only has collected these on-line games but also has worked on an automatic agent that plays the game itself. This agent uses symbolic rules and heuristics to calculate moves it deems in its own interest. The goal of this part of the project is to endow this agent with natural language capacities so that it can interact with human players. We can then evaluate various conversational strategies (e.g., with respect to how effective the conversational agent is at bargaining or at making trades). We also are investigating how various non-linguistic strategies for playing the game interact with conversational strategies. Thus this project will exploit the NLP tools for multi-party dialogue developed at Toulouse.
Screenshot of Glozz
Reinforcement Learning conversational agents (Heriot-Watt)
Parallel to the development of a symbolic game playing agent with conversational capacities, is the development of a game playing agent that uses reinforcement learning to determine moves that are in its interest. This project will also exploit the NLP tools for multi-party dialogue developed at Toulouse and evaluate various conversational strategies
Parallel to the development of a symbolic game playing agent with conversational capacities, is the development of a game playing agent that uses reinforcement learning to determine moves that are in its interest. This project will also exploit the NLP tools for multi-party dialogue developed at Toulouse and evaluate various conversational strategies