Detailseite
Projekt Druckansicht

Bayes'sche Ansätze für eine präferenzenbasierte Antwortgenerierung im Dialog

Fachliche Zuordnung Allgemeine und Vergleichende Sprachwissenschaft, Experimentelle Linguistik, Typologie, Außereuropäische Sprachen
Förderung Förderung von 2013 bis 2021
Projektkennung Deutsche Forschungsgemeinschaft (DFG) - Projektnummer 232196050
 
Erstellungsjahr 2020

Zusammenfassung der Projektergebnisse

The projects were motivated by our view that game-theoretic models of communication benefit from a more elaborated handling of the speaker’s tasks, and answer generation in a dialogue system will benefit from game-theoretic models. A prospective tenant asks questions about apartments the system as a real estate agent is offering. The system answers these questions directly and indirectly, respectively. For example, the question: Does the apartment have a garden? could be answered by the system directly, but it is also able to generate the exemplary answers It has a balcony or If you are interested in growing flowers, it has a balcony. Furthermore, if the customer asks Is there a supermarket? the system can choose between answers that convey vague or exact distances as, for example, There is a supermarket fewer than 200 / 170 / more than 150 m away. The underlying research question is how the system can make optimal strategic use of such indirect answers. Most game theoretic models of communication are based on signaling games, i.e. games in which first the speaker chooses an utterance from a range of alternatives, and the hearer then picks an interpretation. Then the game ends. We considered this inadequate as a model of sales scenarios. If the customer asks Does the house have a garden?, the dialogue does not simply end when the system answers No. If the customer wants a place for growing flowers, it is likely that s/he will next ask for a balcony. The system can anticipate this question and answer to the first question Well, it has a balcony instead of No. Indirect answers come with extra costs, because they are longer, but also save costs as they lead to overall shorter dialogues. This scenario can be described as a strategic situation in which the system has to balance costs and savings of indirect answers. We had to develop an explicit game-theoretic model of this situation, learn probabilities of possible follow up questions from data, determine a threshold that decides when indirect answers are profitable, and evaluate the system that was based this model. This approach was very successful and culminated in two noteworthy publications. The follow-up project extended the linguistic domain. We integrated speech act conditionals and modified numerals as (part of) answers to the dialogue system, which required a corresponding extension of the non-linguistic domain and the assumed requirements. As before, the probabilistically motivated assumptions have been underpinned by experimental studies carried out via the crowdsourcing platforms MTurk and Prolific, and the system behavior had been evaluated as before. In general, we found that indirect answers support the user’s decision for or against an apartment offered by the system as sales agent, since the user’s requirements, expressed by the various forms of indirect answer we have examined in our projects, reduce the number of questions that must be raised in order to come to a decision. From a theoretical perspective, we were able to demonstrate the benefits of using concepts from game theory for computational models of answer generation.

Projektbezogene Publikationen (Auswahl)

 
 

Zusatzinformationen

Textvergrößerung und Kontrastanpassung