Classification-and-Ranking Architecture Based on Intentions for Response Generation Systems

Existing response generation accounts only concern with generation of words into sentences, either by means of grammar or statistical distribution. While the resulting utterance may be inarguably sophisticated, the impact may be not as forceful. We believe that the design for response generation requires more than grammar rules or some statistical distributions, but more intuitive in the sense that the response robustly satisfies the intention of input utterance. At the same time the response must maintain coherence and relevance, regardless of the surface presentation. This means that response generation is constrained by the content of intentions, rather than the lexicons and grammar. Statistical techniques, mainly the over generation-and-ranking architecture works well in written language where sentence is the basic unit. However, in spoken language where utterance is the basic unit, the disadvantage becomes critical as spoken language also render intentions, hence short strings may be of equivalent impact. The bias towards shortstrings during ranking is the very limitation of this approach hence leading to our proposed intention-based classification-and-ranking architecture. In this architecture, response is deliberately chosen from dialogue corpus rather than wholly generated, such that it allows short ungrammatical utterances as long as they satisfy the intended meaning of input utterance. The architecture employs two basic components, which is a Bayesian classifier to classify user utterances into response classes based on their pragmatic interpretations, and an Entropic ranker that scores the candidate response utterances according to the semantic content relevant to the user utterance. The high-level, pragmatic knowledge in user utterances are used as features in Bayesian classification to constrain response utterance according to their contextual contributions, therefore, guiding our Maximum Entropy ranking process to find one single response utterance that is most relevant to the input utterance. The proposed architecture is tested on a mixed-initiative, transaction dialogue corpus of 64 conversations in theater information and reservation system. We measure the output of the intention-based response generation based on coherence of the response against the input utterance in the test set. We also tested the architecture on the second body of corpus in emergency planning to warrant the portability of architecture to cross domains. In the essence, intention-based response generation performs better as compared to surface generation because features used in the architecture extend well into pragmatics, beyond the linguistic forms and semantic interpretations.

Text FSKTM_2008_1 IR.pdf Download (2MB)

Actions (login required)

View Item