Extraction of Semantic Features from Transaction Dialogues

摘要

Existing question and answering systems are mostly limited to answer queries in isolation, although the system's knowledge in determining whether a question is relevant to the previous interaction context is very important as supported in dialogue systems. Nonetheless, the problem with interpreting the input queries in both systems remains universal because utterances are typically short, single-sentenced, and may even be grammatically incorrect. This research presents a set of decision trees that form the rules in extracting semantic features encoded in dialogue utterances based on the Information Structure Theory. The extraction process is illustrated using a transaction dialogue corpus for a theater reservation system called SCHISMA. Because the basis of extraction is the informativeness in utterance within the domain knowledge, the proposed method may be used to improve research in QA systems by taking account the previous interaction context.