Augmenting Conversational Characters with Generated Question-Answer Pairs
Elnaz Nouri, Ron Artstein, Anton Leuski, David R. Traum · National Conference on Artificial Intelligence · 2011
We take a conversational character trained on a set of linked question-answer pairs authored by hand, and augment its training data by adding sets of question-answer pairs which are generated automatically from texts on different topics. The augmented characters can answer questions about the new topics, at the cost of some performance loss on ques- tions about the topics that the original character was trained to answer. for conversational characters which is available for down- load as part of the ICT Virtual Human Toolkit. 1 NPCEd- itor is trained on a knowledge base in the form of linked question-answer pairs, and is able to answer novel ques- tions by selecting the most appropriate response from the available answers in the knowledge base. For each new in- put question, NPCEditor computes a language model for the ideal answer using the linked training data; it then compares the language model of the ideal answer to those of all of the answers in the knowledge base, and selects the closest avail- able answer based on a similarity metric between language models. The use of language models allows NPCEditor to overcome some variation in the phrasing of questions, and retrieve appropriate responses for questions it has not seen in the training data.