Saurabh Chakravarty, Edward Fox
Many documents are constituted by a sequence of question-answer (QA) pairs. Applying existing natural language processing (NLP) methods such as automatic summarization to such documents leads to poor results. Accordingly, we have developed classification meth- ods based on dialog acts to facilitate subsequent application of NLP techniques. This paper describes the ontology of dialog acts we have devised through a case study of a corpus of legal depositions that are made of QA pairs, as well as our development of machine/deep learning classifiers to identify dialog acts in such corpora. We have adapted state-of-the-art text classification methods based on a con- volutional neural network (CNN) and long short term memory (LSTM) to classify the questions and answers into their respective dialog acts. We have also used pre-trained BERT embeddings for one of our classifiers. Experimentation showed we could achieve an F1 score of 0.84 on dialog act classification involving 20 classes. Given such promising techniques to classify questions and answers into dialog acts, we plan to develop custom methods for each di- alog act, to transform each QA pair into a form that would allow for the application of NLP or deep learning techniques for other downstream tasks, such as summarization.
Saurabh Chakravarty, Raja Venkata Satya Phanindra Chava, Edward A. Fox: Dialog Acts Classification for Question-Answer Corpora. ASAIL@ICAIL 2019
- Date of publication:
- ASAIL with International Conference on Artificial Intelligence and Law