[3Yin2] Interactive session 1

[3Yin2-13] Semi-automatic extraction of question and answer parts from inquiry response mails

〇Kohei Hara¹ (1.Shimadzu Corporation)

Keywords:question answering, inquiry response mail, natural language processing, text classification

This paper proposes a semi-automatic extraction method of a list of question and answer (QA) from inquiry response mails to develop a QA system. To develop it, a list that includes a lot of QA pairs is necessary. It costs a lot to make the list. Making it is roughly divided into two parts: "extracting sentences corresponding to Q or A from the mails" and "checking / correcting the excerpt results". In particular, the former part requires reading a large number of mails. In this paper, a bidirectional LSTM was built to classify each sentence in mails as QA or not. In the experiment, it was confirmed that to classify them in mails is more accurate than in other types of texts like inquiry response reports because, in the case of mails, a writing style is unified according to business rules and a diversity of not QA sentences is small.