Automatic question-answer alignment for diverse local assembly minutes

Abstract

In some councils, a format called ‘collective questioning and collective answering’ is adopted.This format is difficult to read for residents, as the question and answer sentences are separated.To address such issues, we propose a method that automatically maps to a ‘one question, one answer’ format. The proposed method first performs segmentation using a model that has been fine-tuned on the minutes dataset using pre-trained BERT.The segmented questions and responses are converted into word weight vectors using Okapi BM25.It is calculated as a perfect match that maximizes the cosine similarity between the vectors of questions and responses.The proposed method performed automatic mapping on local assembly minutes. For assembly minutes, to which regular expressions are not well adapted, the proposed method improved the F-measure of automatic matching compared to the previous studies.