Discussion Structure Prediction Based on a Two-step Method

Proceedings of Recent Advances in Natural Language Processing, pages 538–546 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_061 538 Discussion Structure Prediction Based on a Two-step Method Takumi Himeno and Kazutaka Shimada Department of Artificial Intelligence, Kyushu Institute of Technology, Japan t himeno@pluto.ai.kyutech.ac.jp, shimada@ai.kyutech.ac.jp Abstract Conversations are often held in laboratories and companies. A summary is vital to grasp the content of a discussion for people who did not attend the discussion. If the summary is illus- trated as an argument structure, it is helpful to grasp the discussion’s essentials immediately. Our purpose in this paper is to predict a link structure between nodes that consist of utter- ances in a conversation: classification of each node pair into “linked” or “not-linked.” One approach to predict the structure is to utilize machine learning models. However, the result tends to over-generate links of nodes. To solve this problem, we introduce a two-step method to the structure prediction task. We utilize a machine learning-based approach as the first step: a link prediction task. Then, we apply a score-based approach as the second step: a link selection task. Our two-step methods dramati- cally improved the accuracy as compared with one-step methods based on SVM and BERT. 1 Introduction Meetings are often held in laboratories and compa- nies to come up with new research ideas and man- agement strategies. A summary is vital to grasp the content of a discussion for people who did not attend the discussion. Summaries are suitable for understanding the main points in discussions. As- sume that a summary is illustrated as a discussion structure. The summary is more powerful and help- ful to understand the main points in the discussion because users can immediately capture the flow of the discussion by using links between utterances. For this purpose, we need to predict the discussion structure of each discussion. Argument mining is one of the tasks to con- struct a structure of sentences (Stab and Gurevych, 2017a). It automatically derives the structure of argumentation from unstructured documents such as essays. It consists of four subtasks as follows: component identification, component classification, relation identification, and relation classification. Component identification is a task that extracts ar- gument components from a given document. Argu- ment components denote sentences and paragraphs related to the discussion structure. Component clas- sification is a task that assigns a label, e.g., claim, to each argument component. Relation identifica- tion is a task that predicts whether each pair of argument components is related or not. Relation classification is a task that assigns a label, such as “attack” and “support,” to the related pairs of argument components. In this paper, we deal with relation identifica- tion for constructing a discussion structure in a multi-party conversation. In other words, we con- struct a link prediction model for nodes consisting of some utterances. Methods in previous work of- ten predicted the discussion structures by using machine learning approaches, such as neural net- works. Himeno and Shimada (2020) have reported that such machine learning models tended to over- …

Content truncated to 500 words