仇恨性言論偵測及仇恨性用戶分析

Abstract

近年來，網路仇恨性言論受到各社群平臺的重視，仇恨性言論偵測亦成為國際上之一項研究主題；然而，仇恨性言論偵測在華語使用圈中受到的重視遠不及西方社會，目前仇恨性言論偵測所使用的資料集仍多為推特上的英文留言，忽略了其他語系。有鑑於此，本研究欲建立華文仇恨性言論偵測模型並進行華文仇恨性用戶之分析。本研究提出了新的華文仇恨性言論資料集 PTT_HateSpeech，使用臺灣之電子布告欄系統 PTT 作為資料來源，收集 12 個月的資料，並人工標記了 38,950 則推文。我們根據此資料集訓練了分類模型來進行性別歧視相關之仇恨性言論偵測，最終之平均 F1-score 為 0.5976。透過訓練結果，我們將模型應用在使用者分析上，探討仇恨性用戶和一般用戶在網路使用習慣上之不同，並發現仇恨性用戶有群聚現象，且傾向於使用較為激烈和情緒性之字眼。本研究旨在提供臺灣本地之仇恨性言論資料集，補齊現階段研究中缺乏的華文資料集，並且從自然語言處理的角度分析臺灣網路社群之生態。

Hate speech (HS) is an increasing problem worldwide, and HS detection is an urgent issue. However, only few related studies have been addressed in East Asia society. The majority of research resources are Twitter corpora in English, which may only provide hateful content generated from English users and ignore diverse cultural features. Therefore, in this work, we develop PTT_HateSpeech, a novel sexism HS dataset collected from a Taiwanese bulletin board system named PTT (telnet://ptt.cc), to analyze Chinese linguistic patterns in HS detection task. The dataset contains 38,950 comments across 12 months with hand-annotated “hateful” or “non-hateful”. We train classification models to detect HS, and the average F1-score is 0.5976. Equipped with the proposed model, we conduct further user behavior analysis and sentiment analysis to compare hateful users and normal users. We find out that (1) intensive interactions can be observed in the group of hateful users, and (2) hateful users tend to use fiercer and angrier words, showing low-valence but high-arousal emotions. Our research bridges the gap by covering Taiwanese local data in the investigation, and we present a comprehensive study of Taiwanese online ecology from the field of natural language processing.