運用深度學習方法於中文謠言辨識之比較

Comparison of Applying Deep Learning Methods on Misinformation Detection

邱靖雅(Chiu Ching-Ya)

指導教授 ： 鄭宇庭

國立政治大學/商學院/統計學系/碩士(2022年)

https://doi.org/10.6814/NCCU202200690

摘要

隨著新媒體時代與網際網路的蓬勃發展，資訊流通的速度更快速卻也伴隨社群媒體上大量參雜不實資訊的網路謠言被迅速散播。一般民眾尤其高齡者不易辨認謠言真實與否，在新冠肺炎疫情蔓延之下，誤信不實謠言可能造成不良影響。現今有許多官方與民間的訊息查證平台，如: 衛生福利部疾病管制署－澄清專區、Cofacts和台灣事實查核中心等，將可疑訊息查證結果公布於網頁上供民眾查詢真偽，然而單純以人工方式查核不僅流程耗費大量人力與時間成本，且闢謠速度跟不上網路謠言在群組間轉傳的速度。
因此本研究以Cofacts開源資料庫為中文文本，微調Google BERT、CKIP-BERT和RoBERTa預訓練模型對網路謠言進行「真實訊息」與「虛假訊息」的辨識與分類。根據模型評估指標結果，三個模型皆達到平均85%的準確度，能夠正確判斷85%訊息內容的真偽，其中又以RoBERTa模型的分類能力最佳。說明Google BERT、CKIP-BERT和RoBERTa預訓練模型的分類性能對於本研究所蒐集的網路謠言資料集具有良好的成效。

關鍵字

自然語言處理 ； 中文文本分類 ； BERT ； CKIP-BERT ； RoBERTa ； 假訊息辨識

並列摘要

With the development of the new media era and the Internet, the speed of information spreading is dramatically higher than before. But it is also accompanied by the rapid spread of a large number of online rumors mixed with fake information on social media. It is difficult for the general public, especially the elderly, to identify whether the rumors are true or not. Under the spread of the Covid-19 epidemic, the misleading of the fake news may cause public panic and serious consequences. Nowadays, there are many official or private rumor verification platforms, such as Taiwan Centers for Disease Control - Clarification Zone, Cofacts and Taiwan FactCheck Center, etc., publish suspicious information verification results on the website for public to check the authenticity. Not only does the process cost a lot of manpower and time, but also the speed of refuting rumors cannot keep up with the spreading which online rumors are transmitted among social media.
Therefore, this research uses Cofacts’s open sources as experimental corpus, and fine-tunes the Google BERT, CKIP-BERT and RoBERTa pre-training models to identify and classify "Truth Information" and "Fake Information" on online rumors. According to the results of the model evaluation indicators, the three models have achieved an average accuracy of 85%, and can correctly judge the authenticity of 85% of the message content. Among them, the RoBERTa model has the best classification ability. It shows that the identification performance of Google BERT, CKIP-BERT and RoBERTa pre-trained models have productive results for the rumor data set collected in this search.

並列關鍵字

Natural Language Processing (NLP) ； BERT ； CKIP-BERT ； RoBERTa ； Chinese text classification ； Online rumor identification

參考文獻

一、 中文文獻

1. 王鈞威，(2021)，基於整合RoBERTa與CRF模型之中文文法錯誤診斷系統，朝陽科技大學。

2. 吳晨皓，(2020)，BERT 與 GPT-2 分別應用於刑事案件之罪名分類及判決書生成，國立高雄科技大學。

3. 呂明聲，(2020)，基於深度學習之謠言檢測法: 以食安謠言為例，國立中央大學。

4. 邱彥誠，(2020)，應用人工智慧於股市新聞與情感分析預測股價走勢，國立臺北大學。