使用深度學習技術偵測提及人身安危事件的社群媒體貼文

USING DEEP LEARNING TO DETECT SOCIAL MEDIA POSTS MENTIONING POTENTIALLY HARMFUL EVENTS

社群媒體的使用度越來越廣泛，人們會在上面分享個人的生活紀錄、周圍發生的事情。人們發佈的內容有時候會提及有關人身安全的事件，例如：自殺、憂鬱、霸凌、家暴等。如果將社群媒體上的貼文做自動分類，判斷貼文是否有提及上述的高風險事件，將高嚴重程度的事件通知給平台的管理者、社工單位，讓相關單位追蹤發生的事件，若有必要時則能及時介入，避免發生憾事。
本研究以深度學習方法建立貼文分類系統，自動篩選出提及高風險事件的貼文。研究中將貼文斷句，使用CKIP BERT的編碼器取得句向量後，使用雙向長短期記憶網路（Bi-LSTM）、卷積神經網路（CNN）、殘差網路（ResNet）以及變換器中的編碼器（Transformer encoder）等深度學習模型做預測，並與支持向量機（SVM）做比較。
實驗結果顯示Bi-LSTM的最佳Macro-F1值為65.82%，CNN為57.53%，ResNet為56.17%，Transformer-encoder為57.25%，SVM為50.69%。

The use of social media has become increasingly widespread, as people record and share their lives and the events happening around them. Some posts may mention personal safety issues, such as suicide, depression, bullying, and domestic violence. If posts are automatically categorized to determine whether they mention any of the above high-risk events, these events will be reported to the platformʼs administrators and social work units. This will enable them to track the incidents and intervene in a timely manner if necessary.
In this study, deep learning approaches were used to build post-classification systems that automatically filter out posts mentioning high-risk events. A post was segmented into sentences, and the sentence vectors were obtained using the CKIP BERT encoder. Then, deep learning models such as Bi-directional Long Short-Term Memory (Bi-LSTM), Convolutional Neural Network (CNN), Residual Network (ResNet), and Transformer-encoder were used for prediction and compared with Support Vector Machine (SVM).
The experimental results showed that the best Macro-F1 values for Bi- LSTM, CNN, ResNet, Transformer-encoder, and SVM were 65.82%, 57.53%, 56.17%, 57.25%, and 50.69%, respectively.