Named Entity Emotion Intensity Tagging for Suicidal Ideation Detection From Social Media Texts During MT

Abstract

Detecting suicide ideation from social media texts is challenging as it necessitates analyzing the content and context of the text utterances. Modeling the temporal trajectory of suicide through various stages like stress, depression, thought and strengthening of the thought becomes difficult with content features alone and this needs context cues too. This work proposes a hybrid feature representation incorporating both content and context features. The richness of features is enhanced using adversarial learning. Neural machine translation is done on hybrid feature representation to provide named entities with emotional intensity tagging. The sequence of emotional intensity tagging is mapped by Bi-directional Long short term memory (LSTM) to suicidal ideation label. The proposed integration of emotion tagging to SI detection with Bi-directional LSTM provided an SI detection accuracy of 95.54% which is atleast 3% higher compared to existing works