Multi-Document Transformer for Personality Detection

Personality detection aims to identify the personality traits implied in social media posts. The core of this task is put together information multiple scattered posts depict an overall profile for each user. Existing approaches either encode post individually or assemble arbitrarily into a new document that can be encoded sequentially hierarchically. While first approach ignores connection between posts, second tends introduce unnecessary post-order bias In paper, we propose multi-document Transformer, namely Transformer-MD, tackle above issues. When encoding post, Transformer-MD allows access other user through Transformer-XL’s memory tokens which share same position embedding.Besides, usually defined along different and trait may need attend information, has rarely been touched by existing research. To address concern, dimension attention mechanism on top obtain trait-specific representations multi-trait detection. We evaluate proposed model Kaggle Pandora MBTI datasets experimental results show it compares favorably with baseline methods.