상세 보기
초록
Predicting MBTI based on data from multiple social media posts presents a number of challenges associated with collecting lengthy and varied histories of users where the personality cues associated with those users are dispersed throughout numerous short texts. While recent advances in large language models (LLMs) have shown to be successful at this task, achieving such success has required using either very large backbones or external APIs. We propose a novel architecture called LP-MBTI that aggregates user history in a fixed budget manner through a combination of a small backbone DeBERTa-v3-small post encoder, lightweight post level self attention, and a Perceiver style latent bottleneck with 24 learnable latents. The contribution is not a new attention primitive but rather, a task specific integration of latent bottleneck aggregation and an MBTI aware training recipe utilizing focal loss, R-Drop consistency regularization, and auxiliary 16-type supervision. LP-MBTI achieves 90.55% mean accuracy, 87.39% mean F1 score, and 94.08% mean AUC on the Kaggle MBTI Benchmark; LP-MBTI improves the mean F1 score of EmoPerso by 6.3 percentage points and the mean AUC of Post-to-Personality by 1.4 percentage points, all within a self-contained 160.24M parameter local checkpoint. These results demonstrate that compact latent bottleneck aggregation can compete with LLM scale deployments at multi-post personality prediction. © 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
키워드
- 제목
- LP-MBTI: A 24-latent Perceiver core for memory-light multi-post personality
- 저자
- Huang, Gan; Huang, Han Yao; Kim, Kyung Tae; Liu, Tao
- 발행일
- 2027-01-01
- 유형
- Article
- 권
- 332
- 언어
- ENG
- 출판사
- Elsevier Ltd
- 발행국가
- 영국
- ISSN
- E 1873-6793
P 0957-4174