[논문 리뷰] Multi-Task Learning for Speaker-Role Adaptation in Neural Conversation Models

Yi Luan, Chris Brockett외 3명|arXiv (Cornell University)|2017. 10. 20.

Topic Modeling참고 문헌 26인용 수 65

한 줄 요약

이 논문은 Seq2Seq 대화 모델링과 비대화 데이터로 학습된 오토인코더를 결합한 다중 작업 학습 프레임워크를 제안하여 발화자 역할에 적응시키고 Twitter 데이터에서 perplexity, BLEU, 인간 평가를 개선한다.

ABSTRACT

Building a persona-based conversation agent is challenging owing to the lack of large amounts of speaker-specific conversation data for model training. This paper addresses the problem by proposing a multi-task learning approach to training neural conversation models that leverages both conversation data across speakers and other types of data pertaining to the speaker and speaker roles to be modeled. Experiments show that our approach leads to significant improvements over baseline model quality, generating responses that capture more precisely speakers' traits and speaking styles. The model offers the benefits of being algorithmically simple and easy to implement, and not relying on large quantities of data representing specific individual speakers.

연구 동기 및 목표

• Address lack of speaker-specific conversation data by leveraging both cross-speaker conversation data and non-conversational speaker data.
• Develop a multi-task training regime that shares decoder parameters between a Seq2Seq conversational model and an autoencoder.
• Demonstrate that shared decoder parameters can adapt to target speaker roles without requiring大量 data from individual speakers.

제안 방법

• Use two tasks: a Seq2Seq conversational task over a large general population of speakers and an Autoencoder task over non-conversational data of target speakers.
• Share only the decoder parameters between the Seq2Seq model and the Autoencoder to enable speaker-adaptive generation.
• Train by alternating task batches and selecting models based on Seq2Seq perplexity on a development set.

실험 결과

연구 질문

• RQ1 Can non-conversational, speaker-specific data be used to adapt a general conversational model to speaker roles via multi-task learning?
• RQ2 Does sharing the decoder between Seq2Seq and autoencoder tasks improve the model’s ability to reflect speaker traits and speaking style?
• RQ3 Is the multi-task approach more effective than a baseline Seq2Seq with MMI in perplexity and BLEU on real Twitter data?
• RQ4 Do speaker embeddings (MTask-M) offer advantages over per-speaker specialized models (MTask-S) in efficiency and performance?
• RQ5 How do the models fare in human evaluations regarding capturing target authors’ stylistic and domain characteristics?

주요 결과

• The multi-task models achieve substantial perplexity reductions vs. baseline (dev: Baseline 56.33; MTask-S 32.27; MTask-M 44.96; dev reductions: -42.7% and -20.2% respectively; test: Baseline 61.17; MTask-S 39.83; MTask-M 43.21; test reductions: -34.9% and -29.4%).
• BLEU gains are substantial for both multi-task variants (dev: Baseline 1.32; MTask-S 1.76; MTask-M 2.52; increases of +33.3% and +90.1% respectively; test: Baseline 1.31; MTask-S 1.69; MTask-M 2.25; increases of +29.0% and +71.7%).
• Distinct-1 and distinct-2 show higher diversity for both multi-task models (distinct-1 dev: Baseline 1.69%; MTask-S 2.43%; MTask-M 2.44%; distinct-2 dev: Baseline 6.53%; MTask-S 10.2%; MTask-M 9.79%).
• Human evaluation shows MTask-M achieves statistically significant improvement over the baseline (p = 0.026) in pairwise judgments, and both MTask-S and MTask-M outperform the baseline on average.