[Paper Review] Generating Long and Diverse Responses with Neural Conversation Models

Louis Yuanlong Shao, Stephan Gouws+4 more|arXiv (Cornell University)|Apr 24, 2017

Topic Modeling60 citations

TL;DR

This paper proposes a novel training and decoding framework for neural conversation models that enhances response length, coherence, and diversity. By integrating self-attention into the decoder, introducing the glimpse-model for scalable training, and employing a stochastic beam search with segment-by-segment reranking, the method generates longer, more coherent, and diverse responses—demonstrated through human evaluations showing superior acceptability and quality across varying response lengths.

ABSTRACT

Building general-purpose conversation agents is a very challenging task, but necessary on the road toward intelligent agents that can interact with humans in natural language. Neural conversation models -- purely data-driven systems trained end-to-end on dialogue corpora -- have shown great promise recently, yet they often produce short and generic responses. This work presents new training and decoding methods that improve the quality, coherence, and diversity of long responses generated using sequence-to-sequence models. Our approach adds self-attention to the decoder to maintain coherence in longer responses, and we propose a practical approach, called the glimpse-model, for scaling to large datasets. We introduce a stochastic beam-search algorithm with segment-by-segment reranking which lets us inject diversity earlier in the generation process. We trained on a combined data set of over 2.3B conversation messages mined from the web. In human evaluation studies, our method produces longer responses overall, with a higher proportion rated as acceptable and excellent as length increases, compared to baseline sequence-to-sequence models with explicit length-promotion. A back-off strategy produces better responses overall, in the full spectrum of lengths.

Motivation & Objective

• To address the challenge of short and generic responses in neural conversation models.
• To improve coherence and diversity in long-form dialogue generation.
• To scale sequence-to-sequence models to large-scale web-crawled dialogue datasets.
• To develop a decoding strategy that promotes diversity earlier in the generation process.
• To achieve higher human-rated quality across a spectrum of response lengths.

Proposed method

• Introduces self-attention in the decoder to maintain coherence in long responses.
• Proposes the glimpse-model for efficient training on large-scale datasets of over 2.3 billion messages.
• Develops a stochastic beam-search algorithm with segment-by-segment reranking to inject diversity early in generation.
• Employs a back-off strategy to improve response quality across all lengths.
• Trains end-to-end on a combined dataset of web-mined conversation messages.
• Uses human evaluation to assess response quality, length, and acceptability.

Experimental results

Research questions

• RQ1 Can self-attention in the decoder improve coherence in long dialogue responses?
• RQ2 How can diversity be effectively injected earlier in the response generation process?
• RQ3 Can the glimpse-model enable scalable training on massive dialogue datasets?
• RQ4 Does the proposed decoding strategy outperform baseline models in generating longer, higher-quality responses?
• RQ5 How does response quality vary across different lengths under the proposed method?

Key findings

• The proposed method generates longer responses with a higher proportion rated as acceptable and excellent as length increases, compared to baseline models with explicit length promotion.
• Human evaluations show the method produces better responses across the full spectrum of lengths, especially in coherence and diversity.
• The back-off strategy improves overall response quality, outperforming baselines consistently across all response lengths.
• The glimpse-model enables effective training on over 2.3 billion conversation messages, supporting scalability.
• Stochastic beam search with segment-by-segment reranking successfully enhances diversity in generated responses.
• Self-attention in the decoder contributes to improved coherence in long-form dialogue generation.