[Paper Review] Neural Responding Machine for Short-Text Conversation

Lifeng Shang, Zhengdong Lu+1 more|arXiv (Cornell University)|Mar 9, 2015

Topic Modeling213 citations

TL;DR

This paper proposes the Neural Responding Machine (NRM), a sequence-to-sequence neural network model that generates responses for short-text conversations using an encoder-decoder framework with gated recurrent units (GRUs). Trained on 4.4 million Weibo post-response pairs, NRM outperforms retrieval-based and SMT-based methods, achieving over 75% of responses rated as suitable or neutral, with the hybrid NRM-hyp variant significantly outperforming others in both fluency and relevance.

ABSTRACT

We propose Neural Responding Machine (NRM), a neural network-based response generator for Short-Text Conversation. NRM takes the general encoder-decoder framework: it formalizes the generation of response as a decoding process based on the latent representation of the input text, while both encoding and decoding are realized with recurrent neural networks (RNN). The NRM is trained with a large amount of one-round conversation data collected from a microblogging service. Empirical study shows that NRM can generate grammatically correct and content-wise appropriate responses to over 75% of the input text, outperforming state-of-the-arts in the same setting, including retrieval-based and SMT-based models.

Motivation & Objective

• Address the challenge of generating diverse, fluent, and contextually relevant responses in one-round short-text conversations.
• Overcome limitations of retrieval-based models, which rely on pre-existing responses and struggle with customization and semantic mismatches.
• Improve upon SMT-based methods, which treat response generation as translation and often produce grammatically incorrect or semantically incoherent outputs.
• Develop a neural generative model that learns rich, dynamic representations of input posts to produce varied and appropriate responses.
• Demonstrate that a neural encoder-decoder framework can effectively model the non-parallel, multi-response nature of short-text conversations.

Proposed method

• Employ an encoder-decoder architecture with gated recurrent units (GRUs) to encode input posts into a context vector and decode it into a response.
• Introduce a dynamic context mechanism inspired by Bahdanau et al. (2014), allowing attention over the input sequence during decoding to improve alignment and relevance.
• Propose three variants: NRM-glo (global context), NRM-loc (local context with attention), and NRM-hyp (hybrid of global and local context) for improved representation learning.
• Train the model end-to-end using maximum likelihood estimation on a large-scale Weibo dataset of 4.4 million post-response pairs.
• Use beam search with a beam size of 500 to generate multiple diverse responses per input post, evaluating diversity and fluency.
• Apply a ranking-based evaluation with human annotators to assess response quality across fluency, relevance, and suitability.

Experimental results

Research questions

• RQ1 Can a neural encoder-decoder model effectively generate diverse, fluent, and contextually appropriate responses in one-round short-text conversations?
• RQ2 How does the inclusion of dynamic attention mechanisms during decoding affect response quality compared to static global encoding?
• RQ3 To what extent can a hybrid encoding strategy (combining global and local context) improve response generation over standalone approaches?
• RQ4 How does the performance of the proposed neural model compare to retrieval-based and SMT-based baselines in terms of fluency, relevance, and human-rated suitability?
• RQ5 Can the model generate multiple distinct yet high-quality responses to the same input post, indicating effective density estimation of response space?

Key findings

• The NRM-hyp model, combining global and local context representations, achieved the highest human-rated suitability score, significantly outperforming all baselines (p < 0.05).
• Over 75% of responses generated by NRM variants were rated as 'suitable' or 'neutral' by human annotators, indicating strong fluency and relevance.
• The retrieval-based model performed comparably to NRM-glo but was outperformed by NRM-hyp, with a p-value of 0.062 between NRM-loc and retrieval-based, indicating marginal significance.
• SMT-based models performed significantly worse than both retrieval and NRM models, with 74.4% of responses labeled as unsuitable due to fluency and relevance errors.
• The NRM-hyp model generated multiple diverse, fluent, and relevant responses to the same input post, demonstrating effective coverage of response distribution modes.
• The model successfully avoided common retrieval-based pitfalls such as mismatched named entities (e.g., incorrect restaurant names), producing more general and consistent responses.