[Paper Review] Hierarchical Recurrent Attention Network for Response Generation

Xing Chen, Wei Wu+4 more|arXiv (Cornell University)|Jan 25, 2017

Topic Modeling116 citations

TL;DR

HRAN introduces hierarchical word- and utterance-level attention for multi-turn response generation, outperforming S2SA, HRED, and VHRED on perplexity and human judgments.

ABSTRACT

We study multi-turn response generation in chatbots where a response is generated according to a conversation context. Existing work has modeled the hierarchy of the context, but does not pay enough attention to the fact that words and utterances in the context are differentially important. As a result, they may lose important information in context and generate irrelevant responses. We propose a hierarchical recurrent attention network (HRAN) to model both aspects in a unified framework. In HRAN, a hierarchical attention mechanism attends to important parts within and among utterances with word level attention and utterance level attention respectively. With the word level attention, hidden vectors of a word level encoder are synthesized as utterance vectors and fed to an utterance level encoder to construct hidden representations of the context. The hidden vectors of the context are then processed by the utterance level attention and formed as context vectors for decoding the response. Empirical studies on both automatic evaluation and human judgment show that HRAN can significantly outperform state-of-the-art models for multi-turn response generation.

Motivation & Objective

• Address open-domain multi-turn response generation using conversation context.
• Model the hierarchy of context (words within utterances within the utterance sequence) and the differential importance of context elements.
• Improve response relevance and coherence by using hierarchical attention to select important words and utterances during generation.
• Demonstrate empirical gains over state-of-the-art baselines via automatic metrics and human judgments.

Proposed method

• Encode each utterance with a bidirectional GRU to produce word-level hidden vectors.
• Compute word-level attention dependent on both the decoder state and utterance context to form utterance vectors.
• Encode the sequence of utterance vectors with an utterance-level BRU to produce context representations.
• Apply an utterance-level attention to summarize the context into a context vector for each decoding step.
• Decode the response with a GRU-based language model conditioned on the context vectors, using beam search for generation.
• Train by maximizing the log-likelihood of the ground-truth responses.

Experimental results

Research questions

• RQ1 Can hierarchical word and utterance-level attention improve relevance and coherence in multi-turn response generation?
• RQ2 Does jointly modeling context hierarchy and part-level importance yield measurable gains over existing hierarchical models (HRED, VHRED) and non-hierarchical baselines?
• RQ3 How does HRAN perform on automatic perplexity metrics and human judgments compared to state-of-the-art methods?
• RQ4 What insights do attention visualizations provide about which words and utterances influence generation?

Key findings

• HRAN achieves the lowest perplexity on both validation and test sets compared with S2SA, HRED, and VHRED.
• Validation perplexity: S2SA 43.679, HRED 46.279, VHRED 44.548, HRAN 40.257.
• Test perplexity: S2SA 44.508, HRED 47.467, VHRED 45.484, HRAN 41.138.
• HRAN outperforms baselines in human side-by-side judgments across multiple comparisons.
• Ablation studies show word-level attention and utterance-level attention each contribute to performance gains; removing components degrades results.
• Attention visualizations illustrate HRAN focusing on informative words (e.g., “girl”, “boyfriend”, height numbers) and key utterances in the context.