[Paper Review] Stochastic Language Generation in Dialogue using Recurrent Neural Networks with Convolutional Sentence Reranking

Tsung-Hsien Wen, Milica Gašić+5 more|arXiv (Cornell University)|Aug 7, 2015

Topic ModelingComputer Science34 references78 citations

TL;DR

This paper proposes a neural network-based natural language generation (NLG) system for spoken dialogue systems that jointly uses a recurrent neural network (RNN) generator, a convolutional neural network (CNN) reranker, and a backward RNN reranker to produce fluent, diverse, and semantically accurate utterances. Trained solely on dialogue act-utterance pairs without semantic alignments or handcrafted rules, the model outperforms n-gram and rule-based baselines in both automatic metrics and human evaluations, achieving higher fluency, better slot accuracy, and greater linguistic variation.

ABSTRACT

The natural language generation (NLG) component of a spoken dialogue system (SDS) usually needs a substantial amount of handcrafting or a well-labeled dataset to be trained on. These limitations add significantly to development costs and make cross-domain, multi-lingual dialogue systems intractable. Moreover, human languages are context-aware. The most natural response should be directly learned from data rather than depending on predefined syntaxes or rules. This paper presents a statistical language generator based on a joint recurrent and convolutional neural network structure which can be trained on dialogue act-utterance pairs without any semantic alignments or predefined grammar trees. Objective metrics suggest that this new model outperforms previous methods under the same experimental conditions. Results of an evaluation by human judges indicate that it produces not only high quality but linguistically varied utterances which are preferred compared to n-gram and rule-based systems.

Motivation & Objective

• To reduce the need for extensive handcrafting and semantic annotation in dialogue system NLG by enabling end-to-end training on dialogue act-utterance pairs.
• To improve generation quality and linguistic diversity in spoken dialogue systems without relying on rule-based templates or syntactic trees.
• To develop a scalable, data-driven NLG approach that generalizes across domains and languages with minimal feature engineering.
• To evaluate the model’s performance against n-gram and rule-based baselines using both automatic metrics and human preference judgments.

Proposed method

• A forward RNN language model is trained on delexicalized dialogue utterances, where slot values are replaced by symbolic placeholders, and conditioned on dialogue act features and a gating mechanism to control slot generation.
• A convolutional neural network (CNN) reranker is introduced to validate semantic consistency of candidate utterances, especially for out-of-vocabulary or rare slot-value pairs not covered by delexicalization.
• A backward RNN reranker is added to improve fluency by modeling context from both directions, enhancing sentence-level coherence.
• The system uses a two-stage decoding process: first generating candidate utterances via beam search, then reranking them using a weighted combination of RNN, CNN, and backward RNN scores.
• The model is jointly optimized using a differentiable objective that combines fluency, semantic accuracy, and diversity through a learnable reranking function.
• Post-processing restores slot values from the delexicalized output, enabling generation of natural-sounding, domain-specific utterances.

Experimental results

Research questions

• RQ1 Can a neural NLG model trained only on dialogue act-utterance pairs without semantic alignment produce higher-quality responses than n-gram or rule-based systems?
• RQ2 To what extent does the integration of a CNN reranker improve semantic consistency for rare or unseen slot-value combinations?
• RQ3 How does the backward RNN reranker contribute to fluency and coherence in generated utterances compared to unidirectional RNNs?
• RQ4 What is the impact of training data size on the model’s ability to generate diverse and accurate utterances through top-n sampling?
• RQ5 Does the proposed architecture enable better linguistic variation while maintaining high semantic accuracy and fluency in human evaluations?

Key findings

• The proposed RNN-based NLG model outperforms the n-gram-based baseline in both BLEU score and slot error rate, with a 1.5% improvement on hard test sets containing rare slot values.
• Human judges preferred the neural model’s outputs over rule-based and n-gram baselines in terms of naturalness and linguistic variety, with a statistically significant preference for the RNN model in informativeness and fluency.
• The CNN reranker significantly improves performance on out-of-vocabulary and rare slot-value pairs, achieving a 1.5% BLEU improvement on the hard test set compared to 1% on the full set.
• The backward RNN reranker consistently improves fluency across all n-best selection settings (n=1, 5, 10), with stable gains regardless of diversity trade-offs.
• The model achieves near-optimal performance with only 2,000 training utterances when selecting the 1-best output, but requires over 4,156 utterances to achieve high linguistic diversity via top-5 sampling.
• The system demonstrates strong potential for zero-shot or few-shot adaptation to new domains due to its use of distributed word representations and end-to-end training.