Neural Response Generation via GAN with an Approximate Embedding Layer

By Zhen Xu, Bingquan Liu, Baoxun Wang, Chengjie Sun, Xiaolong Wang, Zhuoran Wang, Chao Qi. 2017

86 citations recorded by the source corpus.

This paper presents a Generative Adversarial Network (GAN) to model single-turn short-text conversations, which trains a sequence-to-sequence (Seq2Seq) network for response generation simultaneously with a discriminative classifier that measures the differences between human-produced responses and machine-generated ones. In addition, the proposed method introduces an approximate embedding layer to solve the non-differentiable problem caused by the sampling-based output decoding procedure in the Seq2Seq generative model. The GAN setup provides an effective way to avoid noninformative responses (a.k.a “safe responses”), which are frequently observed in traditional neural response generators. The experimental results show that the proposed approach significantly outperforms existing neural response generation models in diversity metrics, with slight increases in relevance scores as well, when evaluated on both a Mandarin corpus and an English corpus.

Subfields: Artificial Intelligence, Generative Models, Natural Language Generation, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Jianfeng Gao
2. Jason Aaron Edward Weston
3. Mohit Bansal
4. Yejin Choi
5. Zhe Gan
6. Percy Liang
7. Kyunghyun Cho
8. Noah A. Smith

In Generative Models

1. Ilia Kulikov
2. Kyunghyun Cho
3. Jianfeng Gao
4. Yizhe Zhang
5. Zhe Gan
6. Jason Aaron Edward Weston
7. Verena Rieser
8. Xiaolong Wang

In Natural Language Generation

1. Verena Rieser
2. Jianfeng Gao
3. Ondrej Dusek

In Signal Processing

1. Zhen Xu
2. Zhe Gan
3. Jason Aaron Edward Weston
4. Verena Rieser
5. Sunghun Kim
6. Jianfeng Gao
7. Michel Galley
8. Kyunghyun Cho

Citations in the corpus

• Generating Informative and Diverse Conversational Responses via Adversarial Information Maximization (2018, methodology): “Researchers have applied Generative Adversarial Networks (GANs) [6] to neural response generation [7, 8].”
• Neural Approaches to Conversational AI (2018, positive): “atively bland due to the p(TjS) inference criterion (beam search), MMI rescoring often mitigates rather than completely eliminates the blandness problem. More recently, researchers (Li et al., 2017c; Xu et al., 2017; Zhang et al., 2018e) have used adversarial training and Generative Adversarial Netw”
• Importance of Search and Evaluation Strategies in Neural Dialogue Modeling (2018, background): “Greedy search Greedy search has been the search algorithm of choice among the recent papers on neural dialogue modeling (Gu et al., 2018; Zhao et al., 2017; Xu et al., 2017; Weston et al., 2018; Zhang et al., 2018).”
• LSDSCC: a Large Scale Domain-Specific Conversational Corpus for Response Generation with Diversity Oriented Evaluation Metrics (2018, result): “…2015; Lowe et al., 2017; Wu et al., 2018); b) discrete scores measuring the quality of generated results by human labeling (Shang et al., 2015; Serban et al., 2016; Xu et al., 2017); and c) case study comparing the generated results of different NRG models (Shang et al., 2015; Wang et al., 2017).”
• DialogWAE: Multimodal Response Generation with Conditional Wasserstein Auto-Encoder (2018, methodology): “as shown great success in image generation, adapting it to natural dialog generators is a non-trivial task. This is due to the non-differentiable nature of natural language tokens (Shen et al., 2017; Xu et al., 2017). Li et al. (2017a) address this problem by combining GAN with Reinforcement Learnin”
• Importance of a Search Strategy in Neural Dialogue Modelling (2018, background): “Greedy search Greedy search has been the search algorithm of choice among the recent papers on neural dialogue modelling (Gu et al., 2018; Zhao et al., 2017; Xu et al., 2017; Weston et al., 2018; Zhang et al., 2018).”
• Better Conversations by Modeling, Filtering, and Optimizing for Coherence and Diversity (2018, background): “• D-1, D-2, D-Sent – common metrics used to evaluate the diversity of generated responses (e.g. Li et al., 2016a; Xu et al., 2017; Xing et al., 2017; Dhingra et al., 2017): the proportion of distinct unigrams, bigrams, and sentences in the outputs.”
• Multi-turn Dialogue Response Generation with Autoregressive Transformer Models (2019, methodology): “…into model training have been explored to address this (Li et al. 2016b; Li et al. 2017; Yu et al. 2017; Che et al. 2017; Zhang et al. 2017; Xu et al. 2017; Zhang et al. 2018b; Olabiyi et al. 2019b). v) Training Objective: Most existing dialogue models are trained using maximum likelihood…”