A Hierarchical Latent Structure for Variational Conversation Modeling

Proceedings of NAACL-HLT 2018, pages 1792–1801 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics A Hierarchical Latent Structure for Variational Conversation Modeling Yookoon Park Jaemin Cho Department of Computer Science and Engineering & Center for Superintelligence Seoul National University, Korea yookoonpark@vision.snu.ac.kr, {jaemin895,gunhee}@snu.ac.kr http://vision.snu.ac.kr/projects/vhcr Gunhee Kim Abstract Variational autoencoders (VAE) combined with hierarchical RNNs have emerged as a powerful framework for conversation model- ing. However, they suffer from the notori- ous degeneration problem, where the decoders learn to ignore latent variables and reduce to vanilla RNNs. We empirically show that this degeneracy occurs mostly due to two rea- sons. First, the expressive power of hierar- chical RNN decoders is often high enough to model the data using only its decoding distri- butions without relying on the latent variables. Second, the conditional VAE structure whose generation process is conditioned on a context, makes the range of training targets very sparse; that is, the RNN decoders can easily overfit to the training data ignoring the latent variables. To solve the degeneration problem, we pro- pose a novel model named Variational Hier- archical Conversation RNNs (VHCR), involv- ing two key ideas of (1) using a hierarchical structure of latent variables, and (2) exploiting an utterance drop regularization. With evalu- ations on two datasets of Cornell Movie Dia- log and Ubuntu Dialog Corpus, we show that our VHCR successfully utilizes latent vari- ables and outperforms state-of-the-art models for conversation generation. Moreover, it can perform several new utterance control tasks, thanks to its hierarchical latent structure. 1 Introduction Conversation modeling has been a long interest of natural language research. Recent approaches for data-driven conversation modeling mostly build upon recurrent neural networks (RNNs) (Vinyals and Le, 2015; Sordoni et al., 2015b; Shang et al., 2015; Li et al., 2017; Serban et al., 2016). Ser- ban et al. (2016) use a hierarchical RNN struc- ture to model the context of conversation. Ser- ban et al. (2017) further exploit an utterance latent variable in the hierarchical RNNs by incorporat- ing the variational autoencoder (VAE) framework (Kingma and Welling, 2014; Rezende et al., 2014). VAEs enable us to train a latent variable model for natural language modeling, which grants us several advantages. First, latent variables can learn an interpretable holistic representation, such as topics, tones, or high-level syntactic proper- ties. Second, latent variables can model inher- ently abundant variability of natural language by encoding its global and long-term structure, which is hard to be captured by shallow generative pro- cesses (e.g. vanilla RNNs) where the only source of stochasticity comes from the sampling of output words. In spite of such appealing properties of la- tent variable models for natural language mod- eling, VAEs suffer from the notorious degenera- tion problem (Bowman et al., 2016; Chen et al., 2017) that occurs when a VAE is combined with a powerful decoder such as autoregressive RNNs. This issue makes VAEs ignore latent variables, and eventually behave as vanilla RNNs. Chen et al. (2017) also note this degeneration issue …

Content truncated to 500 words