Generating Dialogue Responses from a Semantic Latent Space

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 4339–4349, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 4339 Generating Dialogue Responses from a Semantic Latent Space Wei-Jen Ko1 Avik Ray2 Yilin Shen2 Hongxia Jin2 1 The University of Texas at Austin 2 Samsung Research America wjko@utexas.edu, {avik.r, yilin.shen, hongxia.jin}@samsung.com Abstract Existing open-domain dialogue generation models are usually trained to mimic the gold response in the training set using cross-entropy loss on the vocabulary. However, a good re- sponse does not need to resemble the gold re- sponse, since there are multiple possible re- sponses to a given prompt. In this work, we hypothesize that the current models are unable to integrate information from multiple seman- tically similar valid responses of a prompt, re- sulting in the generation of generic and unin- formative responses. To address this issue, we propose an alternative to the end-to-end clas- sification on vocabulary. We learn the pair re- lationship between the prompts and responses as a regression task on a latent space instead. In our novel dialog generation model, the rep- resentations of semantically related sentences are close to each other on the latent space. Hu- man evaluation showed that learning the task on a continuous space can generate responses that are both relevant and informative. 1 Introduction The sequence-to-sequence framework and trans- former based models are the most popular choices for designing open-domain neural response gener- ation systems (Vinyals and Le, 2015; Wolf et al., 2018). Those models typically involve maximiz- ing the probability of the ground truth response given the input prompt, trained using a cross en- tropy loss on the vocabulary. However, dialogue response generation is an open-ended, high entropy task, since there can be a wide variety of possi- ble responses to a given prompt. A good response does not have to use similar vocabulary or simi- lar sentence structure as the gold response, thus the end-to-end cross entropy loss is unsuitable for this task. We hypothesize that this fundamental deficiency is the primary reason why dialog gen- eration models tend to generate bland and uninfor- mative responses, such as “I don’t know” (Serban et al., 2016), despite the presence of much more instances of specific responses in the training data than generic responses. The specific issue is the following. A model trained using maximum likelihood objective treats each token of the vocabulary independently. The probabilities of each individual informative word in the vocabulary are low because the answer is open-ended. The model is unable to capture that most of the probability mass are on a group of se- mantically related words. Thus the words with the highest probabilities are often uninformative stop words with high frequency in the training data. A similar effect happens on the utterance level when using beam search decoding. When searching for the most probable utterance, the probability of each candidate sentence is calculated independently, and the model is unable to use the semantic related- ness between different candidate utterances (Qiu …

Content truncated to 500 words