Improving Neural Conversational Models with Entropy-Based Data Filtering

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5650–5669 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 5650 Improving Neural Conversational Models with Entropy-Based Data Filtering Richard Csaky Department of Automation and Applied Informatics Budapest University of Technology and Economics ricsinaruto@hotmail.com Patrik Purgai Department of Automation and Applied Informatics Budapest University of Technology and Economics purgai.patrik@gmail.com Gabor Recski Apollo.AI gabor@apollo.ai Abstract Current neural network-based conversational models lack diversity and generate boring re- sponses to open-ended utterances. Priors such as persona, emotion, or topic provide addi- tional information to dialog models to aid re- sponse generation, but annotating a dataset with priors is expensive and such annotations are rarely available. While previous methods for improving the quality of open-domain re- sponse generation focused on either the un- derlying model or the training objective, we present a method of filtering dialog datasets by removing generic utterances from training data using a simple entropy-based approach that does not require human supervision. We conduct extensive experiments with different variations of our method, and compare dialog models across 17 evaluation metrics to show that training on datasets filtered this way re- sults in better conversational quality as chat- bots learn to output more diverse responses. 1 Introduction Current open-domain neural conversational mod- els (NCM) are trained on pairs of source and tar- get utterances in an effort to maximize the likeli- hood of each target given the source (Vinyals and Le, 2015). However, real-world conversations are much more complex, and a plethora of suitable tar- gets (responses) can be adequate for a given in- put. We propose a data filtering approach where the “most open-ended” inputs - determined by cal- culating the entropy of the distribution over tar- get utterances - are excluded from the training set. We show that dialog models can be improved us- ing this simple unsupervised method which can be applied to any conversational dataset. We con- duct several experiments to uncover how some of the current open-domain dialog evaluation meth- ods behave with respect to overfitting and random data. Our software for filtering dialog data and au- tomatic evaluation using 17 metrics is released on GitHub under an MIT license12. 2 Background Most open-domain NCMs are based on neural net- work architectures developed for machine trans- lation (MT, Sutskever et al. (2014); Cho et al. (2014); Vaswani et al. (2017)). Conversational data differs from MT data in that targets to the same source may vary not only grammatically but also semantically (Wei et al., 2017; Tandon et al., 2017): consider plausible replies to the question What did you do today?. Dialog datasets also con- tain generic responses, e.g. yes, no and i don’t know, that appear in a large and diverse set of con- texts (Mou et al., 2016; Wu et al., 2018). Follow- ing the approach of modeling conversation as a se- quence to sequence (seq2seq, Sutskever et al. (2014)) transduction of single dialog turns, these issues can be …

Content truncated to 500 words