Long-term Control for Dialogue Generation: Methods and Evaluation

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 738 - 753 July 10-15, 2022 ©2022 Association for Computational Linguistics Long-term Control for Dialogue Generation: Methods and Evaluation Ramya Ramakrishnan ASAPP rramakrishnan@asapp.com Hashan Buddhika Narangodage ASAPP hnarangodage@asapp.com Mauro Schilman ASAPP mschilman@asapp.com Kilian Q. Weinberger ASAPP, Cornell kweinberger@asapp.com Ryan McDonald ASAPP rmcdonald@asapp.com Abstract Current approaches for controlling dialogue response generation are primarily focused on high-level attributes like style, sentiment, or topic. In this work, we focus on constrained long-term dialogue generation, which involves more fine-grained control and requires a given set of control words to appear in generated re- sponses. This setting requires a model to not only consider the generation of these control words in the immediate context, but also pro- duce utterances that will encourage the genera- tion of the words at some time in the (possibly distant) future. We define the problem of con- strained long-term control for dialogue genera- tion, identify gaps in current methods for eval- uation, and propose new metrics that better measure long-term control. We also propose a retrieval-augmented method that improves per- formance of long-term controlled generation via logit modification techniques. We show through experiments on three task-oriented di- alogue datasets that our metrics better assess dialogue control relative to current alternatives and that our method outperforms state-of-the- art constrained generation baselines. 1 1 Introduction Despite recent advances in dialogue systems (Ser- ban et al., 2016; Ham et al., 2020), controlling di- alogue generation remains a significant challenge. Response generation in dialogue can be controlled towards different topics and styles (Madotto et al., 2020) or towards a set of hard constraints (i.e., lex- ical control words need to appear in the generated text) (Sha, 2020). We focus on the hard constraint setting, also known as constrained generation, as this provides a more fine-grained method of con- trolling dialogues. For example, consider a customer service use case (Figure 1), in which an agent speaks to a 1Our code is available at https://github.com/a sappresearch/constrained-dialogue-genera tion Figure 1: Examples of short vs. long-term control for dialogue generation. (Left) In short-term control, many control words are generated initially, but the conversa- tion is led away from the desired future. (Right) In long-term control, responses are generated with the fu- ture in mind with words generated at natural points in the conversation. customer about an issue. The goal is to gener- ate a given set of control words in the responses of one of the speakers (agent or customer). Naive constrained generation approaches (Pascual et al., 2020; Miao et al., 2019) use methods like beam search and stochastic search to force the genera- tion of these control words for short-term control, where control words need to appear in a single ut- terance or phrase. Because they do not consider the future, these approaches may generate the words all at once in a single response or not generate them at natural places in the conversation (Figure 1, left). …

Content truncated to 500 words