Pneg: Prompt-based Negative Response Generation for Dialogue Response Selection Task

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10692 - 10703 December 7-11, 2022 ©2022 Association for Computational Linguistics PNEG: Prompt-based Negative Response Generation for Dialogue Response Selection Task Nyoungwoo Lee1, ChaeHun Park2, Ho-Jin Choi3, and Jaegul Choo2 1Scatter Lab 2KAIST AI 3KAIST nyoungwoo@scatterlab.co.kr, {ddehun, hojinc, jchoo}@kaist.ac.kr Abstract In retrieval-based dialogue systems, a response selection model acts as a ranker to select the most appropriate response among several can- didates. However, such selection models tend to rely on context-response content similarity, which makes models vulnerable to adversar- ial responses that are semantically similar but not relevant to the dialogue context. Recent studies have shown that leveraging these ad- versarial responses as negative training sam- ples is useful for improving the discriminating power of the selection model. Nevertheless, collecting human-written adversarial responses is expensive, and existing synthesizing meth- ods often have limited scalability. To overcome these limitations, this paper proposes a simple but efficient method for generating adversarial negative responses leveraging a large-scale lan- guage model. Experimental results on dialogue selection tasks show that our method outper- forms other methods of synthesizing adversar- ial negative responses. These results suggest that our method can be an effective alternative to human annotators in generating adversarial responses. Our dataset and generation code is available at https://github.com/leenw23/ generating-negatives-by-gpt3. 1 Introduction In retrieval-based dialogue systems, the response selection model aims to predict the most appropri- ate response among multiple candidates retrieved for a given conversation context (Zhou et al., 2018; Wu et al., 2019; Chen et al., 2022). The selection model is generally trained to distinguish a related (positive) response from randomly sampled neg- ative responses on training datasets, but such a model generally poses the following problems. First, randomly selected negatives are often too easy to distinguish because they are totally irrele- vant to the dialogue context (Li et al., 2019; Lin et al., 2020). In this case, the model is more likely Dialogue Context Positive Response PNEG Negative Responses (generated) Negative Response Generation by using GPT-3 Few Human-written examples Training Response Selection Model with generated negative responses Original Training Sample Yes! What would you like me to translate? Can you help me to translate these sentences into German? I can speak German, but can't speak French. How well can you speak them? I can speak French and German. How many languages can you speak? I speak three languages, but my English is poor. Can you give me a quick translation of everything? It's urgent. I am going on a trip to Germany next month Figure 1: A conceptual pipeline of prompt-based nega- tive response generation. to predict the response only by relying on the su- perficial content similarity of the context-response pairs (Yuan et al., 2019; Sai et al., 2020; Whang et al., 2021). These models are vulnerable to ad- versarial responses with high content similarity to the dialogue context, and fail to distinguish subtle differences in various contexts in real-world sce- narios. Second, random sampling causes sampling bias, …

Content truncated to 500 words