Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics 2020
DOI: 10.18653/v1/2020.acl-main.185
|View full text |Cite
|
Sign up to set email alerts
|

Negative Training for Neural Dialogue Response Generation

Abstract: Although deep learning models have brought tremendous advancements to the field of opendomain dialogue response generation, recent research results have revealed that the trained models have undesirable generation behaviors, such as malicious responses and generic (boring) responses. In this work, we propose a framework named "Negative Training" to minimize such behaviors. Given a trained model, the framework will first find generated samples that exhibit the undesirable behavior, and then use them to feed neg… Show more

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
1
1
1
1

Citation Types

0
39
0

Year Published

2020
2020
2024
2024

Publication Types

Select...
4
3
2

Relationship

1
8

Authors

Journals

citations
Cited by 39 publications
(39 citation statements)
references
References 20 publications
0
39
0
Order By: Relevance
“…This is because compared to the baseline models, the pretrained models are more sensitive to the contextual input, making them easier to manipulate. This makes the malicious response problem a more relevant issue to solve (He and Glass, 2019b).…”
Section: Implications and Discussionmentioning
confidence: 99%
“…This is because compared to the baseline models, the pretrained models are more sensitive to the contextual input, making them easier to manipulate. This makes the malicious response problem a more relevant issue to solve (He and Glass, 2019b).…”
Section: Implications and Discussionmentioning
confidence: 99%
“…In the area of generative dialogue, a number of works have focused on improving the standard likelihood training approach. Closer to our work is that of He and Glass (2019) which developed the approach of negative training to prevent generic and malicious responses in dialogue models. In terms of improving repetition and specificity, a recent alternative approach is that of control (Fan et al, 2018;Ficler and Goldberg, 2017;Ghazvininejad et al, 2017;See et al, 2019).…”
Section: Related Workmentioning
confidence: 99%
“…Each training instance is processed in the form of "[context] dialogue context [response] response" where response will be predicted given "[context] dialogue context [response]". The objective is a mixture of maximum likelihood and unlikelihood training(He and Glass, 2019;, which we find help reduce repeated and incoherent generations as observed inAdiwardana et al (2020). The unlikelihood training minimizes the likelihood of 1) randomly sampled responses from the corpus and 2) repeated bigrams from the previous generated tokens.…”
mentioning
confidence: 82%