pith. sign in

arxiv: 1906.09556 · v1 · pith:CKJZ24IDnew · submitted 2019-06-23 · 💻 cs.CL

DAL: Dual Adversarial Learning for Dialogue Generation

Shaobo Cui , Rongzhong Lian , Di Jiang , Yuanfeng Song , Siqi Bao , Yong Jiang This is my paper

Pith reviewed 2026-05-25 17:52 UTC · model grok-4.3

classification 💻 cs.CL
keywords dialogue generationadversarial learningdual learningresponse diversitynatural language generationopen-domain dialoguegenerative modelssafe responses
0
0 comments X

The pith

DAL uses the duality between query generation and response generation together with adversarial learning to reduce safe replies and raise both diversity and naturalness in open-domain dialogue.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Dual Adversarial Learning (DAL) as a framework for generating responses in open-domain dialogue systems. It treats the symmetry between turning a response back into a query and turning a query into a response as a way to force the model away from generic safe outputs and toward more varied ones. A separate adversarial component trains a discriminator to distinguish machine responses from human ones, pushing the generator to produce replies that pass as natural. Experiments report gains on automatic diversity metrics and on human ratings of overall quality compared with prior methods. The work positions itself as the first to combine these two mechanisms for this task.

Core claim

DAL is the first work to innovatively utilize the duality between query generation and response generation to avoid safe responses and increase the diversity of the generated responses. Additionally, DAL uses adversarial learning to mimic human judges and guides the system to generate natural responses. Experimental results demonstrate that DAL effectively improves both diversity and overall quality of the generated responses and outperforms the state-of-the-art methods regarding automatic metrics and human evaluations.

What carries the argument

Dual Adversarial Learning framework that jointly exploits query-response duality for diversity and an adversarial discriminator for naturalness.

If this is right

  • Generated responses exhibit higher lexical and semantic diversity by construction.
  • Adversarial training produces responses judged more natural by human raters.
  • The combined system surpasses prior state-of-the-art dialogue models on standard automatic metrics.
  • Both the safe-response problem and the unnatural-response problem are addressed simultaneously within one training procedure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same duality-plus-adversarial pattern could transfer to other paired-sequence tasks such as machine translation or text summarization.
  • Successful deployment would lower dependence on post-hoc human filtering for dialogue quality.
  • If the dual component generalizes, it might reduce the data volume needed to train diverse open-domain agents.

Load-bearing premise

The assumption that the duality between query generation and response generation can be leveraged reliably to avoid safe responses and increase diversity without creating new instabilities, and that the adversarial discriminator serves as a stable proxy for human naturalness judgments.

What would settle it

A controlled human evaluation in which DAL outputs receive lower diversity or naturalness scores than strong baseline models would falsify the central claims.

Figures

Figures reproduced from arXiv: 1906.09556 by Di Jiang, Rongzhong Lian, Shaobo Cui, Siqi Bao, Yong Jiang, Yuanfeng Song.

Figure 1
Figure 1. Figure 1: Dual Adversarial Learning. 2 Related Work 2.1 Dual Learning Many machine learning tasks have emerged in dual forms, such as dual neural machine transla￾tion (dual-NMT) (He et al., 2016), image classifi￾cation and conditional image generation (van den Oord et al., 2016). Dual learning (He et al., 2016) is proposed on the assumption that the dual cor￾relation could be used to improve both the pri￾mal task an… view at source ↗
Figure 2
Figure 2. Figure 2: An example to illustrate why duality promotes diversity. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Case study. Method Human rating Kappa Seq2Seq 0.470 0.56 MMI-anti 0.568 0.46 MMI-bidi 0.523 0.60 Adver-REIN 0.767 0.49 GAN-AEL 0.758 0.52 DAL-Dual (ours) 0.730 0.47 DAL-DuAd (ours) 0.778 0.50 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Time consumed by different methods. 6 Conclusion We propose a novel framework named DAL to al￾leviate two prominent problems (safe responses and unnatural responses) plaguing dialogue gen￾eration. The dual learning proposed in this paper is the first effort to utilize the reverse dependency between queries and responses to reduce the prob￾ability of safe response generation and improve the diversity of the… view at source ↗
read the original abstract

In open-domain dialogue systems, generative approaches have attracted much attention for response generation. However, existing methods are heavily plagued by generating safe responses and unnatural responses. To alleviate these two problems, we propose a novel framework named Dual Adversarial Learning (DAL) for high-quality response generation. DAL is the first work to innovatively utilizes the duality between query generation and response generation to avoid safe responses and increase the diversity of the generated responses. Additionally, DAL uses adversarial learning to mimic human judges and guides the system to generate natural responses. Experimental results demonstrate that DAL effectively improves both diversity and overall quality of the generated responses. DAL outperforms the state-of-the-art methods regarding automatic metrics and human evaluations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a Dual Adversarial Learning (DAL) framework for open-domain dialogue response generation. It claims to be the first to leverage the duality between query generation and response generation to avoid safe responses and increase diversity, while using adversarial learning to mimic human judges and produce natural responses. The paper asserts that DAL outperforms state-of-the-art methods on automatic metrics and human evaluations regarding diversity and overall quality.

Significance. If the experimental claims hold under rigorous validation, the work could be significant for dialogue generation by addressing safe and unnatural responses via a dual-learning plus adversarial mechanism. The claimed novelty in exploiting query-response duality would be a useful contribution if supported by clear evidence isolating its effect.

major comments (3)
  1. [Abstract] Abstract: the central claim of experimental outperformance is stated without any details on datasets, baselines, training procedure, or statistical significance, so the claim cannot be evaluated.
  2. [§3.2] §3.2 (adversarial component): the claim that adversarial learning 'mimics human judges' and guides natural responses rests on the untested assumption that discriminator scores correlate with human naturalness ratings; no correlation analysis, human-discriminator agreement table, or ablation is supplied, which is load-bearing given known instabilities in text GAN training.
  3. [§4] §4 (experiments): no ablation isolating the duality mechanism from the adversarial component is reported, so improvements cannot be attributed to the claimed innovations rather than other factors.
minor comments (2)
  1. [Abstract] Abstract contains a grammatical error ('utilizes' should be 'utilize').
  2. [§3] Notation for the dual generators (query vs. response) is introduced without an explicit equation or diagram in the method overview.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript accordingly to strengthen the presentation and evidence.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of experimental outperformance is stated without any details on datasets, baselines, training procedure, or statistical significance, so the claim cannot be evaluated.

    Authors: We agree that the abstract would benefit from additional context to allow evaluation of the claims. In the revised version, we will expand the abstract to briefly mention the primary dataset (DailyDialog), the main baselines compared against, and that improvements are supported by statistical significance testing. Detailed training procedures and full experimental setup will remain in Section 4 due to space constraints. revision: yes

  2. Referee: [§3.2] §3.2 (adversarial component): the claim that adversarial learning 'mimics human judges' and guides natural responses rests on the untested assumption that discriminator scores correlate with human naturalness ratings; no correlation analysis, human-discriminator agreement table, or ablation is supplied, which is load-bearing given known instabilities in text GAN training.

    Authors: The referee correctly identifies that the manuscript does not provide explicit correlation analysis between discriminator scores and human naturalness ratings. We will add this analysis in a new subsection of the revised manuscript, including Pearson correlation coefficients and a human-discriminator agreement table. We will also elaborate on how the dual learning component contributes to training stability beyond standard GAN approaches. revision: yes

  3. Referee: [§4] §4 (experiments): no ablation isolating the duality mechanism from the adversarial component is reported, so improvements cannot be attributed to the claimed innovations rather than other factors.

    Authors: We agree that the absence of an ablation study isolating the duality mechanism makes it difficult to attribute improvements specifically to the claimed contributions. In the revised manuscript, we will include ablation experiments that separately disable the duality component and the adversarial component, reporting effects on diversity metrics and human evaluation scores. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper proposes the DAL framework for dialogue generation based on duality between query and response generation plus adversarial learning, but contains no equations, derivations, or first-principles results that reduce to inputs by construction. No self-definitional steps, fitted inputs renamed as predictions, or load-bearing self-citations appear in the abstract or described content. Experimental results and human evaluations are presented as external validation rather than tautological outputs. The derivation chain is therefore self-contained with independent empirical content.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no free parameters, axioms, or invented entities can be extracted.

pith-pipeline@v0.9.0 · 5647 in / 1013 out tokens · 30141 ms · 2026-05-25T17:52:08.975867+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · 5 internal anchors

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

    work page

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

    work page

  3. [3]

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  4. [4]

    Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378

    work page 1971

  5. [5]

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets. In NIPS, pages 2672--2680

    work page 2014

  6. [6]

    Di He, Yingce Xia, Tao Qin, Liwei Wang, Nenghai Yu, Tieyan Liu, and Wei-Ying Ma. 2016. Dual learning for machine translation. In NIPS, pages 820--828

    work page 2016

  7. [7]

    Baotian Hu, Zhengdong Lu, Hang Li, and Qingcai Chen. 2014. Convolutional neural network architectures for matching natural language sentences. In NIPS, pages 2042--2050

    work page 2014

  8. [8]

    Xun Huang, Yixuan Li, Omid Poursaeed, John Hopcroft, and Serge Belongie. 2017. Stacked generative adversarial networks. In CVPR, pages 1866--1875. IEEE

    work page 2017

  9. [9]

    Zongcheng Ji, Zhengdong Lu, and Hang Li. 2014. An information retrieval approach to short text conversation. arXiv preprint arXiv:1408.6988

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  10. [10]

    Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, and Jiwon Kim. 2017. Learning to discover cross-domain relations with generative adversarial networks. In ICML, pages 1857--1865

    work page 2017

  11. [11]

    Guillaume Klein, Yoon Kim, Yuntian Deng, Jean Senellart, and Alexander Rush. 2017. Opennmt: Open-source toolkit for neural machine translation. Proceedings of ACL 2017, System Demonstrations, pages 67--72

    work page 2017

  12. [12]

    Alex M Lamb, Anirudh Goyal ALIAS PARTH GOYAL, Ying Zhang, Saizheng Zhang, Aaron C Courville, and Yoshua Bengio. 2016. Professor forcing: A new algorithm for training recurrent networks. In NIPS, pages 4601--4609

    work page 2016

  13. [13]

    Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In NAACL-HLT, pages 110--119

    work page 2016

  14. [14]

    Jiwei Li, Will Monroe, Tianlin Shi, S \.e bastien Jean, Alan Ritter, and Dan Jurafsky. 2017. Adversarial learning for neural dialogue generation. In EMNLP, pages 2157--2169

    work page 2017

  15. [15]

    Chia-Wei Liu, Ryan Lowe, Iulian Serban, Mike Noseworthy, Laurent Charlin, and Joelle Pineau. 2016. How not to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In EMNLP, pages 2122--2132

    work page 2016

  16. [16]

    Zhengdong Lu and Hang Li. 2013. A deep architecture for matching short texts. In NIPS, pages 1367--1375

    work page 2013

  17. [17]

    Thang Luong, Hieu Pham, and Christopher D Manning. 2015. Effective approaches to attention-based neural machine translation. In EMNLP, pages 1412--1421

    work page 2015

  18. [18]

    Lili Mou, Yiping Song, Rui Yan, Ge Li, Lu Zhang, and Zhi Jin. 2016. Sequence to backward and forward sequences: A content-introducing approach to generative short-text conversation. In COLING, pages 3349--3358

    work page 2016

  19. [19]

    Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. 2016. Conditional image generation with pixelcnn decoders. In NIPS, pages 4790--4798

    work page 2016

  20. [20]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311--318

    work page 2002

  21. [21]

    Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. 2017. Automatic differentiation in pytorch. In NIPS-W

    work page 2017

  22. [22]

    Alan Ritter, Colin Cherry, and William B Dolan. 2011. Data-driven response generation in social media. In EMNLP, pages 583--593

    work page 2011

  23. [23]

    Lifeng Shang, Zhengdong Lu, and Hang Li. 2015. Neural responding machine for short-text conversation. In ACL, volume 1, pages 1577--1586

    work page 2015

  24. [24]

    Alessandro Sordoni, Michel Galley, Michael Auli, Chris Brockett, Yangfeng Ji, Margaret Mitchell, Jian-Yun Nie, Jianfeng Gao, and Bill Dolan. 2015. A neural network approach to context-sensitive generation of conversational responses. In NAACL-HLT, pages 196--205

    work page 2015

  25. [25]

    Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In NIPS, pages 3104--3112

    work page 2014

  26. [26]

    Richard S Sutton, David A McAllester, Satinder P Singh, and Yishay Mansour. 2000. Policy gradient methods for reinforcement learning with function approximation. In NIPS, pages 1057--1063

    work page 2000

  27. [27]

    Duyu Tang, Nan Duan, Tao Qin, and Ming Zhou. 2017. Question answering and question generation as dual tasks. arXiv preprint arXiv:1706.02027

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  28. [28]

    Oriol Vinyals and Quoc Le. 2015. A neural conversational model. arXiv preprint arXiv:1506.05869

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  29. [29]

    Hao Wang, Zhengdong Lu, Hang Li, and Enhong Chen. 2013. A dataset for research on short-text conversations. In EMNLP, pages 935--945

    work page 2013

  30. [30]

    Ronald J Williams. 1992. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3-4):229--256

    work page 1992

  31. [31]

    Chen Xing, Wei Wu, Yu Wu, Jie Liu, Yalou Huang, Ming Zhou, and Wei-Ying Ma. 2017. Topic aware neural response generation. In AAAI, pages 3351--3357

    work page 2017

  32. [32]

    Zhen Xu, Bingquan Liu, Baoxun Wang, SUN Chengjie, Xiaolong Wang, Zhuoran Wang, and Chao Qi. 2017. Neural response generation via gan with an approximate embedding layer. In EMNLP, pages 617--626

    work page 2017

  33. [33]

    Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. 2017. Dualgan: Unsupervised dual learning for image-to-image translation. In ICCV, pages 2868--2876. IEEE

    work page 2017

  34. [34]

    Hainan Zhang, Yanyan Lan, Jiafeng Guo, Jun Xu, and Xueqi Cheng. 2018. Reinforcing coherence for sequence to sequence model in dialogue generation. In IJCAI, pages 4567--4573

    work page 2018

  35. [35]

    Hao Zhou, Minlie Huang, Tianyang Zhang, Xiaoyan Zhu, and Bing Liu. 2017. Emotional chatting machine: Emotional conversation generation with internal and external memory. arXiv preprint arXiv:1704.01074

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  36. [36]

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. 2017. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2242--2251. IEEE

    work page 2017

  37. [37]

    Hainan Zhang, Yanyan Lan, Jiafeng Guo, Jun Xu, and Xueqi Cheng. 2018 a . Reinforcing coherence for sequence to sequence model in dialogue generation. In IJCAI, pages 4567--4573

    work page 2018

  38. [38]

    Yizhe Zhang, Michel Galley, Jianfeng Gao, Zhe Gan, Xiujun Li, Chris Brockett, and Bill Dolan. 2018 b . Generating informative and diverse conversational responses via adversarial information maximization. In Advances in Neural Information Processing Systems, pages 1810--1820

    work page 2018