Building Task-Oriented Visual Dialog Systems Through Alternative Optimization Between Dialog Policy and Language Generation

Josh Arnold; Mingyang Zhou; Zhou Yu

REVIEW 4 major objections 5 minor 18 references

Building Task-Oriented Visual Dialog Systems Through Alternative Optimization Between Dialog Policy and Language Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Moving reinforcement learning from word generation to image guessing yields state-of-the-art GuessWhich performance on both task completion and dialog quality.

desk verdict A sensible decoupling idea for dialog RL, but the policy-improvement details are too vague to reproduce. read the letter →

arxiv 1909.05365 v2 pith:S3XL6JDN submitted 2019-09-06 cs.CL cs.AIcs.LG

Mingyang Zhou , Josh Arnold , Zhou Yu This is my paper · ORCID

classification cs.CLcs.AIcs.LG

keywords visualdialogGuessWhichreinforcementlearningtask-orientedimageguessingpolicylanguagegenerationalternativeoptimizationsupervised

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to fix a known failure of reinforcement learning for task-oriented visual dialog: when RL optimizes word-by-word question generation, the agent drifts into unnatural language. It proposes a curriculum that splits the job in two—RL chooses which image to guess at each turn, while a supervised sequence-to-sequence model generates the questions—and alternates the two objectives during fine-tuning. On GuessWhich, the resulting agent (RL-Q-IG) reaches a percentile mean rank of 96.81% with a perplexity of 54.66, and in human games it wins more often and is rated higher on fluency, relevance, comprehension, and diversity than a word-level RL baseline. The point is that task completion and natural language are not in tension when the policy action space is decoupled from the language action space.

What carries the argument

The load-bearing mechanism is an alternating optimization schedule between two objectives. The RL objective trains the policy $\pi_{\theta_r,\theta_g}(i_t|s_t)$, which samples a candidate image from the top-$K$ images nearest to the dialog state $s_t$ in Euclidean distance; the reward is the ranking percentile of the target image, and the policy is updated by cross-entropy toward an alternative action $i^*_t$ with higher estimated value $Q^\pi(s_t,i^*_t)$, following policy improvement theory. The supervised objective is the joint loss $\alpha\sum_{t}\log p(q_t|s_t)+\beta\sum_{t}\operatorname{MSE}(z_{tgt},s_t)$ from pre-training, which keeps the question decoder and response encoder anchored to the human dialog distribution. Every other epoch, the two objectives alternate, so the policy explores task-completing actions while the language model does not degenerate.

What would settle it

Train the same SL-Q-IG model but replace the RL update in Section 3.3.2 with a deterministic update that always chooses the top-K image closest to the dialog state, and keep the alternating supervised objective unchanged; if the percentile mean rank stays near 96.81%, then the RL policy-improvement mechanism is not what produces the reported gain.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that the action space of reinforcement learning should be the image guess, not the output vocabulary. By treating the sequence of image guesses as a partially observable Markov decision process, the agent can learn an optimal question-asking strategy from a reward signal—the ranking percentile of the target image—while the question decoder remains supervised and therefore stays close to human language. The authors further show that alternating the RL objective with the supervised language objective prevents the shared response encoder from drifting away from the human dialog distribution; the ablation RL-Q-IG-NA, trained without the alternating language objective, achieves only 0.07% higher PMR but its perplexity jumps to 363.88. In human-AI games, the full framework maintains its win rate under human-generated captions (67.6%) where the supervised pre-trained agent drops to 40.6%, indicating the RL policy helps in unseen scenarios.

Load-bearing premise

The load-bearing premise is that, at nearly every turn, the agent can actually find a different candidate image that its own scoring says is better than the image it would otherwise guess; if that search fails, the task-completion gains attributed to reinforcement learning would not materialize.

Editorial extensions

If this is right

Task completion and language quality do not have to be traded off: separating RL on image guesses from supervised question generation yields both high percentile mean rank and low perplexity.
Word-level RL is the source of language degeneration in this setting: the RL-Q baseline's perplexity explodes to 544.97, while RL-Q-IG stays at 54.66 by keeping the decoder supervised.
The RL policy is most valuable in unseen scenarios: with human-generated captions, RL-Q-IG holds a 67.6% win rate while the supervised pre-trained agent drops to 40.6%.
An agent can be trained to ask better questions by rewarding the image-guessing trajectory only, without shaping rewards for individual words.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper reports one alternating schedule (every other epoch); varying the frequency of the supervised objective would map a task-completion versus language-quality trade-off curve and would test whether the schedule itself, rather than the action-space separation, is responsible for the gains.
The paper leaves the policy-improvement step at the level of finding an alternative image guess with higher estimated value; making that search explicit would let the same update transfer to other interactive retrieval settings, such as product search or image recommendation, where an agent asks questions and ranks candidate items.
A cheaper check on the RL contribution is to replace the policy-improvement update with a deterministic nearest-neighbor guess while keeping the alternating supervised objective; if percentile mean rank remains near 96.81%, the reported improvement is attributable to the action-space reformulation rather than to the RL update itself.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The short version: this paper has the right instinct — separate the RL policy from the language generator — but it leaves the load-bearing RL step vague enough that I couldn't reimplement it from the text. That's a real weakness, not a fatal one.

What's new: instead of running RL over the vocabulary, they run it over the image-guessing action space (top-K nearest images to the dialog state), then alternate that objective with the supervised language-model loss. That's a clean way to avoid the language degeneration you see in word-level RL (Das et al.). The idea builds on Guo et al.'s iterative retrieval, but the alternating curriculum is their own.

The experiments are more thorough than most: they compare against word-level RL, the SL-only version, and two ablations (RL without the language loss, and word-level RL on top of their architecture). The PMR numbers are consistent with the claim, and the perplexity numbers show the alternating loss prevents the language drift that RL-Q-IG-NA suffers. The human eval, while small and without significance tests, shows the full model wins more games and rates better on relevance and comprehension.

Now the soft spots. Section 3.3.2 is the biggest one. They say they 'explore a different action i*_t' with higher Q-estimate, then train toward it with cross-entropy, but they never specify how Q is estimated, how many candidates are explored, or how rollouts are done. That's not a side detail — it's the core of the claimed RL improvement. Without it, the PMR gain from RL (96.09 to 96.81) could come from the extra epochs or the answer-bot behavior, not from the policy-improvement mechanism they describe. Second, the reward is the ranking percentile of the target image, which is exactly the PMR metric they report, so the PMR result is partly tautological. That's not a fatal flaw — it's fine to optimize what you care about — but they shouldn't frame PMR as an independent finding. The dialog quality improvements are not forced by the reward, so those are the more convincing evidence.

Citations are fine; they build on Das et al. and Guo et al. without overselling novelty. Math and data are otherwise solid, though code would help.

I'd send this to a serious referee. The idea deserves attention, the experiments are decent, and the main fix is a precise description of the RL step plus code or pseudo-code. Without that, the paper is more of a proof-of-concept than a reproducible method.

Should we read it? Yes, for a discussion of action spaces in dialog RL. I'd cite it for the decoupling idea if I were working in that area.

Referee Report

4 major / 5 minor

Summary. The paper proposes a framework for task-oriented visual dialog that separates dialog-policy learning from language generation. The Q-bot is first trained with a supervised objective combining question generation and image-feature prediction; it is then fine-tuned by alternating an RL objective over an image-guessing action space (the top-K nearest images to the dialog state vector) with the supervised language-generation objective. On the GuessWhich benchmark the authors report that their model, RL-Q-IG, reaches 96.81% PMR and 54.66 perplexity, outperforming word-level RL and supervised baselines, and that it also wins more human-AI games and receives higher human quality ratings, especially when the game is initialized with human-generated captions. Ablations with RL applied only to policy learning (RL-Q-IG-NA) and with word-level RL (RL-Q-IG-W) are used to isolate the contributions of the alternating objective and the reduced action space.

Significance. If the method is reproducible as described, the core idea is valuable: applying RL to a compact, semantically meaningful action space (the image-guessing action) and decoupling it from token-level generation is a plausible way to avoid the language degeneration commonly observed in word-level RL dialog systems. The paper also provides a useful ablation study showing that the alternating supervised objective substantially improves perplexity with only a negligible PMR cost, and the human evaluation goes beyond fixed AI-AI games by including human-generated captions. The main weaknesses are that the RL policy-improvement procedure is not specified at the level needed to reproduce the result, the PMR reward and PMR evaluation metric overlap, and several headline claims in the human evaluation lack statistical support.

major comments (4)

[Sec. 3.3.2, Eq. (2)] The policy-improvement step is underspecified in a way that is load-bearing for the central claim. The paper does not state how Q_pi(s_t, i) is estimated for actions not actually taken: is a separate value network trained, are rollouts simulated, or is some heuristic used? It also does not describe how candidate actions i*_t are explored, how many rollouts are averaged, or how the 'larger Q' action is selected. Moreover, updating the policy by cross-entropy toward one higher-Q action is not, by itself, a policy-improvement step in the sense of the policy-improvement theorem, which requires the new policy to be the maximizing action (or a distribution with sufficient coverage over actions). Without these details, the reported PMR gain from 96.09% (SL-Q-IG) to 96.81% (RL-Q-IG) cannot be attributed to the claimed RL mechanism rather than to additional epochs, supervised fine-tuning, or an ad hoc exploration heuristic.
[Sec. 3.3.2 (Rewards) and Table 2] The reward function is defined as the ranking percentile of the target image with respect to the dialog state vector, and the PMR metric reported in Table 2 is the same quantity averaged over the test rounds. The PMR improvement of the RL-trained model over its supervised counterpart is therefore a direct measure of how well the model optimizes its own training reward, not an independent measure of task-completion quality. Please state this overlap explicitly, and preferably supplement PMR with additional task-completion metrics (e.g., retrieval accuracy, success@k, or human win rate) so that the evaluation is not entirely aligned with the training objective.
[Sec. 5.2, Tables 3 and 4] The claim that 'RL-Q-IG significantly outperforms the baseline RL agent in all criteria' is not supported by any significance test, confidence interval, or inter-annotator agreement statistic. The human-generated-caption condition is based on only 30 games per setting, and the win-rate difference between RL-Q-IG and SL-Q-IG in Table 3 is modest (67.5% vs 62.7%) without error bars. Please add appropriate statistical tests (e.g., bootstrap or paired tests over games and/or evaluators) or temper the word 'significantly'.
[Sec. 4 and Eq. (1)] Several training hyperparameters that likely affect the reported results are missing: alpha and beta in Eq. (1), the top-K size of the image action space, the discount factor gamma, the number of RL epochs, and the exact alternating schedule between the RL and SL objectives. Since no code or checkpoints are released, the textual description is the only specification of the method, and these omissions make exact re-implementation difficult.

minor comments (5)

[Introduction] There is a typo in the phrase 'various image guessing tasks such us GuessWhich'; it should be 'such as'.
[Sec. 3.1] The dialog-history notation [q1, a2, ..., qt, at] appears to contain an indexing typo; presumably it should be [q1, a1, ..., qt, at].
[Sec. 5.2 and Appendix B] The text refers to 'Table 9' when discussing dialog examples, but Table 9 appears only in Appendix B; the table numbering should be made consistent or the reference clarified.
[Eq. (1)] Equation (1) uses n both as the dialog length and as the upper limit of the summation over turns; please define the indexing convention clearly.
[Sec. 4.4] The human evaluation section would benefit from reporting the number of ratings collected per criterion, the agreement between evaluators, and whether the evaluators were blind to model identity in the rating tasks.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reward and perplexity metrics align with the training objectives, but this is standard objective-metric alignment, not a hidden derivation.

full rationale

The paper's claimed derivation chain is an empirical training pipeline rather than a formal derivation from first principles. The RL reward in Sec. 3.3.2 is the ranking percentile of the target image, and PMR is the same ranking percentile, so RL directly optimizes the reported task metric; similarly, the supervised language-model loss is log-likelihood and perplexity is its exponential. This is transparent objective-metric alignment, not a case where a fitted parameter is renamed as a prediction: no quantity is simultaneously used as both input and output in a way that would force the reported result. The RL policy-improvement step is underspecified (no explicit Q-estimator or exploration rule is given), which is a reproducibility risk, but that is not circularity. Baselines are external (Das et al. 2017b), human evaluation provides independent evidence of dialog quality, and self-citations appear only in related work and are not load-bearing. Therefore no significant circularity is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a small set of training hyperparameters and domain assumptions about the simulator and reward signal. No new physical or conceptual entities are introduced. The main unstated load is in the RL policy improvement algorithm, which is not specified in enough detail to fully audit.

free parameters (5)

alpha (Eq. 1) = not reported
Weight for supervised question-generation loss in the joint SL objective; chosen by hand, no value or tuning details given.
beta (Eq. 1) = not reported
Weight for image-feature prediction MSE loss in the joint SL objective; chosen by hand.
top-K action space size = not reported
Number of nearest images used as the RL action space; controls exploration/exploitation; no value given.
discount factor gamma = not reported
Discount factor for RL return in the reward definition; no value given.
SL epochs and RL epochs = 40 and 20
Training schedule; set to 40 supervised epochs and 20 RL fine-tuning epochs by hand.

assumptions (4)

domain assumption The A-Bot trained on VisDial is a sufficiently good simulator for RL training of the Q-Bot.
RL fine-tuning uses the answer bot as the environment; if the simulator is not representative of real users, the policy may not transfer, though human eval partially mitigates this.
domain assumption Ranking percentile of the target image is a valid reward signal for task success.
The reward is defined as the target image's rank among candidates based on dialog state; this encodes the task objective but may be a noisy reward at early turns.
standard math Standard deep learning and RL background: seq2seq, LSTM, VGG-16 features, policy improvement.
The method relies on standard neural architectures and RL machinery; no proofs are provided.
domain assumption The VisDial dataset is an adequate training distribution for language generation.
Supervised language generation is trained on VisDial; the paper acknowledges that caption-only guessing already achieves high PMR, implying limited headroom.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Building Task-Oriented Visual Dialog Systems Through Alternative Optimization Between Dialog Policy and Language Generation." pith.science (2026). https://pith.science/paper/S3XL6JDN

@misc{pith2026190905365,
  author       = {Pith},
  title        = {Pith review of: Building Task-Oriented Visual Dialog Systems Through Alternative Optimization Between Dialog Policy and Language Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S3XL6JDN}},
  note         = {Machine review of arXiv:1909.05365}
}

read the original abstract

Reinforcement learning (RL) is an effective approach to learn an optimal dialog policy for task-oriented visual dialog systems. A common practice is to apply RL on a neural sequence-to-sequence (seq2seq) framework with the action space being the output vocabulary in the decoder. However, it is difficult to design a reward function that can achieve a balance between learning an effective policy and generating a natural dialog response. This paper proposes a novel framework that alternatively trains a RL policy for image guessing and a supervised seq2seq model to improve dialog generation quality. We evaluate our framework on the GuessWhich task and the framework achieves the state-of-the-art performance in both task completion and dialog quality.

Figures

Figures reproduced from arXiv: 1909.05365 by the authors.

**Figure 2.** The percentile mean rank (PMR) over the 5- [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** The web interface for human-AI guessing game. The left image is a target image randomly sampled [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: The user interface for human relevancy experiment. The conversation from the three different models on [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 11 canonical work pages

[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...
[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...
[3]

Prithvijit Chattopadhyay, Deshraj Yadav, Viraj Prabhu, Arjun Chandrasekaran, Abhishek Das, Stefan Lee, Dhruv Batra, and Devi Parikh. 2017. http://arxiv.org/abs/1708.05122 Evaluating visual conversational agents via cooperative human-ai games . CoRR, abs/1708.05122

work page Pith review arXiv 2017
[4]

Moura, Devi Parikh, and Dhruv Batra

Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jos\'e M.F. Moura, Devi Parikh, and Dhruv Batra. 2017 a . V isual D ialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

work page 2017
[5]

Moura, Stefan Lee, and Dhruv Batra

Abhishek Das, Satwik Kottur, Jos\'e M.F. Moura, Stefan Lee, and Dhruv Batra. 2017 b . Learning cooperative visual dialog agents with deep reinforcement learning. In Proceedings of the IEEE International Conference on Computer Vision (ICCV)

work page 2017
[6]

Xiaoxiao Guo, Hui Wu, Yu Cheng, Steven Rennie, and Rog \' e rio Schmidt Feris. 2018. http://arxiv.org/abs/1805.00145 Dialog-based interactive image retrieval . CoRR, abs/1805.00145

work page Pith review arXiv 2018
[7]

Dauphin, Devi Parikh, and Dhruv Batra

Mike Lewis, Denis Yarats, Yann N. Dauphin, Devi Parikh, and Dhruv Batra. 2017. http://arxiv.org/abs/1706.05125 Deal or no deal? end-to-end learning for negotiation dialogues . CoRR, abs/1706.05125

arXiv 2017
[8]

Jiwei Li, Will Monroe, Alan Ritter, Michel Galley, Jianfeng Gao, and Dan Jurafsky. 2016. http://arxiv.org/abs/1606.01541 Deep reinforcement learning for dialogue generation . CoRR, abs/1606.01541

arXiv 2016

Show all 18 references

[9]

Antoine Raux, Brian Langner, Dan Bohus, Alan W Black, and Maxine Eskenazi. 2005. Let’s go public! taking a spoken dialog system to the real world. In in Proc. of Interspeech 2005

2005
[10]

Courville, and Joelle Pineau

Iulian Vlad Serban, Alessandro Sordoni, Yoshua Bengio, Aaron C. Courville, and Joelle Pineau. 2015. http://arxiv.org/abs/1507.04808 Hierarchical neural network generative models for movie dialogues . CoRR, abs/1507.04808

2015 arXiv
[11]

Weiyan Shi and Zhou Yu. 2018. http://arxiv.org/abs/1804.10731 Sentiment adaptive end-to-end dialog systems . CoRR, abs/1804.10731

2018 arXiv
[12]

Simonyan and A

K. Simonyan and A. Zisserman. 2015. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations

2015
[13]

Sutton and Andrew G

Richard S. Sutton and Andrew G. Barto. 1998. Introduction to Reinforcement Learning, 1st edition. MIT Press, Cambridge, MA, USA

1998
[14]

Courville

Harm de Vries, Florian Strub, Sarath Chandar, Olivier Pietquin, Hugo Larochelle, and Aaron C. Courville. 2016. http://arxiv.org/abs/1611.08481 Guesswhat?! visual object discovery through multi-modal dialogue . CoRR, abs/1611.08481

2016 arXiv
[15]

Williams and Steve Young

Jason D. Williams and Steve Young. 2007. https://doi.org/10.1016/j.csl.2006.06.008 Partially observable markov decision processes for spoken dialog systems . Comput. Speech Lang., 21(2):393--422

2007 doi
[16]

Jiaping Zhang, Tiancheng Zhao, and Zhou Yu. 2018. http://arxiv.org/abs/1805.03257 Multimodal hierarchical reinforcement learning policy for task-oriented visual dialog . CoRR, abs/1805.03257

2018 arXiv
[17]

Junjie Zhang, Qi Wu, Chunhua Shen, Jian Zhang, Jianfeng Lu, and Anton van den Hengel. 2017. http://arxiv.org/abs/1711.07614 Asking the difficult questions: Goal-oriented visual question generation via intermediate rewards . CoRR, abs/1711.07614

2017 arXiv
[18]

Tiancheng Zhao, Kaige Xie, and Maxine Esk \' e nazi. 2019. http://arxiv.org/abs/1902.08858 Rethinking action spaces for reinforcement learning in end-to-end dialog agents with latent variable models . CoRR, abs/1902.08858

2019 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[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...

[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...

[3] [3]

Prithvijit Chattopadhyay, Deshraj Yadav, Viraj Prabhu, Arjun Chandrasekaran, Abhishek Das, Stefan Lee, Dhruv Batra, and Devi Parikh. 2017. http://arxiv.org/abs/1708.05122 Evaluating visual conversational agents via cooperative human-ai games . CoRR, abs/1708.05122

work page Pith review arXiv 2017

[4] [4]

Moura, Devi Parikh, and Dhruv Batra

Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jos\'e M.F. Moura, Devi Parikh, and Dhruv Batra. 2017 a . V isual D ialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

work page 2017

[5] [5]

Moura, Stefan Lee, and Dhruv Batra

Abhishek Das, Satwik Kottur, Jos\'e M.F. Moura, Stefan Lee, and Dhruv Batra. 2017 b . Learning cooperative visual dialog agents with deep reinforcement learning. In Proceedings of the IEEE International Conference on Computer Vision (ICCV)

work page 2017

[6] [6]

Xiaoxiao Guo, Hui Wu, Yu Cheng, Steven Rennie, and Rog \' e rio Schmidt Feris. 2018. http://arxiv.org/abs/1805.00145 Dialog-based interactive image retrieval . CoRR, abs/1805.00145

work page Pith review arXiv 2018

[7] [7]

Dauphin, Devi Parikh, and Dhruv Batra

Mike Lewis, Denis Yarats, Yann N. Dauphin, Devi Parikh, and Dhruv Batra. 2017. http://arxiv.org/abs/1706.05125 Deal or no deal? end-to-end learning for negotiation dialogues . CoRR, abs/1706.05125

arXiv 2017

[8] [8]

Jiwei Li, Will Monroe, Alan Ritter, Michel Galley, Jianfeng Gao, and Dan Jurafsky. 2016. http://arxiv.org/abs/1606.01541 Deep reinforcement learning for dialogue generation . CoRR, abs/1606.01541

arXiv 2016