NeurIPS 2020

A Simple Language Model for Task-Oriented Dialogue

Review 1

Summary and Contributions: This paper is about using a unified language model to build task-oriented dialog systems

Strengths: The paper has compared with many recent dialog work.

Weaknesses: Mostly is the innovation is not that huge. It is hard to see how much improvement this is compared to Sololist and ARDM

Correctness: The method looks correct

Clarity: Yes

Relation to Prior Work: Missing reference: SOLOIST: Few-shot Task-Oriented Dialog with A Single Pre-trained Auto-regressive Model Qingyang Wu, Yichi Zhang, Yu Li, and Zhou Yu. 2019. Alternating recurrent dialog model with large-scale pre-trained language models

Reproducibility: Yes

Additional Feedback:

Review 2

Summary and Contributions: This paper proposed a single end-to-end model SimpleTOD for task-oriented dialog by leveraging existing pre-trained models. Empirical results on the MultiWOZ dataset demonstrated that the proposed methods had outperformed all prior methods for DST tasks and dialog act/response generation in the e2e setting.

Strengths: Leveraging pre-trained models to solve problems in task-oriented dialogues has become a trend in recent years. By taking advantage of pre-trained models, this paper achieves SOTA performance in the DST task on the MultiWOZ dataset.

Weaknesses: • This paper did not mention SOLOIST, which has achieved SOTA in policy optimization and the end2end setting for MultiWOZ 2.0, as illustrated in the MultiWOZ repository.  • Although SimpleTOD produces SOTA results for the DST task on MultiWOZ, the gain over the previous SOTA model is small.  • No human evaluation is provided for response generation. • Since there are existing prior and concurrent works that leverage pre-trained models like GPT-2, based on my understanding, the primary differentiator (novelty) in this paper is limited. • If there is an experiment on datasets other than MultiWOZ, the paper's quality will be better.

Correctness: The claims and methods seem correct.

Clarity: This paper is well written in general. But I think more discussions should be made to compare the approach with other papers that employ pre-trained models.

Relation to Prior Work: [6][17] uses similar idea as this paper, but the differences are not well discussed. Besides, the paper 'SOLOIST' was never mentioned in the paper.

Reproducibility: Yes

Additional Feedback: ------------------------------------------------------------------------------------------ A response after the rebuttal period: After reading authors' reply, I agree that aforementioned papers are considered concurrent work. With human evaluation added in the final version as the authors mentioned, I think this work should be good.

Review 3

Summary and Contributions: Post AR: I see the other reviewers mostly are concerned with novelty in the face of SOLOIST and DAMD. I emphasize the contemporaneousness of SOLOIST and feel satisfied with discussion of DAMD. I still very much want to see this published. The authors consider the task of task-oriented dialogue on multiwoz, where an assistant needs to help a user find a train/restaurant/whatever in the city center of Cambridge. The authors propose a fully end-to-end model, in which the entire task is structured as sequence prediction using a transformer. The authors report state of the art results on multiwoz, while previous end-to-end models tend to perform worse. The authors provide some analysis of several aspects of it (special tokens, pretraining ablation, oracle belief state, etc) to help clarify source of the results. Their model is also conceptually much simpler than most of the prior attempts.

Strengths: I found the work quite compelling, and would like to see it published. With respect to Ham et al., the analysis and state of the art results are primary areas of novelty, but these two works MAY be considered contemporary, and so they may be viewed as sharing the core contribution of applying GPT-2 to task oriented dialogue for both response generation AND belief state tracking. One particularly strong aspect of this paper is that I read it on arxiv a month or two ago, and our lab was interested in the results, so we internally replicated them. We found that we got similar numbers without too much effort, and that the results carried over with other pretraining sources too. To this end, the paper is more sound than the VAST majority of NeurIPS papers: in a short amount of time, a third party was able to replicate the results using only the description from the paper.

Weaknesses: - Only use one dataset, but many task oriented datasets exist. Another dataset would strengthen this model a bit. - Weak differences with Ham et al. or Soloist. All three works should probably be perceived as contemporaneous, and I don't think should detract from the novelty of this work, as they came out within 30 days of each other. I would rather the authors include detailed explanations to help explain the primary differences, for the benefit of future readers.

Correctness: Yes. As emphasized in Strengths, my lab has even replicated these results, indicating that they are very robust.

Clarity: Yes, I find the paper mostly easy to read. The explanation of MultiWoZ may be unclear to someone unfamiliar with the details of the dataset, especially around some of the aspects of the database results, etc. The authors may consider finding a way to make Figure 1 more central in the text to make this part clear. The authors would also benefit greatly from putting Figure 2 of the appendix in the main text, as that makes the special tokens analysis much easier to understand.

Relation to Prior Work: The only problematic one is the differentiation with Ham et al., which does pretty much the same thing, but got worse results. The authors should more clearly delineate the differences between their work, and attempt to explain the performance differences. Otherwise, prior work is reasonably covered. They drop some potential related work, (for example, pretrained on reddit for dialogue a year before Henderson did), but the bibliography is reasonable.

Reproducibility: Yes

Additional Feedback: I need to say that I read this paper on arxiv a month or two before reviewing it for NeurIPS, so this is not a double blind review. However, I feel it's a good paper, and that my lab has internally reproduced it already lends it strong credibility. I feel that the authors would benefit from dropping section 3.3 and using the extra space to include a copy of the Figure 2 from the appendix.

Review 4

Summary and Contributions: The authors propose SimpleTOD, which can replace modular task-oriented dialogue models to unified causal language model in an end-2-end manner. There are three sub-tasks in the task-oriented dialogue. They are dialogue state tracking, action prediction, and response generation. SimpleTOD treats all three sub-tasks as sequence generation. Whole up to dialogue context C_t is used as the first input to the model, and the model generates dialogue state B_t at turn t. The dialogue state B_t is used for database search. The DB search results are (domain, slot, value) triplets, but they use aggregated result D_t that includes the number of rows satisfying the conditions and booking information. And then, the model receives concatenated C_t, B_t, and D_t as input to generate action A_t. At last, all the conditions are combined to generate system response S_t. The experimental results show that SimpleTOD achieves state-of-the-art performance in the MultiWOZ dataset. Especially their model is evaluated below various settings regarding noisy label-cleaning in the dialogue state tracking task. They argue their model is robust to noisy annotations. In action and response generation benchmark, It also achieves better performance than previous SOTA models. They emphasize the importance of usage of oracle information in the end-2-end evaluation. So, They compare their model with only one model DAMD, which doesn’t use oracle information except for DB search results. Additional analysis shows interesting observations. Especially the usage of special tokens to the input is a very important factor in this case.

Strengths: They propose a simple unified framework to solve task-oriented dialogue with an end-to-end manner. It is convincing all sub-tasks are cascaded continuously and exchange positive signals to each other. From this point of view, their choice on causal language modeling as their framework is very natural and a good point. It could reduce efforts to train each component for task-oriented dialogue. They evaluate their model on truly end-2-end setting without any oracle information for the first time. The model achieves state-of-the-art performance in the multi-domain task-oriented dialogue dataset.

Weaknesses: Their contribution lies in simplicity. It is good applications however fundamentally not novel. For more convincing results, The author should argue the modular pipeline problems and show the effectiveness of the end-to-end model compared with a seperated model more precisely. With this comparison, I want to know more about the advantages of multi-task learning approach and correlations between the sub-tasks since they insist on inherent dependencies between the sub-tasks. However, this analysis is missing decisively.

Correctness: Their method is simple and clear.

Clarity: The overall description of the causal language model is well written. However, its formal representation regarding DB search result D_t and each special tokens are ambiguous. It is confusing that they described their own transformer models too simple, but their main results are based on the pre-trained GPT-2 model (There are differences between these models e.g., activation function (ReLU vs. Gelu), position encoding (sinusoidal vs. learned)).

Relation to Prior Work: There is a lack of explanation for the baseline model. For fair comparison regarding robustness on the noisy annotation in the dialogue state tracking, they should show it on other models too. When it comes to end-to-end evaluation, they described it well except for precise analysis on the oracle information.

Reproducibility: Yes

Additional Feedback: Regarding oracle database search, I don’t fully understand their analysis. Especially DAMD+augmentation outperforms other models when it uses two oracle information compared with not using that information in Table 7. Even though there are noisy annotations to search result, It is not a scalable setting to expand new domains or new DB instances. I think it needs more explorations why SimpleTOD model is vulnerable from the oracle information compared with other baseline models in Table 7. It would have been better if the inference speed and performance trade-off were also compared when choosing pre-trained model. It is Because autoregressive generation has time-consuming natures, especially with long sequence length. ------------------------------------------------------------------------------------------ A response after the rebuttal period: Thank you for responding to the review. I have agreed that it is impressive to show that a simplified input scheme with a simple generative model outperforms all previous models on DST. I believe the paper is worthy of being reported to NeurIPS community. ------------------------------------------------------------------------------------------