Summary and Contributions: This paper propose a hybrid generation models by integrating the information retrieval strategy (non-parametric memory) with seq2seq model(parametric memory). This model can work on more general knowledge-intensive tasks, such as QA, question generation and fact verification, other than only extractive downstream tasks. It outperforms seq2seq models and task-specific retrieve-and-extract models on three open domain QA tasks. In generation tasks, it can generate more specific, diverse and factual language than a seq2seq baseline model.
Strengths: 1.The motivation of using the the information retrieval strategy to improve the the generation models is sound. The paper did extensive experiments on several Knowledge-Intensive NLP Tasks, which show the effectiveness of the proposed approach.
Weaknesses: Using the the information retrieval strategy to improve the the generation models is straightforward and certainly not novel .Although, the paper did extensive experiments on several Knowledge-Intensive NLP Tasks and showed the effectiveness of the proposed approach. The contribution of the model is not very specific, since that RAG and its components are not proposed by authors.
Clarity: 1.A figure or example about PAG-Sequence Model and PAG-Token Model is needed, which can help people understand more about the difference between these two models. 2.The reason why it need not fine tune document encoder is not clear, which needs experimental result to support it.
Relation to Prior Work: yes
Summary and Contributions: This paper proposes a retrieval-augmented generation framework to enhance the performance of NLP tasks required knowledge. A pre-trained document retriever (DPR) is introduced to conduct the query understanding and document retrieval. A pre-trained seq2seq encoder decoder framework (BART) is employed to conduct the context (query + retrieved documents) understanding and generation. Two information exploiting strategies are designed for the decoder. Experimental results show that the proposed framework can obtain better performance on the tasks of knowledge depended tasks.
Strengths: The proposed idea and approach are interesting, and the experimental results are also very good. The paper is in well-written and easy to follow. We believe that the retrieval+generation framework can be extended to more tasks such as machine translation and dialogue system.
Weaknesses: One tiny weakness is that the core technical components are just borrowed directly from the existing works, such as DPR and BART. And retrieval+extraction/generation framework have also been conducted by some previous works such as Danqi Chen’s work.
Relation to Prior Work: yes
Additional Feedback: This is an interesting work and can benefit the community, and it should be accepted and introduced to public (though the pre-trained version is on arxiv).
Summary and Contributions: This paper presents a new model for NLP tasks that require the use of external knowledge. In particular, the proposed approach combines a non-parametric memory (a dense vector index of Wikipedia) and a parametric memory (a pre-trained seq2seq transformer) to enrich the input query and address the considered task. Experiments conducted on a wide range of different tasks confirm the significance of the proposed approach. Although the method was primarily designed for open-domain question answering, it show nice results also on related NLP tasks, such as fact checking. *** AFTER THE REBUTTAL *** I thank the authors for the answers provided in the rebuttal.
Strengths: + Novelty of the approach + Extensive experimental evaluation + Relevance to the NeurIPS community
Weaknesses: - The description of the model is quite concise (due to space restrictions) - Reproducibility of the results (link to code and models is hidden for double blind review, so it is hard to judge reproducibility)
Correctness: The methodology is solid and correct
Clarity: The paper is very well written, just concise in the part of model description
Relation to Prior Work: The paper is well positioned against the state-of-the-art, but more experimental comparison would have strengthened the paper
Additional Feedback: * Although the experimental section is very rich, I wonder whether it would have been possible to include more comparisons with some memory-based neural network. For example, one could think of applying a generator on top of a standard Memory-Augmented Neural Network, or some other similar approach. * In Section 2.4, it is stated that the model is trained "without any direct supervision on what document should be retrieved". I wonder whether it would be possible also to train the model with such an information, in the same way memory networks use strong supervision. Typos: - "can't" -> "cannot" - In footnote 1 there is a repetition of "available" - Pag. 4, "answers are extracted spans from retrieved documents" -> remove "spans"? - Pag. 5, "quuestion" -> "question"
Summary and Contributions: This paper proposes a retrieval augmented seq2seq model for question answering and related knowledge-intensive NLP tasks. The model is combination of a pre-trained BART and a dense passage retriever via joint probabilistic model. Two specific formulations, referred to as RAG-Sequence and RAG-Token, are proposed to let the model select relevant document(s) to generate answers. Experiments are conducted on a range of tasks including open-domain question answering and fact verification, showing that the RAG model achieve state-of-the-art or competitive performances. The design of the model share some similarity with REALM model, which is also a retrieval augmented encoder-only model. REALM can be pre-trained on unsupervised documents and then fine-tuned for extractive question answering tasks. The RAG model proposed in this paper have an encoder-decoder structure which makes the model more flexible to handle a broader range of tasks. However, the retriever and seq2seq module of RAG model are pre-trained in a separated manner and only joint learned in the fine-tuning stage.
Strengths: The RAG model proposed in this paper benefits from both the parametric and non-parametric model and provide a promising framework for the open-domain question answering and knowledge-related NLP tasks. Although the idea may not be totally novel, it is the first model that integrates seq2seq pre-trained model with the neural retrieval model, which can be trained in an end-to-end fashion. The performance of the model is strong on several question answering tasks, proving it as a general framework for these tasks. This paper should be of great interest to the NeruIPS community, especially for researchers in NLP and IR field.
Weaknesses: The whole framework is in some sense the encoder-decoder version of the REALM model, while the two components (BART and DPR) cannot be pre-trained end-to-end, which prevent their synergy in learning the knowledge in the pre-training stage. For generation tasks, it would be more interesting to compare RAG with GPT-2 or T5 models rather than BART.
Clarity: The paper is well organized and easy to follow.
Relation to Prior Work: Yes