Repetition In Repetition Out: Towards Understanding Neural Text Degeneration from the Data Perspective

There are a number of diverging hypotheses about the neural text degeneration problem, i.e., generating repetitive and dull loops, which makes this problem both interesting and confusing. In this work, we aim to advance our understanding by presenting a straightforward and fundamental explanation from the data perspective. Our preliminary investigation reveals a strong correlation between the degeneration issue and the presence of repetitions in training data. Subsequent experiments also demonstrate that by selectively dropping out the attention to repetitive words in training data, degeneration can be significantly minimized. Furthermore, our empirical analysis illustrates that prior works addressing the degeneration issue from various standpoints, such as the high-inflow words, the likelihood objective, and the self-reinforcement phenomenon, can be interpreted by one simple explanation. That is, penalizing the repetitions in training data is a common and fundamental factor for their effectiveness. Moreover, our experiments reveal that penalizing the repetitions in training data remains critical even when considering larger model sizes and instruction tuning.

Paper

Similar papers

Peer review

Reviewer Pdys7/10 · confidence 4/52023-06-13

Summary

This paper studies the cause of neural text degeneration, i.e. language models tend to generate repetitive loops. They design an experiment showing that text degeneration is correlated with the amount of repetitive text in the training data. Motivated by this finding, they propose repetitive dropout, which applies dropout on the attention weights over repetitive context. They experiment with 3 datasets and demonstrate that their proposed method can greatly reduce repetition. They also re-inspect previous hypotheses: 1. high in-flow words: they show that merging repetitive words contributes to a large portion of the effect of merging high in-flow words. 2. the maximum likelihood objective: they show that even when the model is trained with the maximum likelihood objective, as long as repetitions are penalized, the model does not degenerate much. Thus, the maximum likelihood objective is not the cause of neural text degeneration. 3. self-reinforcement of repetition: the argue that a. It is also caused by the repetitive words in the training data. b. The self-reinforcement loops are broken by their proposed method. Finally, they categorize repetitions into three categories and study their frequency along with their effect on models’ degeneration behavior. They find that the *theme* category has the greatest effect.

Strengths

1. Most parts of the paper are easy to follow. 2. Neural text degeneration has been discovered since 2019, but its cause is still unclear. Finding the root cause may interest audiences in this field. 3. Despite the hypothesis being simple, as far as I know, it hasn’t been studied. The experiment design in Section 4 is clever. 4. They proposed a simple method that can reduce repetition effectively. 5. They inspected a few previous hypotheses and the results are aligned with their hypothesis (in general). 6. They inspected the effect of three types of repetitions in the training data, which in my opinion is insightful. In sum, in my opinion, their analyses of the cause of neural text degeneration is comprehensive and the conclusion is convincing.

Weaknesses

In my opinion, because the main message of this paper is clear enough, the following issues may not be very crucial. 1. The authors only compare their method with one baseline in Table 1. There are other mitigations for text degeneration, e.g. [14, 22, 11] cited in the paper. 2. Section 6.2 is relatively hard to follow. 1. Line 248, “The core idea of many previous works is to penalize a specific set of data to alleviate the degeneration.” is vague, though I can roughly figure out what it means after reading the following parts. 2. Having a brief introduction for the definition of high-inflow words will make the paper more self-contained. 3. For the Likelihood Objective part, the paragraph at line 283 needs a clearer topic sentence. The connection between the experiment design in paragraph at line 288 and the argument in the previous paragraph is not very clear to me either. 3. At Line 312, the author mentioned “we find that the model trained by repetition dropout can break the self-reinforcement loop”. Some evidence should be provided. 4. At Line 223, “we hypothesize that the exposure bias issue is an important factor for the left repetition, because we find that lower-order n -gram2…”, I couldn't understand the explanation. Some details should be provided too. repetitions generally appear earlier than higher-order repetitions 4. Given the popularity of large language models, the study on GPT-2-sized models may be less useful for the general audience.

Questions

1. If repetition in the training data is the cause of degeneration, why not simply remove those repetitive text from the training data? This seems to me a more straightforward method than applying attention dropout. 2. If the authors promise to address the issues in the weakness section above, I would be happy to increase my score, because I think the main message is clear and interesting enough. 3. Having some analyses on large language models would make this work more impactful, e.g. 1. If I understood correctly, the analyses in Section 6.3 can also be done for LLMs. 2. I also wonder what would be an effective way to fine-tune an existing (large) model so it doesn't generate repetitive text.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations are discussed in Section 7.

Reviewer jpnv6/10 · confidence 3/52023-07-06

Summary

In this paper, the authors demonstrate that repetition in the training data is a major cause of the neural text repetition problem. They first show a strong correlation between the repetition ratio of training data and generated text. Based on the observation, they propose repetition dropout to prohibit the model from learning repetition in data during training. Experimental results show that repetition dropout significantly addresses the repetition problem compared to previous approaches and penalizing repetition in training data is a key factor for reducing the problem. In analysis, the authors demonstrate that their method specifically reduces repetition caused by subject matter rather than grammar or highly frequent phrases.

Strengths

- The proposed method and supporting experiments are well-motivated and the findings are interesting. - They significantly reduce the neural text repetition problem compared to previous work.

Weaknesses

- To demonstrate that the distribution of generated texts is close to that of human texts, authors could further utilize metrics such as MAUVE [1]. - Experiments on various parameter sizes would be beneficial for further understanding such as "Does increasing model size enhance the robustness to the repetition problem" or "How are the previous and proposed methods effective as the model size increases?" - A case study on samples generated from the models would provide further insights. [1] Pillutla, Krishna, et al. "Mauve: Measuring the gap between neural text and human text using divergence frontiers." Advances in Neural Information Processing Systems 34 (2021): 4816-4828.

Questions

- Is there a possible reason for a relatively high repetition ratio in OpenWebText2?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

As written in the Limitation section, experiments on various scales could be conducted in future work.

Reviewer 6VM26/10 · confidence 4/52023-07-07

Summary

The paper explores the issue of degeneration in text generation, which refers to the generation of repetitive words and dull loops by neural language models. The authors focus on the impact of repetition in the training data and propose a method to address this issue. Specifically, they suggest dropping out repetitive words during the training data pre-processing stage. The main conclusion drawn from their investigation is that penalizing repetitive n-grams in the training data is crucial for the effectiveness of existing methods in preventing degeneration. By highlighting the importance of addressing repetition in the training data, the paper contributes to understanding the factors influencing degeneration in text generation.

Strengths

The paper is generally well-written and is easy to follow. The problem of generating repetitive words and dull loops has been widely observed and extensively discussed in the field, which makes this problem important and worth exploring. The conclusion drawn by the authors, that training data plays a pivotal role in the occurrence of degeneration, is logical and well-supported.

Weaknesses

The paper primarily focuses on data pre-processing rather than novel training techniques. This limited scope may diminish its technical novelty. Many researchers working on text generation may have already explored similar data pre-processing ideas, making the proposed method less impactful or original. Also, the conclusion of the paper lacks novel insights and seems to affirm an intuitive understanding. While it is reasonable to assume that data quality has a direct impact on degeneration, this insight does not offer any new or surprising findings. As a result, the conclusion may be considered weak in terms of offering novel contributions or expanding the current understanding of the issue. Explore additional factors that could contribute to degeneration in text generation, such as the number of training data and model size. Investigating these factors may provide deeper insights into the phenomenon and allow for a more comprehensive understanding of degeneration. This expansion could strengthen the paper and enhance its impact within the field. For me, it would be more interesting to study other factors that impact degeneration, such as the number of training data, the model size and the model architecture. Investigating these factors may provide deeper insights into the phenomenon and allow for a more comprehensive understanding of degeneration. With more training data and more parameters, large language models can capture semantics and thus prevent degeneration. For example, degeneration has been greatly resolved for ChatGPT due to its expressiveness and more general understanding of texts, In my opinion, tackling degeneration in data pre-processing is not very meaningful because the root cause is still the model lacks of capability to understand and generate reasonable texts.

Questions

Does dropping some repetitive n-grams sometimes distort the semantics of the document?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

There are other potential factors that can impact the degeneration. The paper should have a more comprehensive discussion about all the factors.

Reviewer 6H9c7/10 · confidence 4/52023-07-09

Summary

This paper explains the repetition in model generated text from a data standpoint, pointing out that there is a strong correlation between the degeneration issue and the presence of repetitions in training data. The authors find out that penalizing repetitions in data can alleviate degeneration, and propose a method, repetition dropout, that is to apply dropout in transformers’ sublayers during training time. The proposed method achieve significant improvement in terms of REP-n scores.

Strengths

- This is a very well written paper! The preliminary study does a great job in introducing and clearly motivating the problem. - The proposed repetition dropout is also a very simple technique, yet still improve the performance in REP-n stats. - The comparison between different methods and objectives in section 6 is very interesting and thorough.

Weaknesses

Currently it seems that the repetition dropout mask is applied on each instance, but what about repeated text in different instance? How should we apply it in current LM training paradigm?

Questions

NA

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

NA

Authorsrebuttal2023-08-14

Response to Reviewer Pdys

Dear Reviewer Pdys, We are grateful for your insightful feedback and your recognition of our efforts to address the concerns. Thank you so much for increasing the scores. Regards, Authors

Reviewer 6H9c2023-08-15

Thank you for your reply

Hi, Thank you for your reply! >confirm that by "repeated text" you are referring to repetitive n-grams, and by "instance" you mean a sentence Yes, your interpretation is correct. I am satisfied of the answers to my question, and my rating will stay the same.

Authorsrebuttal2023-08-19

Thanks for your kind confirmation, and we are glad that the previous response addressed your concern.

Reviewer 6VM22023-08-17

Raise my rating to 6

Thanks for the rebuttal, which addressed my primary concerns. The observation that scaling up a model size can somewhat mitigate repetition while still presenting a significant challenge for larger models is interesting. I also notice that the models discussed in the rebuttal are not instruction-tuned models. I'm curious if there's a specific rationale behind this choice. Although I recognize ChatGPT is not comparable due to its limited accessibility, I believe it could be insightful to include a comparison of instruction-tuned models (i.e. alpaca/vicuna/Llama2) and the model without instruction-tuning (i.e. LLama). Also, it would be interesting to analyze how rep scores connect to the downstream generative task performance metrics like ROUGE scores for summarization.

Authorsrebuttal2023-08-19

Many thanks for your kind reply to our previous response and raising the scores! We really enjoy the discussion with you, and we are happy that our response addressed your primary concerns, **Q: I also notice that the models discussed in the rebuttal are not instruction-tuned models. I'm curious if there's a specific rationale behind this choice.** The main reason of this choice is to make a fair comparison with previous works, because most previous works only focus on the standard language model. **Q: I believe it could be insightful to include a comparison of instruction-tuned models (i.e. alpaca/vicuna/Llama2) and the model without instruction-tuning (i.e. LLama).** Table 1: Results of Llama2-7B on instruction-following data. The column "Rep-2 of FT Data" indicates the rep-2 score of the training data used for fine-tuning. The rest Rep-2, Rep-3, and Rep-4 scores are evaluated on the generated text by different methods. The "FT" means fine-tuning. | No. | Methods | | Rep-2 of FT Data | | Rep-2 | Rep-3 | Rep-4 | |-----|-------------------------------|---|------------------|---|-------|-------|-------| | 1 | Llama2 w/o FT | | -- | | 47.79 | 41.97 | 38.52 | | 2 | FT Llama2 on Alpaca | | 5.54 | | 15.08 | 10.91 | 8.93 | | 3 | FT Llama2 on Alpaca + WT-103 50K | | 9.67 | | 41.63 | 35.64 | 32.29 | | 4 | FT Llama2 on WT-103 | | 10.31 | | 54.10 | 49.77 | 36.80 | Thanks for this valuable suggestion. We agree that analyzing the effect of instruction-tuning on degeneration is meaningful. We conduct experiments on three datasets: 1. **Alpaca**: The instruction-tuning dataset used by Alpaca [1]. 2. **WT-103 50K**: We randomly sample 50k sentences from Wikitext-103 and convert them to the instruction-following data. More details are at the end of this response. 3. **Alpaca + WT-103 50K**: The mixture of Alpaca and WT-103 50K As shown in Table 1, the “Llama2 w/o FT” (Line 1) indicates the LLM without instruction-tuning, and “FT Llama2 on Alpaca” (Line 2) means the Llama2 with instruction-tuning. We can find that the instruction-tuning process does alleviate the degeneration issue. [1]: https://crfm.stanford.edu/2023/03/13/alpaca.html However, we hypothesize that the alleviation of degeneration is caused by that the training data of **Alpaca** has less repetitions. As shown in Table 1, the rep-2 scores of the **Alpaca**, **Alpaca + WT-103 50K**, and **WT-103 50K** datasets are 5.54, 9.67, and 10.31, respectively. We can find that the degeneration issue becomes severer if we fine-tune the model on instruction-following data with higher repetition rate (Line 2-4 in Table 1). This observation further demonstrates that the degeneration issue has a high correlation with the repetitions in training data during the instruction-tuning process, which is consistent with the finding in our paper. Implementation details of our experiments: 1. **Fine-tuning strategy**: we use the QLoRA to fine-tune the Llama2-7B model, due to the limited computational resources. 2. **Decoding strategy**: greedy search 3. **Test Data**: The test set of Wikitext-103 in the instruction-following format. 4. **Data pre-processing**: To ensure a fair comparison, we convert the wikitext-103 dataset to a instruction-following dataset by using the following template: ``` { "instruction": "Please continue writing based on the following prefix. The text to be continued should be relevant, fluent and informative.", "input": PREFIX, # prefix of a sentence "output": COMPLETION # the completion of the prefix } ``` **Q: Also, it would be interesting to analyze how rep scores connect to the downstream generative task performance metrics like ROUGE scores for summarization.** Thanks for your suggestion. The downstream tasks, e.g., summarization, may also suffer from the degeneration issue. We agree it is interesting to evaluate our methods and findings on those tasks. We will leave the further investigation to future work.

Reviewer jpnv2023-08-17

Thank you for the experimental results which have addressed my curiosity. I could notice the repetition problem does not resolve by simply scaling the language model and Rep-Dropout is effective for the larger models. I think further studies on scaling and decoding strategies would also be meaningful. I raised the score after reading the rebuttal.

Authorsrebuttal2023-08-19

Thanks for your constructive feedback and raising the overall score! We agree that it is meaningful to investigate the scaling and decoding strategies, and we also planned to compare those different factors in a unified framework. We will leave the detailed investigations to our future work.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC