PLANNER: Generating Diversified Paragraph via Latent Language Diffusion Model

Autoregressive models for text sometimes generate repetitive and low-quality output because errors accumulate during the steps of generation. This issue is often attributed to exposure bias - the difference between how a model is trained, and how it is used during inference. Denoising diffusion models provide an alternative approach in which a model can revisit and revise its output. However, they can be computationally expensive and prior efforts on text have led to models that produce less fluent output compared to autoregressive models, especially for longer text and paragraphs. In this paper, we propose PLANNER, a model that combines latent semantic diffusion with autoregressive generation, to generate fluent text while exercising global control over paragraphs. The model achieves this by combining an autoregressive "decoding" module with a "planning" module that uses latent diffusion to generate semantic paragraph embeddings in a coarse-to-fine manner. The proposed method is evaluated on various conditional generation tasks, and results on semantic generation, text completion and summarization show its effectiveness in generating high-quality long-form text in an efficient manner.

Paper

Similar papers

Peer review

Reviewer H5ac7/10 · confidence 4/52023-07-05

Summary

The paper presents a two-stage latent text diffusion model that uses an autoencoder to condense lengthy texts into a limited number of paragraph embeddings, and a continuous time diffusion model that359 learns the distribution of these embeddings. The paper presents detailed experiments on open-ended generation tasks and shows that the proposed model alleviates the issue of repetition and advances generation diversity across different tasks.

Strengths

The paper is well-motivated. The examples in Figure 1 with the self-reinforcing effect are impressive. The proposed method first summarizes the paragraph information with an autoencoder and then applies the diffusion process to control the token generation process, which is novel from my point. The paper conducts extensive experiments on open-ended generation tasks and provides some analyses of the running time.

Weaknesses

1. Could you provide more experimental analyses on the learned paragraph embedding? I think that paragraph embedding is used to learn a high-level concept for planning but ignores the details of sentence/phrase structure to avoid copying similar phrases from previous prompts and repeating them. Could you provide more experiments to show the quality of learned paragraph embedding? Directly extracting the first K tokens from the encoder is not the best choice for me. 2. In experiments such as Table 1, the proposed method with a greedy decoding mode is comparable to baselines with top-p sampling in terms of diversity. A question is whether PLANNER combined with top-p can lead to better results. Are the proposed methods compatible with stochastic decoding methods?

Questions

Please see the weakness.

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

4 excellent

Contribution

3 good

Limitations

Please see the weakness.

Reviewer 6a1h6/10 · confidence 4/52023-07-06

Summary

This paper presents a 2-stage generative model for text. The first stage involves training a VAE over the data to obtain an effective encoder and decoder modules. The second stage uses the output of the frozen VAE encoder as the hidden state of the input text and learns a regular continuous diffusion model over the encoded hidden states of training data instances. For generation, the diffusion model produces a hidden state from noise and this hidden state is fed into the VAE decoder to generate text. This model is evaluated on sentiment-guided generation, text completion, and summarization and compared against baselines that include token-based diffusion models (instead of hidden state text diffusion models in this work), VAE/Enc-dec models, and standard autoregressive models (like fine tuned GPT-2).

Strengths

– Although the approach seems straightforward, to the best of my knowledge, surprisingly I am not aware of other latent sequence/paragraph embedding based diffusion models for text generation. Hence, this fills a gap in the research on diffusion models for text. – The experimental design is reasonable and informative. – This approach seems to be more fluent, diverse, and effective than popular token-based diffusion models following the results. – Although typically less fluent than the autoregressive models, it still performs competitively when considering other metrics like diversity, reference overlap, and control.

Weaknesses

– I have some concerns about evaluation, particularly related to the quality of baseline models. For example, on CNN-DM summarization task, publicly available results (https://paperswithcode.com/sota/abstractive-text-summarization-on-cnn-daily) show that T5 achieve R-L that is ~7-10 points better than reported in this paper (and is hence better than the R-L proposed model achieves). Therefore, I am doubtful about the quality of training/tuning of the baseline models reported in the paper. – Choosing “first k hidden state vectors” from the encoder in the VAE seems arbitrary. What is the motivation behind this hyperparameter? What is the effect of this k? – The paper alludes to learning “paragraph embeddings” which capture high level semantic properties which enables “coarse-to-fine” generation and allows for “planning”. However, no evidence is presented in this paper about compositional/planning capabilities of the model. While it is true that paragrpahs/tokens are generated from a lower dimensional manifold, it doesn’t automatically imply that these embeddings are linguistically interesting or allow for controlled planning-oriented generation. At an abstract conceptual level, how different are these representations from VAE based representation for example? – Although this is a generative model, the training objective is arbitrary and doesn’t seem linked to maximizing the likelihood/ learning the distribution of the training data. This is because the VAE and the diffusion components are separately trained and hence preclude a clean interpretation of the training objective. – I am not convinced that the “Distributional smoothness metric” reflect the nature of the manifold as is claimed in the paper. – Runtime analysis seems to compare unbatched autoregressive models with batched diffusion models. Sorting by length and then padding for batching is a standard practice for autoregressive models.

Questions

See above.

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

2 fair

Presentation

3 good

Contribution

3 good

Limitations

See above.

Reviewer r7NC5/10 · confidence 4/52023-07-06

Summary

In this paper, the authors propose PLANNER, a model that combines latent semantic diffusion with autoregressive generation to generate fluent text while exercising global control over paragraphs. The proposed method is evaluated on various conditional generation tasks, and the results show its effectiveness in generating high-quality text.

Strengths

1. The paper addresses the issue of repetitive and low-quality output generated by autoregressive models and proposes a novel approach using latent semantic diffusion. 2. The combination of autoregressive decoding and latent diffusion allows for efficient generation over paragraph generation. 3. The proposed method is evaluated on various tasks and shows improved generation quality compared to autoregressive and text diffusion baselines.

Weaknesses

1. I understand that revisiting and revising the generated sentences can alleviate exposure bias, as errors can be reduced through further editing at the token level. However, revisiting and revising the latent space does not seem reasonable to me. From my perspective, the exposure bias occurs in the process of picking words out, but the authors merely employ GPT-2 without making any changes in the phase. 2. The paper claims that they can generate longer text and paragraphs, but there is no further analysis about relationship between length of generated sentence and the performance. 3. The evaluation of CNN/Daily Mail and XSum is performed on 256 subsampled examples from the test set. This manner is not convenient for the subsequent works to follow, and this also hinders reviewers from making fair comparisons between this work and others. 4. I have reservations about the ability of the variational paragraph embedder to learn effective representations of sentences with different lengths. What would happen if there is a significant difference in sentence lengths, such as one being very short (e.g., 5 tokens) and the other being much longer (e.g., 512 tokens or more)?

Questions

1. Unlike autoregressive models based on conditional probabilities, diffusion models are unable to ascertain the optimal sentence from the generated set. How did you address this problem in your paper? 2. As the reverse diffusion goes on in latent space, does the latent representation z get closer to the representation of the ground-truth one in cosine distance or other distances? 3. What is the influence of paragraph embeddings number, k in line 119?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

2 fair

Presentation

3 good

Contribution

2 fair

Reviewer MZNE7/10 · confidence 4/52023-07-15

Summary

This paper proposes to combine latent semantic diffusion with autoregressive generation to alleviate the issues of exposure bias in training / inference of text based language models, and computational and performance cost of purely diffusion approaches. They improve the diffusion process by applying it to the latent semantic space instead of the token / embedding space. For this, they learn some "semantic tokens" for encoding paragraph level information and then use a decoder to map these to the raw text space.

Strengths

Working with the diffusion model in the semantic spaces opens up the door for controllable generation. They also provide an extensive study of the requirements for a good latent space for paragraph diffusion models. The paper is well written and easy to follow. The ideas employed to fix / ensure local smoothness (by perturbing data) and distributional smoothness (by using VAE) are simple and useful. They propose a novel (?) metric called AuBLEU to evaluate the denoising capabilities of the model. I believe is this generally suitable for other works and could be impactful. However, it does not feel to be properly justified / grounded. Hparams are provided. Human eval results are significant. The other results are sound and robust evaluation is performed. The paper is just lacking ablations on the design choices. The analysis is complete and justifies the main claims made by the authors.

Weaknesses

The changes employed by the authors, especially during the training stage are not properly ablated. It is not clear if the proposed fixes provide benefits. Some of the design choices are not experimentally justified (eg: line 182) Evaluation is performed on just a sample of the test set (this is the first time I have seen something like this in a paper and I'm not sure how to take it - I'm not super comfortable as this makes your technique essentially un-comparable). This also might not be robust. (line 196). No intention of providing code and it might be very hard to reproduce because of the many changes especially in the training setup. Nitpicks: typo line 181

Questions

Can you talk a bit more about the design choices and how much impact they had on the results? For eg: how much impact does fixing "distributional smoothness" have? Have you performed other ablation experiments to justify claims? Why did you subsample the test sets?

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

4 excellent

Limitations

Addressed in appendix.

Reviewer MZNE2023-08-14

I acknowledge the rebuttal and appreciate the full evaluation. Can you add statistical tests to your evaluation results as well? I thank the authors for the clarification and would like to stick with my initial score.

Authorsrebuttal2023-08-17

Thank you for reading our response and your additional suggestions! We will perform the statistical analysis on the evaluation results in our next revision.

Reviewer 6a1h2023-08-16

Thanks for the response

I am willing to agree that the results I mentioned might have been achieved by a much larger model. However, my overall impression remains similar as before so I am keeping my current score.

Reviewer r7NC2023-08-18

Thank you for the clarifications. I will raise my rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC