Source-linked AI summary
ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training
Weizhen Qi, Yu Yan, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang, Ming Zhou
TL;DR
ProphetNet targets the limited future planning and long-term dependency modeling of one-step autoregressive Seq2Seq training. It combines future n-gram prediction with n-stream self-attention, and reports best or state-of-the-art results across the evaluated summarization and question-generation benchmarks. One stated data caveat is that REALNEWS substitutes for the unreleased CC-News corpus.
Problem
One-step autoregressive training can overfit strong local correlations while underfitting global coherence and long-term dependencies.
Method
ProphetNet combines future n-gram prediction, n-stream self-attention, and mask-based auto-encoder denoising for Seq2Seq pre-training.
Results
ProphetNet achieves the best performance on abstractive summarization and question generation tasks, including new state-of-the-art results on CNN/DailyMail and Gigaword.
Takeaways & Limitations
Future n-gram prediction provides extra guidance for planning future tokens while ProphetNet remains compatible with standard next-token inference and fine-tuning.
Takeaways & Limitations
Because CC-News is not officially released, the experiments use the similar public news corpus REALNEWS instead.
Abstract
from arXiv · showhide
This paper presents a new sequence-to-sequence pre-training model called ProphetNet, which introduces a novel self-supervised objective named future n-gram prediction and the proposed n-stream self-attention mechanism. Instead of optimizing one-step-ahead prediction in the traditional sequence-to-sequence model, the ProphetNet is optimized by n-step ahead prediction that predicts the next n tokens simultaneously based on previous context tokens at each time step. The future n-gram prediction explicitly encourages the model to plan for the future tokens and prevent overfitting on strong local correlations. We pre-train ProphetNet using a base scale dataset (16GB) and a large-scale dataset (160GB), respectively. Then we conduct experiments on CNN/DailyMail, Gigaword, and SQuAD 1.1 benchmarks for abstractive summarization and question generation tasks. Experimental results show that ProphetNet achieves new state-of-the-art results on all these datasets compared to the models using the same scale pre-training corpus.
1 Introduction
ProphetNet addresses the tendency of autoregressive Seq2Seq models to overfit strong local correlations by jointly predicting future n-grams. Its n-stream self-attention enables efficient future-token prediction during training while retaining standard next-token inference and fine-tuning.
- Motivation: Autoregressive Seq2Seq models can capture local coherence while underfitting global coherence and long-term dependencies.Strong local correlations and one-step teacher forcing provide no explicit bias toward future-token planning.
- ProphetNet: ProphetNet adds future n-gram prediction to traditional one-step-ahead language modeling.The additional objective predicts multiple future tokens and explicitly encourages future planning.
- Architecture: N-stream self-attention provides one main stream and n predicting streams for simultaneous future-token prediction.Each predicting stream attends to main-stream hidden states to predict a different future token.
- Architecture: Shared main-stream parameters allow ProphetNet to disable n-stream attention during inference and predict only the next token.This preserves the original Transformer Seq2Seq inference and fine-tuning behavior.
- Experiments: ProphetNet uses mask-based auto-encoder denoising with 16GB and 160GB pre-training datasets before downstream fine-tuning.The experiments evaluate abstractive summarization and question generation tasks.
2 ProphetNet
ProphetNet extends Transformer Seq2Seq pre-training with future n-gram prediction and n-stream self-attention. It predicts multiple future tokens during training while retaining compatibility with conventional next-token inference and fine-tuning.
- 2.2 Future N-gram Prediction: ProphetNet introduces future n-gram prediction, replacing one-step prediction with simultaneous prediction of the next n continuous tokens at each time step.The objective combines the traditional language-modeling loss with n−1 future-token prediction losses.
- 2.2 Future N-gram Prediction: The future n-gram loss encourages future-token planning and reduces overfitting to strong local correlations.The loss is combined with the traditional language-modeling objective using weights α_j, with γ described as the attenuation coefficient.
- 2.3 N-Stream Self-Attention: Its decoder uses n extra predicting streams alongside the main Transformer self-attention stream, with the i-th stream modeling the i-th future token.The main stream preserves masked self-attention, while predicting streams use previous main-stream hidden states to predict future positions.
- 2.3 N-Stream Self-Attention: Predicting-stream parameters are shared with the main stream, allowing all predicting streams to be disabled for conventional Transformer decoding during inference or fine-tuning.This design supports simultaneous future-token training without changing the next-token decoding form used after training.
- 2.4 Seq2Seq Pre-training on Denoising Task: For pre-training, ProphetNet masks token spans in an auto-encoder denoising task and recovers the next n future tokens within each masked span.The task follows token-span masking and differs from MASS by recovering multiple future tokens rather than one next token at each step.
3 Experiments and Results
ProphetNet is evaluated through pre-training and fine-tuning on abstractive summarization and answer-aware question generation, using standard and large-scale corpora. Across these experiments, it improves over prior models, including without pre-training, while ProphetNetbase-2gram offers a faster alternative to 3-gram prediction.
- Downstream tasks: ProphetNet is fine-tuned for abstractive summarization on CNN/DailyMail and Gigaword, and for answer-aware question generation on SQuAD 1.1.The experiments use standard Seq2Seq task formulations and evaluate two SQuAD data splits.
- Standard-scale pre-training: ProphetNet achieves the best performance on all reported CNN/DailyMail ROUGE metrics.The evaluation reports F1 scores for ROUGE-1, ROUGE-2, and ROUGE-L on the test set.
- Standard-scale pre-training: ProphetNet outperforms previous models on all reported Gigaword metrics.Comparisons include OpenNMT, Re3Sum, MASS, and UniLM.
- Question generation: ProphetNet outperforms all previous methods on SQuAD 1.1 question generation with significant improvement.The same model and inference hyper-parameters are used for the two data splits with swapped development and test sets.
- Large-scale pre-training: 160GB pre-training yields state-of-the-art results on CNN/DailyMail and Gigaword, exceeding PEGASUS baselines on Gigaword with smaller pre-training corpora.ProphetNet uses about one-fifth and one-twentieth of the cited PEGASUS pre-training corpus sizes, respectively.
- Without pre-training: Without pre-training, the proposed method significantly improves performance over an architecture-matched original Transformer Seq2Seq model.The comparison uses the same architecture hyper-parameters for ProphetNet and the original Transformer.
- N-gram comparison: ProphetNetbase-2gram performs comparably to ProphetNetbase-3gram while training 15% faster, motivating its use in subsequent experiments.Both variants outperform MASSbase and ProphetNetbase-1gram.
4 Related Work
Prior work spans autoregressive, bidirectional, and denoising objectives for language modeling and sequence-to-sequence generation, while ProphetNet-related research emphasizes incorporating future information into generation.
- GPT uses left-to-right language modeling to predict subsequent tokens, whereas BERT and SpanBERT recover masked tokens or spans.
- BART uses an encoder-decoder denoising objective that reconstructs original sentences from corrupted inputs.
- Natural language generation methods typically generate one token per step using left-to-right or right-to-left language models.
- Recent generation research incorporates future information through approaches including actor-critic modeling with a value function.
5 Conclusion
ProphetNet is a sequence-to-sequence pre-training model that predicts future n-grams at each time step. It achieves the best reported performance on abstractive summarization and question generation, including new state-of-the-art results on CNN/DailyMail and Gigaword with about one-third the prior model's pre-training epochs.
- ProphetNet learns to predict future n-grams at each time step in sequence-to-sequence pre-training.
- ProphetNet achieves the best performance on abstractive summarization and question generation tasks.
- ProphetNet achieves new state-of-the-art results on CNN/DailyMail and Gigaword using about 1/3 the pre-training epochs of the previous model.