Source-linked AI summary
Generative Adversarial Network for Abstractive Text Summarization
Linqing Liu, Yao Lu, Min Yang, Qiang Qu, Jia Zhu, Hongyan Li
TL;DR
Abstractive summarization still faces generic, less readable outputs and training–evaluation mismatches. This paper jointly trains a reinforcement-learning generator and a summary discriminator adversarially, improving ROUGE over its pretrained generator and reported readability.
Problem
Abstractive summarization remains challenged by generic outputs, limited readability, and maximum-likelihood training that mismatches evaluation and testing conditions.
Method
The paper jointly trains a generator that predicts summaries with reinforcement learning and a discriminator that distinguishes generated summaries from human summaries.
Results
ROUGE-1, ROUGE-2 and ROUGE-L increase by 1.10, 0.84 and 1.00 absolute points over the pretrained generator, while the model achieves the best ROUGE-1 and ROUGE-2 scores.
Takeaways & Limitations
The model generates more abstractive, readable and diverse summaries, with reported improvements in ROUGE and readability.
Abstract
from arXiv · showhide
In this paper, we propose an adversarial process for abstractive text summarization, in which we simultaneously train a generative model G and a discriminative model D. In particular, we build the generator G as an agent of reinforcement learning, which takes the raw text as input and predicts the abstractive summarization. We also build a discriminator which attempts to distinguish the generated summary from the ground truth summary. Extensive experiments demonstrate that our model achieves competitive ROUGE scores with the state-of-the-art methods on CNN/Daily Mail dataset. Qualitatively, we show that our model is able to generate more abstractive, readable and diverse summaries.
Introduction
Abstractive summarization aims to generate concise summaries that capture salient source ideas, but existing systems face generic outputs, readability problems, and exposure bias. The paper addresses these challenges with adversarial joint training of a generator and discriminator.
- Abstractive summarization generates concise summaries that capture salient source ideas and may contain phrases absent from the source.
- Neural sequence-to-sequence models often produce trivial, generic summaries dominated by high-frequency phrases.
- Generated summaries can have limited grammaticality and readability.
- Maximum-likelihood training mismatches evaluation metrics and exposes models to different decoder inputs during training and testing.
- Exposure bias leads to error accumulation at test time.
- The proposed adversarial framework jointly trains a generator and discriminator to address these challenges.
Our model
The model alternates a generator that predicts summaries from source text with a discriminator that distinguishes human from machine summaries. Reinforcement learning uses discriminator feedback alongside maximum-likelihood training to update the generator.
- Training procedure: The generator is pretrained on source-summary pairs, while the discriminator learns from human summaries and outputs from the pretrained generator.
- Generative Model: The generator encodes source text with a bidirectional LSTM and predicts summaries with an attention-based LSTM decoder.
- Generative Model: A switching pointer-generator can either generate from a fixed vocabulary or copy rare and unseen words from the input.
- Discriminative Model: The discriminator uses a CNN with multiple filter sizes, max-over-time pooling, and a classifier to estimate whether a sequence is human-generated.
- Adversarial training: The generator and discriminator are trained alternately, with discriminator feedback acting as a reward that is updated iteratively.
- Updating model parameters: The generator objective combines policy-gradient loss and maximum-likelihood loss as J = βJpg + (1 −β)Jml.
Experiments
Experiments on CNN/Daily Mail compare the adversarial model with established abstractive summarizers using ROUGE and human readability judgments. The model improves ROUGE over its pretrained generator and is reported to produce more readable summaries.
- Dataset: The CNN/Daily Mail corpus contains 287,226 training pairs, 13,368 validation pairs, and 11,490 test pairs.
- Experimental Results: The model is compared with ABS, pointer-generator coverage networks, and DeepRL.
- Experimental Results: 1.10, 0.84 and 1.00 absolute points increase ROUGE-1, ROUGE-2 and ROUGE-L, respectively, over the pretrained generator.
- Experimental Results: The approach achieves the best ROUGE-1 and ROUGE-2 scores among the compared methods.
- Human evaluation: Human evaluation ranks summaries for readability on 50 randomly selected test examples using a five-point scale.
- Human evaluation: The authors report that the model significantly improves summary readability.
Conclusion
The paper proposes an adversarial process for abstractive text summarization. Its model generates summaries that are more abstractive, readable, and diverse.
- The paper proposes an adversarial process for abstractive text summarization.
- The model generates more abstractive summaries.
- The model generates more readable and diverse summaries.