[Paper Review] Neural Sequence Prediction by Coaching.

Wenhu Chen, Guanlin Li+4 more|arXiv (Cornell University)|Jun 28, 2017

Topic Modeling1 citations

TL;DR

This paper proposes the Generative Bridging Network (GBN), a novel training framework that improves sequence prediction by introducing a bridge module to mitigate data sparsity and overfitting in maximum likelihood estimation. By minimizing the KL divergence between the generator's output and a bridge distribution conditioned on ground truth, GBN enhances model confidence, language smoothness, and training efficiency, yielding significant gains in machine translation and abstractive summarization tasks.

ABSTRACT

In order to alleviate data sparsity and overfitting problems in maximum likelihood estimation (MLE) for sequence prediction tasks, we propose the Generative Bridging Network (GBN), in which a novel bridge module is introduced to assist the training of the sequence prediction model (the generator network). Unlike MLE directly maximizing the conditional likelihood, the bridge extends the point-wise ground truth to a bridge distribution conditioned on it, and the generator is optimized to minimize their KL-divergence. Three different GBNs, namely uniform GBN, language-model GBN and coaching GBN, are proposed to penalize confidence, enhance language smoothness and relieve learning burden. Experiments conducted on two recognized sequence prediction tasks (machine translation and abstractive text summarization) show that our proposed GBNs can yield significant improvements over strong baselines. Furthermore, by analyzing samples drawn from different bridges, expected influences on the generator are verified.

Motivation & Objective

• To address data sparsity and overfitting in sequence prediction models trained via maximum likelihood estimation.
• To improve model generalization and training stability by replacing direct likelihood maximization with a bridge-based optimization objective.
• To enhance model confidence, fluency, and learning efficiency through distinct bridge variants: uniform, language-model, and coaching GBN.
• To empirically validate the effectiveness of the proposed framework on standard sequence generation benchmarks.

Proposed method

• Introduce a bridge module that transforms the point-wise ground truth into a distribution, enabling more robust training signals.
• Optimize the generator by minimizing the KL divergence between its output and the bridge distribution, rather than maximizing likelihood directly.
• Design three variants: uniform GBN for confidence regularization, language-model GBN for fluency, and coaching GBN to reduce learning burden.
• Train the generator end-to-end using the bridge-based objective, with the bridge distribution conditioned on the ground truth sequence.
• Use the bridge distribution to guide the generator toward more diverse and plausible outputs during training.
• Apply the framework to sequence prediction tasks such as machine translation and abstractive summarization.

Experimental results

Research questions

• RQ1 Can replacing direct likelihood maximization with a bridge-based objective reduce overfitting and data sparsity in sequence modeling?
• RQ2 How does the bridge distribution influence the generator’s confidence and output quality?
• RQ3 To what extent can different bridge designs—uniform, language-model, and coaching—improve model performance and training dynamics?
• RQ4 Does the bridge module enhance fluency and diversity in generated sequences without requiring additional training data?
• RQ5 How do the generated samples from different bridge distributions affect the final generator behavior?

Key findings

• The proposed GBN framework achieves significant improvements over strong baselines in both machine translation and abstractive text summarization tasks.
• The coaching GBN variant effectively reduces the learning burden on the generator, leading to faster convergence and better performance.
• The language-model GBN enhances output fluency by incorporating n-gram language modeling signals into the training objective.
• The uniform GBN variant successfully regularizes model confidence, reducing overconfidence in predictions.
• Analysis of samples from different bridges confirms their expected influence on the generator, validating the design principles of the framework.
• The bridge-based training objective leads to more diverse and plausible outputs compared to standard MLE training.