[Paper Review] Sticking to the Facts: Confident Decoding for Faithful Data-to-Text Generation

Ran Tian, Shashi Narayan+2 more|arXiv (Cornell University)|Oct 19, 2019

Topic Modeling72 references48 citations

TL;DR

The paper introduces a confidence-based decoding framework for encoder-decoder models to reduce hallucinations in data-to-text generation by coupling an attention-derived confidence score with a tailored base language model and a variational Bayes training objective.

ABSTRACT

We address the issue of hallucination in data-to-text generation, i.e., reducing the generation of text that is unsupported by the source. We conjecture that hallucination can be caused by an encoder-decoder model generating content phrases without attending to the source; so we propose a confidence score to ensure that the model attends to the source whenever necessary, as well as a variational Bayes training framework that can learn the score from data. Experiments on the WikiBio (Lebretet al., 2016) dataset show that our approach is more faithful to the source than existing state-of-the-art approaches, according to both PARENT score (Dhingra et al., 2019) and human evaluation. We also report strong results on the WebNLG (Gardent et al., 2017) dataset.

Motivation & Objective

• Motivate and address hallucination in data-to-text generation where outputs may be unfaithful to the source.
• Propose a confidence score that ties attention to source information with a base language model to judge fidelity of each generated token.
• Develop a variational Bayes training framework to learn the confidence score from data and promote confident subsequences during training.
• Evaluate fidelity and fluency on WikiBio and WebNLG, comparing to state-of-the-art baselines.

Proposed method

• Define a confidence score C_t(y_t) that combines an attention-based signal A_t and a base language model probability P_B(y_t | y_<t).
• Modify attention to allow “not attending” by normalizing with a constant in the denominator and by excluding source information from RNN hidden state inputs.
• Introduce a tailorable base language model RNN_B that down-weights source-associated inputs to learn soft templates.
• Train with a variational Bayes objective by sampling a confident subsequence Z from Q(z|y,x) and maximizing a bound that favors confident, faithful tokens; use Monte Carlo estimates to approximate expectations.
• Use calibration and a <null> token mechanism at inference to re-rank and suppress unconfident tokens, improving precision without sacrificing fluency.

Experimental results

Research questions

• RQ1 Can a confidence-guided decoding strategy reduce hallucinations in data-to-text generation without sacrificing fluency?
• RQ2 How should attention and language modeling components be restructured to support faithful generation from structured sources?
• RQ3 Can a variational Bayes framework learn a reliable confidence score for token-level faithfulness from data?
• RQ4 Do calibration and <null> token strategies further improve fidelity while maintaining or improving fluency?
• RQ5 Are these methods effective across datasets with varying source-reference divergence (WikiBio vs. WebNLG)?

Key findings

• The confidence-based decoding approach yields higher faithfulness (precision) and F1 on PARENT compared to baselines on WikiBio and WebNLG datasets.
• Calibrating generation with the confidence score improves token selection without raising perplexity, and can boost recall with controlled precision.
• Variational Bayes sub-sequence sampling focuses training on confident, source-supported tokens, while the base language model learns soft templates that reduce unfaithful generation.
• The <null> token strategy and length penalty during inference help balance recall and precision, enhancing overall faithful generation without substantial loss of fluency.
• The approach increases model sensitivity to the source, as shown by higher token-level changes when source vectors are ablated during decoding, indicating reliance on source information for faithful outputs.