Controlled Hallucinations: Learning to Generate Faithfully from Noisy Data

TL;DR

This paper introduces a technique to control hallucinations in neural text generation from noisy data, improving output faithfulness without changing model architecture, validated on the noisy WikiBio dataset.

Contribution

The paper presents a simple, effective method to treat hallucinations as controllable aspects of generated text without altering model architecture or dismissing input.

Findings

Improved faithfulness in generated text on WikiBio dataset

Effective control over hallucinations demonstrated in evaluations

Method works without modifying existing model architectures

Abstract

Neural text generation (data- or text-to-text) demonstrates remarkable performance when training data is abundant which for many applications is not the case. To collect a large corpus of parallel data, heuristic rules are often used but they inevitably let noise into the data, such as phrases in the output which cannot be explained by the input. Consequently, models pick up on the noise and may hallucinate--generate fluent but unsupported text. Our contribution is a simple but powerful technique to treat such hallucinations as a controllable aspect of the generated text, without dismissing any input and without modifying the model architecture. On the WikiBio corpus (Lebret et al., 2016), a particularly noisy dataset, we demonstrate the efficacy of the technique both in an automatic and in a human evaluation.