Controlled Hallucinations: Learning to Generate Faithfully from Noisy Data
This addresses the issue of unreliable text generation for applications with limited or noisy training data, representing an incremental improvement in handling data noise.
The paper tackles the problem of neural text generation models hallucinating unsupported text when trained on noisy parallel data, and introduces a technique to control hallucinations without modifying the model architecture, achieving improved performance on the WikiBio corpus in both automatic and human evaluations.
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.