Calibrating Sequence likelihood Improves Conditional Language Generation
This addresses a fundamental issue in conditional language generation for NLP applications, offering a method to improve sequence ranking without heuristics, though it appears incremental as it builds on existing MLE training paradigms.
The paper tackles the problem of conditional language models trained with maximum likelihood estimation (MLE) not accurately ranking generated sequences by quality, leading to degraded output with large beam sizes and reliance on heuristics. It introduces sequence likelihood calibration (SLiC), which aligns model-generated sequence likelihoods with references, eliminating the need for decoding heuristics and achieving or matching state-of-the-art results across tasks like abstractive summarization and question generation, even with modest-sized models.
Conditional language models are predominantly trained with maximum likelihood estimation (MLE), giving probability mass to sparsely observed target sequences. While MLE trained models assign high probability to plausible sequences given the context, the model probabilities often do not accurately rank-order generated sequences by quality. This has been empirically observed in beam search decoding as output quality degrading with large beam sizes, and decoding strategies benefiting from heuristics such as length normalization and repetition-blocking. In this work, we introduce sequence likelihood calibration (SLiC) where the likelihood of model generated sequences are calibrated to better align with reference sequences in the model's latent space. With SLiC, decoding heuristics become unnecessary and decoding candidates' quality significantly improves regardless of the decoding method. Furthermore, SLiC shows no sign of diminishing returns with model scale, and presents alternative ways to improve quality with limited training and inference budgets. With SLiC, we exceed or match SOTA results on a wide range of generation tasks spanning abstractive summarization, question generation, abstractive question answering and data-to-text generation, even with modest-sized models.