Calibrating Sequence likelihood Improves Conditional Language Generation
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.
PDF AbstractDatasets
Task | Dataset | Model | Metric Name | Metric Value | Global Rank | Benchmark |
---|---|---|---|---|---|---|
Abstractive Text Summarization | CNN / Daily Mail | Pegasus | ROUGE-1 | 47.36 | # 3 | |
ROUGE-2 | 24.02 | # 1 | ||||
ROUGE-L | 44.45 | # 3 | ||||
Text Summarization | Reddit TIFU | PEGASUS 2B + SLiC | ROUGE-1 | 32.03 | # 1 | |
ROUGE-2 | 11.13 | # 2 | ||||
ROUGE-L | 25.51 | # 1 | ||||
Text Summarization | SAMSum | PEGASUS 2B + SliC | ROUGE-1 | 54.37 | # 4 | |
ROUGE-2 | 29.88 | # 4 | ||||
ROUGE-L | 45.89 | # 6 |