LLMs have shown impressive few-shot performance across many tasks. However, they still struggle when it comes to generating complex output structures, such as those required for Information Extraction. This limitation stems from the fact that LLMs, without finetuning, tend to generate free text rather than precise structures that follow a specific grammar. In this work, we propose to enrich the decoding step with formal grammar constraints. During beam search, only valid token continuations compliant with the grammar production rules are considered. This enforces the generation of valid sequences exclusively. Our framework is highly general and flexible, allowing any Context-Free Grammar (CFG) to be integrated into our custom constrained beam search implementation. We demonstrate that the outputs of many NLP tasks can be represented as formal languages, making them suitable for direct use in our framework. For task where the output space is dependent on the input, we propose input-dependent grammars to constrain the generation. We conducted experiments with two challenging tasks involving large alphabets in their grammar (Wikidata entities and relations): information extraction and entity disambiguation. Our results with LLaMA models clearly indicate that grammar-constrained decoding outperforms few-shot prompting without constraints, and even competes with task-specific finetuned models. These findings suggest that integrating grammar-based constraints during decoding holds great promise in making LLMs reliably produce structured outputs, especially in setting where training data is scarce and finetuning is expensive.

翻译：暂无翻译