Controlling hallucinations at word level in data-to-text generation

Abstract Data-to-Text Generation (DTG) is a subfield of Natural Language aiming at transcribing structured data in natural language descriptions. The field has been recently boosted by the use neural-based generators which exhibit on one side great syntactic skills without need hand-crafted pipelines; other side, quality generated text reflects training data, realistic settings only offer imperfectly aligned structure-text pairs. Consequently, state-of-art neural models include misleading statements –usually called hallucinations—in their outputs. control this phenomenon today major challenge for DTG, and problem addressed paper. Previous work deal with issue instance level: using an alignment score each table-reference pair. In contrast, we propose finer-grained approach, arguing that hallucinations should rather be treated word level. Specifically, Multi-Branch Decoder able to leverage word-level labels learn relevant parts instance. These are obtained following simple efficient scoring procedure based co-occurrence analysis dependency parsing. Extensive evaluations, via automated metrics human judgment standard WikiBio benchmark, show accuracy our effectiveness proposed Decoder. Our model reduce hallucinations, while keeping fluency coherence texts. Further experiments degraded version ToTTo could successfully used very noisy settings.