Logic2Text: High-Fidelity Natural Language Generation from Logical Forms

摘要: Previous works on Natural Language Generation (NLG) from structured data have primarily focused surface-level descriptions of record sequences. However, for complex data, e.g., multi-row tables, it is often desirable an NLG system to describe interesting facts logical inferences across records. If only provided with the table, hard existing models produce controllable and high-fidelity generations. In this work, we formulate level as generation forms in order obtain controllable, high-fidelity, faithful We present a new large-scale dataset, \textsc{Logic2Text}, 10,753 involving common logic types paired underlying forms. The show diversified graph structure free schema, which poses great challenges model's ability understand semantics. experiment (1) Fully-supervised training full datasets, (2) Few-shot setting, hundreds examples; compare several popular analyze their performances. hope our dataset can encourage research towards building advanced capable natural, faithful, human-like generation. code are available at https URL.