Automatic Logical Forms improve fidelity in Table-to-Text generation

\Table-to-text systems generate natural language statements from structured data like tables. While end-to-end techniques suffer from low factual correctness (fidelity), a previous study reported fidelity gains when using manually produced graphs that represent the content and semantics of the target text called Logical Forms (LF). Given the use of manual LFs, it was not clear whether automatic LFs would be as effective, and whether the improvement came from the implicit content selection in the LFs. We present TlT, a system which, given a table and a set of pre-selected table values, first produces LFs and then the textual statement. We show for the first time that automatic LFs improve the quality of generated texts, with a 67% relative increase in fidelity over a comparable system not using LFs. Our experiments allow to quantify the remaining challenges for high factual correctness, with automatic selection of content coming first, followed by better Logic-to-Text generation and, to a lesser extent, improved Table-to-Logic parsing.