Deterministic natural language generation from meaning representations for machine translation

Abstract

This paper describes a deterministic method for generating natural language suited to being part of a machine translation system with meaning representations as the level for language transfer. Starting from Davidsonian/Penman meaning representations, syntactic trees are built following the Penn Parsed Corpus of Modern British English, from which the yield (i.e., the words) can be taken. The novel contribution is to highlight exploiting the presentation of meaning content to inform decisions regarding the selection of language constructions: active vs. passive, argument subject vs. expletive it vs. existential there, discourse vs. intra-sentential coordination vs. adverbial clause vs. participial clause vs. purpose clause, and infinitive clause vs. finite clause vs. small clause vs. relative clause vs. it cleft.

Neural sentence generation from formal semantics

“…Recent rule-based approaches to generation using formal semantics and higher-order logic include a type-theoretic system based on Grammatical Framework (GF) (Ranta, 2011) and a system called Treebank Semantics based on event semantics (Butler, 2016).…”

RIGOTRIO at SemEval-2017 Task 9: Combining Machine Learning and Grammar Engineering for AMR Parsing and Generation

“…Inspired by Butler (2016), we use the Stanford JavaNLP utilities Tregex and Tsurgeon (Levy and Andrew, 2006) for the pilot implementation of AMR-to-AST conversion. 5 The difference of our approach is that we convert AMR graphs to abstract instead of concrete syntax trees, and the choice of GF allows for further multilingual text generation, preserving grammatical and semantic accuracy.…”

Is Shortest Always Best? The Role of Brevity in Logic-to-Text Generation

“…Logic-to-Text Generation Logic-to-text generation is the task of generating NL text, starting from a logical formalism (e.g., propositional logic, description logic, or first-order logic). Although the bulk of recent work on NLG (see e.g., Gatt and Krahmer (2018) for a survey) has focused on other areas, generating text from logic nonetheless has a long tradition, with approaches ranging from rule-based methodologies (Wang, 1980;De Roeck and Lowden, 1986;Calder et al, 1989;Shieber et al, 1989;Shemtov, 1996;Carroll and Oepen, 2005;Mpagouli and Hatzilygeroudis, 2009;Coppock and Baxter, 2010;Butler, 2016;Flickinger, 2016;Kasenberg et al, 2019) to statistical (Wong and Mooney, 2007;Lu and Ng, 2011;Basile, 2015) and neural models (Manome et al, 2018;Hajdik et al, 2019;Chen et al, 2020;Liu et al, 2021;Wang et al, 2021;Lu et al, 2022).…”