AggGen: Ordering and Aggregating while Generating

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 1419–1434 August 1–6, 2021. ©2021 Association for Computational Linguistics 1419 AGGGEN: Ordering and Aggregating while Generating Xinnuo Xu†, Ondˇrej Duˇsek‡, Verena Rieser† and Ioannis Konstas† †The Interaction Lab, MACS, Heriot-Watt University, Edinburgh, UK ‡Charles University, Faculty of Mathematics and Physics, Prague, Czechia xx6, v.t.rieser, i.konstas@hw.ac.uk odusek@ufal.mff.cuni.cz Abstract We present AGGGEN (pronounced ‘again’), a data-to-text model which re-introduces two ex- plicit sentence planning stages into neural data- to-text systems: input ordering and input ag- gregation. In contrast to previous work us- ing sentence planning, our model is still end- to-end: AGGGEN performs sentence planning at the same time as generating text by learn- ing latent alignments (via semantic facts) be- tween input representation and target text. Ex- periments on the WebNLG and E2E challenge data show that by using fact-based alignments our approach is more interpretable, expressive, robust to noise, and easier to control, while retaining the advantages of end-to-end sys- tems in terms of fluency. Our code is avail- able at https://github.com/XinnuoXu/ AggGen. 1 Introduction Recent neural data-to-text systems generate text “end-to-end” (E2E) by learning an implicit mapping between input representations (e.g. RDF triples) and target texts. While this can lead to increased fluency, E2E methods often produce repetitions, hallucination and/or omission of important con- tent for data-to-text (Duˇsek et al., 2020) as well as other natural language generation (NLG) tasks (Cao et al., 2018; Rohrbach et al., 2018). Tradi- tional NLG systems, on the other hand, tightly control which content gets generated, as well as its ordering and aggregation. This process is called sentence planning (Reiter and Dale, 2000; Duboue and McKeown, 2001, 2002; Konstas and Lapata, 2013; Gatt and Krahmer, 2018). Figure 1 shows two different ways to arrange and combine the rep- resentations in the input, resulting in widely differ- ent generated target texts. In this work, we combine advances of both paradigms into a single system by reintroducing William Anders dateOfRetirement 1969-09-01 Apollo 8 commander Frank Borman William Anders member_of Apollo 8 Apollo 8 backup_pilot Buzz Aldrin Apollo 8 operator NASA Input DBpedia Triples William Anders served as a crew member on Apollo 8 operated by nasa. The backup pilot was Buzz Aldrin. Frank Borman was also an Apollo 8 commander. William Anders retired on September 1st, 1969. William Anders retired on 1969-09-01. He was a crew member of nasa 's Apollo 8. Frank Borman was also a commander with Buzz Aldrin as the backup pilot. operator backup_pilot commander dateOfRetirement dateOfRetirement operator backup_pilot commander dateOfRetirement commander member_of backup_pilot operator member_of member_of William Anders, who retired on September 1st, 1969, was a crew member on Apollo 8 and served under commander Frank Borman. Apollo 8 was operated by NASA with Buzz Aldrin as backup pilot. Human-authored Text Sentence Plan 1: Generated Target Text 1: Sentence Plan 2: Generated Target Text 2: Figure 1: Two different sentence plans with their cor- responding generated target texts …

Content truncated to 500 words