DCU/TCD-FORGe at WebNLG’23: Irish rules! (WegNLG 2023)

Proceedings of the Workshop on Multimodal, Multilingual Natural Language Generation and Multilingual WebNLG Challenge (MM-NLG 2023), pages 87–92 September 12, 2023 ©2023 Association for Computational Linguistics DCU/TCD-FORGe at WebNLG’23: Irish rules! Simon Mille ADAPT, Dublin City University simon.mille@adaptcentre.ie Elaine Uí Dhonnchadha Trinity College, Dublin uidhonne@tcd.ie Stamatia Dasiopoulou Independent Researcher stamatia.dasiopoulou@gmail.com Lauren Cassidy ADAPT, Dublin City University lauren.cassidy@adaptcentre.ie Brian Davis ADAPT, Dublin City University brian.davis@adaptcentre.ie Anya Belz ADAPT, Dublin City University anya.belz@adaptcentre.ie Abstract In this paper, we describe the submission of Dublin City University (DCU) and Trinity Col- lege Dublin (TCD) for the WebNLG 2023 shared task. We present a fully rule-based pipeline for generating Irish texts from DB- pedia triple sets which comprises 4 compo- nents: triple lexicalisation, generation of non- inflected Irish text, inflection generation, and post-processing. Our whole pipeline is avail- able at https://github.com/mille-s/DCU_ TCD-FORGe_WebNLG23 1 Introduction The WebNLG dataset is a benchmark for data-to- text Natural Language Generation (NLG) consist- ing of {input, output} pairs, where the input is a set of n triples (1 ≤ n ≤ 7) and the output a set of m texts that verbalise each triple set. The triples are extracted from DBpedia and represent relation- ships between DBpedia resources, namely subjects (DB-Subj) and objects (DB-Obj), via respective properties e.g. for the property country: Texas | country | United_States. The WebNLG’23 shared task focuses on four languages for which the existing resources are limited: Irish, Welsh, Breton and Maltese; we submitted Irish outputs only. Large Language Models are becoming more and more used for NLG, but it is well known that they are heavily dependent on the quantity and quality of data they are trained on. On the other hand, rule-based systems although limited in terms of coverage and/or fluency are usually eas- ier to adapt to low-resource languages. For our DCU/TCD-FORGe submission, our pipeline con- sists of 4 fully rule-based modules, which are de- scribed in the remainder of the paper: Lexicalisa- tion of input triples: instantiation of predicate- argument templates with DB-Subj and DB-Obj values (Section 2); Generation of non-inflected Irish sentences: a sequence of graph transduc- ers that progressively specify the linguistic struc- tures into their surface-oriented form (FORGe, Section 3); Generation of morphological inflec- tions: finite-state transducers to produce inflections via two-level morphology (Section 4); and Post- processing: cleaning and formatting of the text (Section 5). DCU/TCD-FORGe uses disk space of ∼8MB and runs with less than 1GB of RAM; it generates the WebNLG test set (1,779 texts) in ∼15 min (∼0.5 sec/text) and achieved 0.167 BLEU. 2 Lexicalisation of Input Triples In this section, we detail the steps involved in the lexicalisation of the WebNLG inputs: we start with the predicate-argument (PredArg) templates and their instantiation (2.1), and continue with the adap- tations carried out for Irish generation (2.2). 2.1 PredArg templates and their instantiation For lexicalisation, we follow the approach of the FORGe submission at WebNLG’19 (Mille et al., 2019), i.e. we use PredArg templates in the Prop- Bank style (Kingsbury and Palmer, 2002) that cor- …

Content truncated to 500 words