Automatic Normalisation of Early Modern French

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3354–3366 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3354 Automatic Normalisation of Early Modern French Rachel Bawden1 Jonathan Poinhos2 Eleni Kogkitsidou2 Philippe Gambette2 Benoˆıt Sagot1 Simon Gabay3 1Inria, Paris, France 2LIGM (UMR 8049), Universit´e Gustave Eiffel, CNRS, 77454 Marne-la-Vall´ee, France 3Universit´e de Gen`eve, Switzerland firstname.lastname@{inria.fr,univ-eiffel.fr,unige.ch}, Abstract Spelling normalisation is a useful step in the study and analysis of historical language texts, whether it is manual analysis by experts or automatic analysis using downstream natural language processing (NLP) tools. Not only does it help to homogenise the variable spelling that often exists in historical texts, but it also facilitates the use of off-the-shelf contemporary NLP tools, if contemporary spelling conventions are used for normalisation. We present FREEMnorm, a new benchmark for the normalisation of Early Modern French (from the 17th century) into contemporary French and provide a thorough comparison of three different normalisation methods: ABA, an alignment-based approach and MT-approaches, (both statistical and neural), including extensive parameter searching, which is often missing in the normalisation literature. Keywords: Digital Humanities, Normalisation, Spelling, Modern French, Machine Translation, Historical 1. Introduction Computational approaches have recently been playing an increasing role in the humanities (Gabay, 2021), especially concerning the study of textual documents. Historical documents are particularly interesting, as they are an invaluable source of historical information and are crucial witnesses of language evolution. Whether documents are to be studied manually by philologists and literary experts or analysed automat- ically using downstream natural language processing (NLP) tasks such as part-of-speech (PoS) tagging and parsing, a useful preliminary step is normalisation, which consists in modernising the spelling of the doc- uments to conform to contemporary spelling conven- tions. Normalisation has the effect of (i) reducing spelling variation present in historical documents, of- ten written at a time spelling was not standardised, and (ii) reducing the gap between the historical state of the language and the contemporary state. Importantly, this allows us to apply off-the-shelf NLP tools to old texts and limit the performance drop that can usually be ex- pected, for example for tagging and parsing (Petters- son et al., 2013b) or geographical named entity recog- nition (Kogkitsidou and Gambette, 2020). There has been a considerable amount of previous re- search in historical spelling normalisation, with a range of methods being developed, including manually de- veloped rules (Porta et al., 2013; Baron and Rayson, 2009; Riguet, 2019), those exploiting edit distances and other external resources such as lexicons (Mitankin et al., 2014) and machine translation (MT) approaches, both statistical (Scherrer and Erjavec, 2013; Domingo and Casacuberta, 2018a) and neural (Bollmann and Søgaard, 2016; H¨am¨al¨ainen et al., 2018). Despite this, questions still remain regarding which method is the most effective, particularly between statistical MT (SMT) and neural MT (NMT) approaches. There has for example been little research in optimising these models for the particular task, which could lead to false conclusions being drawn about which model is best; …

Content truncated to 500 words