Priscilla Postma, Rina Donker, Ruth Stam, Athalia Roorda, Andreas van Cranenburgh and Gertjan van Noord
Non-canonical language can be handled in an NLP pipeline using normalization of the input (e.g., MoNoise; van der Goot & van Noord, CLINjournal 2017) or domain adaptation of the pipeline (e.g., Hupkes & Bod, LREC 2016); we focus on the former. MoNoise shows that normalization is effective for social media language. We consider a different domain: Dutch literature from Project Gutenberg. We work with 9 fragments that make up the OpenBoek corpus (van den Berg et al., CLIN 2021). The fragments consist of 10,000+ tokens from texts first published 1860-1920, both translated and originally Dutch.
MoNoise consists of several modules: a lookup table, automatic spelling correction (aspell), and word embeddings; we aim to explore these techniques on our data in future work. Here we report results of a rule-based approach implemented with a sed script (i.e., regular expressions) for normalizing frequently occurring non-standard spellings.
The output consists of instructions to the Alpino parser (van Noord, TALN 2006) to treat words with non-canonical orthography as if they occur with modern spelling. The advantage of this approach is that the resulting parse trees contain the original tokens, and existing annotation layers (such as coreference) do not have to be re-aligned. Consider the following sentence from Couperus, Eline Vere (ch. I, § II):
18-1|- Is het [ @alt zo zoo ] goed ? vroeg zij met bevende stem , [ @alt ene eene ] , van te voren bestudeerde poze aannemende .
Here [ @alt zo zoo ] indicates that the original token zoo should be treated as zo. Besides doubled vowels, other frequent spelling normalizations are de/den, zei/zeide, and mensen/menschen. When multiple alternatives are given the parser considers the input as a lattice and uses the sequence of tokens that generates the most likely parse. Parse trees for the above sentence show that the automatic spelling normalization is not perfect (the correct normalization of eene is een with POS lid rather than ene), but it does lead to a correct bracketing of the NP eene … poze. Furthermore, it turns out that a comma is missing after bestudeerde in the Project Gutenberg etext we use (EBook-No. 19563); the DBNL version of this text (coup002elin01_01) does have this comma—this underscores the importance of professionally edited critical editions.
We will perform an intrinsic evaluation of our spelling normalization pipeline with manually corrected texts and report F1 scores (Reynaert, LREC 2008). We also perform an extrinsic evaluation of downstream tasks: part-of-speech tagging, mention detection, and coreference resolution. Scores for the latter two tasks on Multatuli, Max Havelaar:
mentions lea pron
recall prec f1 recall prec f1 CoNLL acc
original 89.96 81.29 85.40 54.80 47.07 50.64 65.76 55.00
normalized 90.18 82.22 86.02 54.82 45.96 50.00 65.48 54.20
The mention score is improved, which makes sense given that parsing of NPs seems to improve after spelling normalization, but there is a decrease in the coreference metrics, which warrants further investigation.