Difficulties in Handling Mathematical Expressions in Universal Dependencies

What it is about

Levine surveys how mathematical expressions are handled in UD English corpora. A regular expression over numbers and operators finds 886 sentences in UD 2.11, which is 0.05% of the sentences, spread over 51 corpora. In a sample of ACL papers, 68% of documents contain some mathematical expression and 3.7% of sentences do. The paper compares three ways to analyse such expressions: as multi-word expressions, as code-switching, or as natural language. It argues for the third, with a short list of recommendations, for example treating simple arithmetic as English-like coordination and predicate operators such as = and < as predicates. Precision of the search query on a sample is 0.93 and recall 0.88.

Ideas

• I1. en/src/tok.rs and en/src/ud.rs: treat variables joined by adjacency (xy) as conj and arithmetic operators as cc or predicate heads, as the paper suggests; measure LAS on the 886-sentence math subset with the en ud-eval command, before and after (test only if the source is not GUM).
• I2. en/seeds/lexicon: add predicate-operator entries (=, <, >) with a verbal tag, then measure UPOS accuracy of en tag-eval on math-bearing sentences.
• I3. annot silver: count math-bearing sentences per 1k sentences in corpus/, so the share of math in silver is known before training; a prevalence check like the paper's.

For Mova

Math is rare in the English treebanks but common in the ACL-style text that Mova's RAG work reads, so the handling matters more for documents than for the tree-bank score. Relevant chapters are languages/en/07-punctuation-capitals-numbers and 12-pronouns-determiners-numerals. Measure first the prevalence in Mova's own corpus, and the LAS of en on the math subset of EWT test.

Topics: Universal Dependencies

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/2023.law-1.3

@inproceedings{levine2023difficulties,
  title = {{Difficulties in Handling Mathematical Expressions in Universal Dependencies}},
  author = {Levine, Lauren},
  booktitle = {Proceedings of the 17th Linguistic Annotation Workshop (LAW-XVII)},
  year = {2023},
  doi = {10.18653/v1/2023.law-1.3},
  url = {https://doi.org/10.18653/v1/2023.law-1.3}
}

References

5 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Jack Rueter, Niko Partanen, and Flammie A. Pirinen. 2021. Numerals and what counts. In Proceedings of the Fifth Workshop on Universal Dependencies (UDW, SyntaxFest 2021), pages 151–159, Sofia, Bulgaria. Association for Computational Linguistics. find
2. Mark Davies. 2010. The Corpus of Contemporary American English as the first reliable monitor corpus of English. Literary and Linguistic Computing, 25(4):447–464. Abbreviation acl cc compound conj fixed flat:foreign nmod:npmod nsubj nummod obj obl punct xcomp Ivan A Sag, Timothy Baldwin, Francis Bond, Ann Copestake, and Dan Flickinger. 2002. Multiword expressions: A pain in the neck for NLP. In International conference on intelligent text processing and computational linguistics, pages 1–15. Springer. find our card
3. Manuela Sanguinetti, Cristina Bosco, Lauren Cassidy, Özlem Çetinoğlu, Alessandra Teresa Cignarella, Teresa Lynn, Ines Rehbein, Josef Ruppenhofer, Djamé Seddah, and Amir Zeldes. 2022. Treebanking user-generated content: a UD based overview of guidelines, corpora and unified recommendations. Language Resources and Evaluation, pages 1–52. find
4. Nathan Schneider and Amir Zeldes. 2021. Mischievous nominal constructions in Universal Dependencies. In Proceedings of the Fifth Workshop on Universal Dependencies (UDW, SyntaxFest 2021), pages 160–172, Sofia, Bulgaria. Association for Computational Linguistics. *Table 1: Abbreviations of dependency relations discussed in this paper.* Amir Zeldes. 2017. The GUM corpus: creating multilayer resources in the classroom. Language Resources and Evaluation, 51(3):581–612. find our card
5. B Daniel Zeman. 2021. Date and time in Universal Dependencies. In Proceedings of the Fifth Workshop on Universal Dependencies (UDW, SyntaxFest 2021), pages 173–193, Sofia, Bulgaria. Association for Computational Linguistics. A Relation clausal modifier of noun coordinating conjunction compound conjunct fixed multiword expression foreign words NP as adverbial modifier nominal subject numeric modifier object oblique nominal punctuation open clausal complement Dependency Relation Proportions for Mathematical Operators Query to Identify Mathematical Expressions This section gives a description of the search criteria in the query we used to identify mathematical expression in UD corpora. Our query identified sentences containing the at least one of following combinations of numerical values and Unicode mathematical operators and symbols: *Table 2: Dependency relation proportions for the most frequent mathematical operators in Universal Dependencies corpora (UD version 2.11, contains 243 treebanks, 138 languages).* 1. At least 1 token that is included in one of the following Unicode blocks: • Mathematical Operators • Supplemental Mathematical Operators • Mathematical Alphanumeric Symbol find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/ud-2023-levine-difficulties-handling-mathematical-expressions-universal/