UDPipe: Trainable Pipeline for Processing CoNLL-U Files Performing Tokenization, Morphological Analysis, POS Tagging and Parsing

What it is about

UDPipe is one binary with one model per language that turns raw text into CoNLL-U: tokenisation, morphological analysis, tagging, lemmatisation and dependency parsing. The tokeniser is a GRU over characters that predicts sentence and token boundaries together. When a treebank lacks SpaceAfter=No, the paper fills it in by comparing counts of word pairs with and without a space in a raw corpus. Tagging uses a suffix-based guesser of analyses with an averaged perceptron, and parsing uses the Parsito transition parser with a small neural classifier, a search-based oracle and a beam of 5. The authors released models for 32 of the 37 UD 1.2 treebanks, and the whole pipeline needs no dictionaries or feature engineering. The paper presents the tool and its trainability, and gives tokenisation results against baselines (Table 1) and parsing results against Ammar et al. (2016).

Ideas

• I1. en/src/tok.rs: for treebanks whose converter cannot align SpaceAfter, fill the missing SpaceAfter=No by the paper's count rule over a raw corpus, and measure token F1 against the aligned original on EWT with the spacing removed.
• I2. en/src/morph.rs: test a suffix guesser for words missing from the dictionary (analyses per fixed-length suffix, with the shortest lemma rule) and measure UPOS accuracy on out-of-dictionary tokens of EWT dev through en tag-eval.
• I3. en beam-curve: check on EWT dev that the beam of 8 is on the flat part of the curve (the 2017 UDPipe paper finds that a large fixed beam beats per-treebank beam tuning), and drop the beam to save time if it is.

For Mova

The ideas touch the tokeniser en/src/tok.rs, the lexicon and guesser in en/src/morph.rs and en/src/ptag.rs, and the beam in en/src/parse.rs. The chapters are languages/en/07-punctuation-capitals-numbers and languages/en/10-word-classes. Measure first I3, since en beam-curve already exists and the answer changes only the speed budget.

Topics: Universal Dependencies · Treebanks · Part-of-speech tagging

Cite

The paper itself, not this card. BibTeX from the library's record of the work.

@inproceedings{straka2016udpipe,
  title = {{UDPipe: Trainable Pipeline for Processing CoNLL-U Files Performing Tokenization, Morphological Analysis, POS Tagging and Parsing}},
  author = {Straka, Milan and Hajič, Jan and Straková, Jana},
  booktitle = {Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16)},
  year = {2016},
  url = {https://aclanthology.org/L16-1680/}
}

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/ud-2016-straka-udpipe-trainable-pipeline-processing-conll/