Source-linked AI summary
A Universal Part-of-Speech Tagset
Slav Petrov, Dipanjan Das, Ryan McDonald
TL;DR
The paper addresses the need for comparable POS categories across languages and for grammar induction without target-language gold POS tags. It introduces a twelve-category universal tagset with mappings from 25 treebanks, then demonstrates competitive multilingual grammar induction using automatically projected tags.
Problem
Language-specific POS tagsets hinder comparable multilingual evaluation, while unsupervised grammar induction commonly assumes gold POS tags.
Method
The paper defines twelve universal POS categories, maps 25 treebank tagsets to them, and uses projected universal tags in a Bayesian grammar-induction model.
Results
The resource yields common POS data for 22 languages and supports competitive unsupervised grammar induction without target-language gold POS tags.
Takeaways & Limitations
Universal POS categories generalize across language boundaries and provide a shared resource for multilingual tagging and grammar-induction experiments.
Takeaways & Limitations
The approach adopts a pragmatic tagset despite disagreement about the exact definition of universal POS categories, and plain-text grammar induction remains difficult.
Abstract
from arXiv · showhide
To facilitate future research in unsupervised induction of syntactic structure and to standardize best-practices, we propose a tagset that consists of twelve universal part-of-speech categories. In addition to the tagset, we develop a mapping from 25 different treebank tagsets to this universal set. As a result, when combined with the original treebank data, this universal tagset and mapping produce a dataset consisting of common parts-of-speech for 22 different languages. We highlight the use of this resource via two experiments, including one that reports competitive accuracies for unsupervised grammar induction without gold standard part-of-speech tags.
1 Introduction
The paper proposes a twelve-category universal POS tagset and mappings from 25 treebank tagsets, yielding common POS annotations across 22 languages. It demonstrates the resource in multilingual tagging comparisons and unsupervised grammar induction without target-language gold POS tags.
- The resource defines twelve universal POS categories and maps 25 fine-grained treebank tagsets to them, producing data for 22 languages.
- The resource supports unsupervised and cross-lingual tagger development, multilingual training, and downstream applications without language-specific annotation rules.
- Common categories support cross-language POS accuracy comparisons, since language-specific tagsets can differ in granularity and annotation guidelines.
- The experiments compare POS tagging accuracies across 25 treebanks and use projected universal POS tags for unsupervised grammar induction across multiple languages.
2 Tagset
The tagset pragmatically defines twelve cross-lingual POS categories and operationally maps each treebank’s fine-grained tags to them. The mappings accommodate language-specific annotation differences and remain open to expert refinement.
- The universal tagset contains NOUN, VERB, ADJ, ADV, PRON, DET, ADP, NUM, CONJ, PRT, punctuation, and X.
- Categories are defined operationally by studying each treebank’s POS definitions and annotation guidelines before constructing mappings.
- PennTreebank verb tags map to VERB, while ambiguous TO tags can cause some prepositions to receive the universal PRT tag.
- The mapping handles language-specific category differences, including Korean stative verbs mapped to ADJ and Bulgarian treebanks lacking a determiner category.
- The released resource is an open-source project with version control, allowing experts to refine mappings and add treebanks and languages.
3 Experiments
Experiments evaluate universal-tagset tagging across treebanks and grammar induction using automatically projected POS tags. The induced system remains competitive without target-language gold POS tags, while universal-tagset tagging reduces cross-language variance.
- 3.1 Language Comparisons: Universal-tagset tagging reduces accuracy variance across 25 treebanks from 10.4 to 5.1.
- 3.1 Language Comparisons: Japanese tagging exceeds 99% accuracy with golden segmentation, whereas Turkish tagging reaches 90.2% in the reported comparison.
- 3.1 Language Comparisons: The best tagging results train on original treebank categories and map predictions to universal tags afterward, because universal transitions are less informative.
- 3.2 Grammar Induction: The grammar-induction experiment replaces the usual gold-POS assumption with POS tags automatically projected from English.
- 3.2 Grammar Induction: The experiments use universal syntactic rules within a probabilistic Bayesian grammar-induction model and evaluate directed dependency accuracy.
- 3.2 Grammar Induction: USR-I underperforms USR-G but surpasses DMV for every language and remains competitive with PGI despite using automatically induced universal tags.
4 Conclusions
The paper proposes a twelve-category universal POS tagset with mappings from 25 language-specific tagsets. Experiments show that these categories generalize across languages for unsupervised grammar induction, achieving competitive parsing accuracies without gold POS tags.
- Twelve universal POS categories were mapped from 25 language-specific tagsets.
- Competitive parsing accuracies were achieved on unsupervised grammar induction without relying on gold POS tags.