Tagging Icelandic Text using a Linguistic and a Statistical Tagger

Hrafn LoftssonNAACL 2007

TL;DR

This paper introduces IceTagger, a novel linguistic rule-based tagger for Icelandic, a morphologically complex language. IceTagger achieves a higher tagging accuracy (91.54%) than the state-of-the-art statistical tagger TnT (90.44%), demonstrating the effectiveness of linguistic rules combined with heuristics for this challenging task.

What Problem Does This Paper Address?

Part-of-speech (POS) tagging, the process of assigning grammatical tags to words in a text, is a fundamental task in Natural Language Processing (NLP). However, languages like Icelandic, with their rich morphology (complex word structures and inflections), pose significant challenges for traditional tagging methods. Statistical taggers, while powerful, can struggle with the vast number of possible word forms and their subtle grammatical nuances. Linguistic rule-based taggers, on the other hand, can be difficult and time-consuming to develop. This paper addresses the need for an accurate and efficient POS tagger for Icelandic by proposing a hybrid approach that combines linguistic rules with heuristics and a morphological analyzer.

Methods

The paper presents IceTagger, a linguistic rule-based tagger (LRBM) for Icelandic. Its core components include:

  • Local Rules: Approximately 175 rules are used for initial disambiguation. These rules operate on a small window of words around the target word and eliminate inappropriate tags based on lexical forms or tag features.
  • Heuristics: A set of global heuristics are applied for further disambiguation. These heuristics go beyond the immediate word neighborhood to perform tasks like identifying grammatical functions, guessing prepositional phrases, and enforcing feature agreement (e.g., subject-verb agreement).
  • Unknown Word Guesser (IceMorphy): This morphological analyzer is used to handle words not present in the tagger's lexicon. It generates all appropriate tags for a given word by performing morphological and ending analysis. A key feature is its ability to fill "tag profile gaps" – missing tags in a word's potential tag set – by inferring them from morphological suffixes and case information.

The paper also compares IceTagger to TnT, a well-known second-order (trigram) Hidden Markov Model statistical tagger.

Datasets

The evaluation uses the Icelandic Frequency Dictionary (IFD) corpus. The authors created ten pairs of training and test corpora from IFD. Nine of these test corpora were used for evaluation, while the tenth served as a development corpus for IceTagger. The training corpora (90% of IFD) were used to deduce the lexicons for TnT and IceTagger.

Key Insights and Ideas

  • Linguistic Rules Can Outperform Pure Statistical Methods: For morphologically rich languages like Icelandic, a well-crafted set of linguistic rules, especially when combined with heuristics, can achieve higher accuracy than purely statistical approaches like TnT.
  • Heuristics are Crucial for Disambiguation: The global heuristics in IceTagger play a significant role in improving accuracy by enforcing grammatical agreement and identifying broader linguistic structures, going beyond local context.
  • Morphological Analysis is Key for Unknown Words: IceMorphy's ability to intelligently guess tags for unknown words by leveraging morphological information significantly boosts the performance of the overall system, even improving the statistical TnT tagger when its lexicon gaps are filled.
  • Development Time is Manageable: The authors highlight that developing their LRBM (including tokenization, morphological analysis, and tagging) took only 7 man-months, suggesting that linguistic rule-based approaches can be developed in a reasonable timeframe, especially with a focus on heuristics.

Key Results

  • IceTagger Accuracy: Achieved an average tagging accuracy of 91.54%.
  • TnT Accuracy: Achieved an average tagging accuracy of 90.44%.
  • IceTagger vs. TnT: IceTagger made 11.5% fewer errors than TnT.
  • TnT with IceMorphy: When TnT's lexicon gaps were filled by IceMorphy, its accuracy increased to 91.18%. In this scenario, IceTagger still made 4.1% fewer errors than the enhanced TnT.
  • Base Tagger Accuracy: A simple baseline tagger (assigning most frequent tag) achieved 76.27% accuracy.
  • Speed: TnT is significantly faster, tagging about 50,000 tokens/sec, while IceTagger tags about 2,700 tokens/sec.

Limitations and Future Work

  • Speed: IceTagger is considerably slower than TnT, which is a limitation for real-time applications.
  • Component Improvement: The authors suggest that further improvements can be made by enhancing individual components of both IceTagger and IceMorphy.
  • Tagger Combinations: They also plan to explore various tagger combination methods to further boost Icelandic tagging accuracy.

Summary updated

BibTeX
@inproceedings{loftsson2007tagging,
  author = {Hrafn Loftsson},
  title = {{Tagging Icelandic Text using a Linguistic and a Statistical Tagger}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2007},
}
Figure