---
title: "NUS-PT — paper by Chan et al. (2007) — summary and ideas"
description: "This SemEval-2007 system paper from the National University of Singapore tackles supervised word… 10 references, cited by 24 papers here."
authors: ["Yee Seng Chan", "Hwee Tou Ng", "Zhi Zhong"]
year: 2007
venue: "Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007)"
source: "https://aclanthology.org/S07-1054/"
licence: "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)"
licence_url: "https://creativecommons.org/licenses/by-nc-sa/3.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/chan-2007-nus-pt-exploiting-parallel-texts/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/chan-2007-nus-pt-exploiting-parallel-texts/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2007](https://md.research.elamur.ai/free/en/papers/2007/) / NUS-PT: Exploiting Parallel Texts for Word Sense Disambiguation in the English All-Words Tasks

# NUS-PT: Exploiting Parallel Texts for Word Sense Disambiguation in the English All-Words Tasks

**Authors:** [Yee Seng Chan](https://md.research.elamur.ai/free/en/authors/yee-seng-chan/), [Hwee Tou Ng](https://md.research.elamur.ai/free/en/authors/hwee-tou-ng/) and Zhi Zhong · **Year:** 2007 · **Venue:** [Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007)](https://md.research.elamur.ai/free/en/venues/international-workshop-on-semantic-evaluations/) ([International Workshop on Semantic Evaluations 2007](https://md.research.elamur.ai/free/en/venues/international-workshop-on-semantic-evaluations/2007/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY-NC-SA 3.0 (ACL Anthology, before 2016)](https://creativecommons.org/licenses/by-nc-sa/3.0/)

**Links:** [Publisher's page](https://aclanthology.org/S07-1054/) · [PDF](https://aclanthology.org/S07-1054.pdf) · [OpenAlex](https://openalex.org/W2033441998)

**Citations:** 67 citations (OpenAlex, 2026-10-14) · [24](https://md.research.elamur.ai/free/en/papers/chan-2007-nus-pt-exploiting-parallel-texts/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
This SemEval-2007 system paper from the National University of Singapore tackles supervised word sense disambiguation (WSD) for two English all-words tasks. The coarse-grained task covers every content word in five documents using a coarsened WordNet inventory. The fine-grained task covers verbs and the head nouns of their arguments in three documents, using fine-grained WordNet senses. Because larger training sets helped in earlier SENSEVAL rounds, the authors built training data from three sources: SemCor, the DSO corpus, and examples harvested automatically from six English-Chinese parallel corpora. In those corpora, a word's aligned Chinese translation acts as its sense label. For each word they trained an SVM, following Lee and Ng (2002), with local collocation, POS, and surrounding-word features. All sense labels had to be mapped across WordNet versions to WordNet-2.1. The system scored 0.825 on the coarse task, first among official entries, and 0.587 on the fine task, second. Adding the parallel-text examples raised both scores slightly. This shows that translations can serve as cheap sense annotation for frequent, highly polysemous words.

## Key points
- Six English-Chinese corpora from LDC supplied the parallel data: Sinorama, Xinhua News, the English translation of the Chinese Treebank, and three Hong Kong collections (Hansards, News, and Laws).
  - The pipeline was sentence alignment, Chinese word segmentation, and GIZA++ word alignment.
  - Chinese translations were then assigned by hand to each English sense.
- Harvesting from parallel text was limited to frequent Brown Corpus words: 730 nouns, 326 adjectives, and 190 verbs, which together cover about 60% of noun, adjective, and verb tokens.
  - Caps were 1,000 examples per noun or adjective and 500 per verb.
  - Up to 500 DSO examples were added for each of the 70 DSO verbs.
  - All added examples were sampled to follow each word's SemCor sense distribution.
- The SVM features were:
  - 11 local collocations (C−1,−1 … C1,3);
  - 7 POS tags (P−3 … P3);
  - bag-of-unigram context, which may cross sentence boundaries.
  - Syntactic relation features were dropped for efficiency.
  - Words with no training data received the WordNet first sense.
- Results with and without parallel text (PT):
  - **Coarse task:** 0.817 without PT, 0.825 with PT. The first-sense baseline was 0.789, and the system ranked first among 16 participants once the organizer's own system was excluded.
  - **Fine task:** 0.578 without PT, 0.587 with PT, ranking second of 14.
- Senses were mapped as follows:
  - WordNet-1.7.1 to 2.1 using the Daudé et al. (2000) mappings, with manual fixes for the frequent words.
  - DSO's WordNet-1.5 senses mapped to 2.1 entirely by hand.
- The weakest document was d004, a Wikipedia article on computer programming (0.761). There, *program* and *programming* appeared in senses with no training examples, causing 47 errors, or 2.1% of the 2,269 instances.

## Interesting ideas
- A word's translation in an aligned bilingual corpus can label its sense without manual annotation, turning existing parallel text into training data for meaning distinctions.
- When adding automatically harvested examples, sample them to match the sense proportions of a trusted gold corpus, so the noisy source adds volume without skewing the priors.
- Spend manual correction effort only on the most frequent items when mapping between resource versions, and trust the automatic mapping for the long tail.
- Breaking scores down per document exposed domain-specific senses with zero training examples as a concentrated source of error, which a single overall score hid.
- A clear, documented fallback (the WordNet first sense) handles words with no training data, so every instance still gets an answer that can be traced.

## For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

- **I1** (learn): Ukrainian-English parallel text could yield silver WordNet-sense labels for frequent polysemous English words, giving the frame and world stages a disambiguation signal that can be explained by pointing to the aligned translation.
- **I2** (learn): when mixing silver data from witness parsers or LLM teachers with gold treebanks or sense data, resample to gold label distributions so the cheap data does not shift what the model treats as the default reading.
- **I3** (world): when compiling WordNet, FrameNet, or VerbNet knowledge across versions, hand-check the mappings for high-frequency lemmas and classes first, since those dominate real text.
- **I4** (learn): evaluation reports could list, per text genre, the senses or frames observed in test data that have no training support, making coverage gaps visible and explainable.
- **I5** (frames): when VerbNet role assignment has no evidence for a verb, default to its primary sense or class and record that a fallback was used, keeping the step explainable.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [National University of Singapore](https://md.research.elamur.ai/free/en/universities/national-university-of-singapore/) — University, Singapore, Singapore · 180 works · 34,304 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works

🧭 Topics (2): [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,263 works · [🔀 Word sense disambiguation](https://md.research.elamur.ai/free/en/topics/word-sense-disambiguation/) · 83 works

## Cite

Cite as: Yee Seng Chan, Hwee Tou Ng and Zhi Zhong. 2007. NUS-PT: Exploiting Parallel Texts for Word Sense Disambiguation in the English All-Words Tasks. Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007).

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

```bibtex
@inproceedings{chan2007nus,
  title = {{NUS-PT: Exploiting Parallel Texts for Word Sense Disambiguation in the English All-Words Tasks}},
  author = {Yee Seng Chan, Hwee Tou Ng and Zhi Zhong},
  booktitle = {Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007)},
  year = {2007},
  doi = {10.3115/1621474.1621528},
  url = {https://doi.org/10.3115/1621474.1621528}
}
```

- [← previous Person Name Entity Recognition for Arabic](https://md.research.elamur.ai/free/en/papers/shaalan-2007-person-name-entity-recognition-arabic/)
- [next → SemEval-2007 Task 04: Classification of Semantic Relations between Nominals](https://md.research.elamur.ai/free/en/papers/girju-2007-semeval-2007-task-04-classification/)
- [↑ back to papers of 2007](https://md.research.elamur.ai/free/en/papers/2007/)

## References

10 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Yee Seng Chan and Hwee Tou Ng. 2005. Scaling up word sense disambiguation via parallel texts. In Proc. of AAAI05, pages 1037–1042. [find](https://scholar.google.com/scholar?q=Yee+Seng+Chan+and+Hwee+Tou+Ng%2E+2005%2E+Scaling+up+word+sense+disambiguation+via+parallel+texts%2E+In+Proc%2E+of+AAAI05%2C+pages+1037%E2%80%931042%2E)
2. Jordi Daude, Lluis Padro, and German Rigau. 2000. Mapping WordNets using structural information. In Proc. of ACL00, pages 504–511. [find](https://scholar.google.com/scholar?q=Jordi+Daude%2C+Lluis+Padro%2C+and+German+Rigau%2E+2000%2E+Mapping+WordNets+using+structural+information%2E+In+Proc%2E+of+ACL00%2C+pages+504%E2%80%93511%2E)
3. Bart Decadt, Veronique Hoste, Walter Daelemans, and Antal van den Bosch. 2004. GAMBL, genetic algorithm optimization of memory-based WSD. In Proc. of SENSEVAL-3, pages 108–112. [find](https://scholar.google.com/scholar?q=Bart+Decadt%2C+Veronique+Hoste%2C+Walter+Daelemans%2C+and+Antal+van+den+Bosch%2E+2004%2E+GAMBL%2C+genetic+algorithm+optimization+of+memory%2Dbased+WSD%2E+In+Proc%2E+of+SENSEVAL%2D3%2C+pages+108%E2%80%93112%2E)
4. Veronique Hoste, Anne Kool, and Walter Daelemans. 2001. [Classifier optimization and combination in the English all words task](https://md.research.elamur.ai/free/en/papers/hoste-2001-classifier-optimization-combination-english-all/). In Proc. of SENSEVAL-2, pages 83–86.
5. Yoong Keok Lee and Hwee Tou Ng. 2002. [An empirical evaluation of knowledge sources and learning algorithms for word sense disambiguation](https://md.research.elamur.ai/free/en/papers/lee-2002-empirical-evaluation-knowledge-sources-learning/). In Proc. of EMNLP02, pages 41–48.
6. Jin Kiat Low, Hwee Tou Ng, and Wenyuan Guo. 2005. A maximum entropy approach to Chinese word segmentation. In Proc. of the Fourth SIGHAN Workshop on Chinese Language Processing, pages 161–164. [find](https://scholar.google.com/scholar?q=Jin+Kiat+Low%2C+Hwee+Tou+Ng%2C+and+Wenyuan+Guo%2E+2005%2E+A+maximum+entropy+approach+to+Chinese+word+segmentation%2E+In+Proc%2E+of+the+Fourth+SIGHAN+Workshop+on+Chinese+Language+Processing%2C+pages+161%E2%80%93164%2E)
7. George A. Miller, Martin Chodorow, Shari Landes, Claudia Leacock, and Robert G. Thomas. 1994. [Using a semantic concordance for sense identification](https://md.research.elamur.ai/free/en/papers/miller-1994-semantic-concordance-sense-identification/). In Proc. of HLT94 Workshop on Human Language Technology, pages 240–243.
8. Hwee Tou Ng and Hian Beng Lee. 1996. [Integrating multiple knowledge sources to disambiguate word sense: An exemplar-based approach](https://md.research.elamur.ai/free/en/papers/ng-1996-integrating-multiple-knowledge-sources-disambiguate/). In Proc. of ACL96, pages 40–47. [↗](https://doi.org/10.48550/arxiv.cmp-lg/9606032)
9. Hwee Tou Ng, Bin Wang, and Yee Seng Chan. 2003. [Exploiting parallel texts for word sense disambiguation: An empirical study](https://md.research.elamur.ai/free/en/papers/ng-2003-exploiting-parallel-texts-word-sense/). In Proc. of ACL03, pages 455–462.
10. Franz Josef Och and Hermann Ney. 2000. [Improved statistical alignment models](https://md.research.elamur.ai/free/en/papers/och-2000-improved-statistical-alignment-models/). In Proc. of ACL00, pages 440–447.

*Updated 1 November 2026*