Elamur Research⌕
☰
home / papers / 2007 / Learning Multilingual Subjective Language via Cross-Lingual Projections

Learning Multilingual Subjective Language via Cross-Lingual Projections

Summary

The paper asks how to build subjectivity analysis tools (telling subjective sentences from objective ones) for a language with few resources by reusing English resources. Romanian is the test case, but the methods use no Romanian-specific knowledge. The authors try two bridges.

  • Lexicon translation. They translate the English OpinionFinder subjectivity lexicon with bilingual dictionaries and feed the result to a rule-based classifier.
  • Annotation projection. They label the English side of a parallel corpus automatically with OpinionFinder, carry the labels over to the Romanian sentences, and train a Naive Bayes classifier on them.

Human annotation studies show that word-level subjectivity often disappears in translation, mostly because of sense ambiguity. Sentence-level subjectivity survives much better. The projection-based classifier performs best, at 67.85 F-measure on a 504-sentence Romanian gold standard. The baseline there is 54.16%, and the English high-coverage classifier reaches 74.70. The work matters because it gives a cheap recipe for starting subjectivity resources in any language that has a dictionary or parallel text linking it to English.

Key points

  • Translation input. The source is the OpinionFinder lexicon: 6,856 entries, 990 of them multi-word expressions, each tagged strong/weak and with part of speech. It was translated mainly with a 41,500-entry English–Romanian dictionary, with a volunteer-built 4,500-entry UDP dictionary as a fallback.
  • Translation output. Multi-word expressions missing from the dictionaries were translated word by word. They were kept only if they appeared at least three times on the Web (AltaVista counts), which reduced 990 to 264. The final Romanian lexicon has 4,983 entries.
  • Annotation of translated entries. Two native speakers judged 150 random entries in Web contexts. Agreement was 0.80 (κ = 0.70), and 27 entries were wrong translations. Of the 123 correct entries, 49.6% were subjective, 23.6% mainly objective and the rest mixed.
  • Effect of clue strength. English strong clues turned objective in Romanian less often than weak ones (14.3% vs. 39.0%).
  • Projection training data. After removing the test sentences, 10,628 training sentences were labeled automatically. Projections from OpinionFinder's high-coverage classifier beat those from its high-precision one (67.85 vs. 64.68 F-measure). The rule-based lexicon classifier still had the best precision on subjective sentences.

Interesting ideas

  • I1. Web frequency as a filter for guessed translations. Composing a phrase word by word and keeping it only if it appears on the Web a few times is a cheap way to drop invented phrases. It could clean up any dictionary-based transfer of multi-word terms.
  • I2. Evaluating a lexicon without a target-language corpus. Annotators judge each entry by its typical Web usage, with an extra label for wrong translations. This separates translation errors from loss of meaning and works where no labeled corpus exists.
  • I3. Using source-side confidence when transferring resources. Strong English clues kept their subjectivity far more often than weak ones. Transferring only high-reliability items, or weighting them more, could improve projected lexicons for other tasks.
  • I4. Projecting sentence labels rather than word lists. Whole-sentence translations carry context that resolves word ambiguity. For meaning-sensitive properties, annotation projection over parallel text may be more reliable than translating isolated words.
  • I5. Preferring more noisy labels over fewer clean ones. The larger, noisier training set from the high-coverage labeler beat the smaller, cleaner high-precision set. This trade-off is worth testing whenever weak supervision is used to bootstrap a new language.

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

Fields: Natural language processing

Cited by

93 papers of our library cite it (of the 140884 whose bibliographies we read); 3 of them have a page here, 90 not yet.

Cite

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

@inproceedings{mihalcea2007learning,
  title = {{Learning Multilingual Subjective Language via Cross-Lingual Projections}},
  author = {Mihalcea, Rada and Banea, Carmen and Wiebe, Janyce},
  booktitle = {Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics},
  year = {2007},
  url = {https://aclanthology.org/P07-1123/}
}

References

23 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Alina Andreevskaia and Sabine Bergler. Mining wordnet for fuzzy sentiment: Sentiment tag extraction from WordNet glosses. In Proceedings of EACL 2006. find
  2. Cecilia Ovesdotter Alm, Dan Roth, and Richard Sproat. 2005. find
  3. Emotions from text: Machine learning for text-based emotion prediction. In Proceedings of HLT/EMNLP 2005. find our card
  4. Krisztian Balog, Gilad Mishne, and Maarten de Rijke. 2006. Why are they excited? identifying and explaining spikes in blog mood levels. In EACL-2006. find
  5. Andrea Esuli and Fabrizio Sebastiani. 2006. Determining term subjectivity and term orientation for opinion mining. In Proceedings the EACL 2006. find
  6. Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In Proceedings of ACM SIGKDD. find
  7. Yi Hu, Jianyong Duan, Xiaoming Chen, Bingzhen Pei, and Ruzhan Lu. 2005. A new method for sentiment classification in text retrieval. In Proceedings of IJCNLP. find
  8. Radu Ion. 2007. Methods for automatic semantic disambiguation. Applications to English and Romanian. Ph.D. thesis, The Romanian Academy, RACAI. find
  9. Hiroshi Kanayama and Tetsuya Nasukawa. 2006. Fully automatic lexicon expansion for domain-oriented sentiment analysis. In Proceedings of EMNLP 2006. find
  10. Soo-Min Kim and Eduard Hovy. 2006. Identifying and analyzing judgment opinions. In Proceedings of HLT/NAACL 2006. find
  11. Levon Lloyd, Dimitrios Kechagias, and Steven Skiena. 2005. Lydia: A system for large-scale news analysis. In Proceedings of SPIRE 2005. find
  12. George Miller, Claudia Leacock, Tangee Randee, and Ross Bunker. 1993. A semantic concordance. In Proceedings of the DARPA Workshop on Human Language Technology. find
  13. Bo Pang and Lillian Lee. 2004. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings of ACL 2004. find our card
  14. Bo Pang, Lillian Lee, and Shivakumar Vaithyanathan. 2002. Thumbs up? Sentiment classification using machine learning techniques. In Proceedings of EMNLP 2002. find our card
  15. Ellen Riloff and Janyce Wiebe. 2003. Learning extraction patterns for subjective expressions. In Proceedings of EMNLP 2003. find our card
  16. Hiroya Takamura, Takashi Inui, and Manabu Okumura. 2006. Latent variable models for semantic orientations of phrases. In Proceedings of EACL 2006. find
  17. Peter Turney. 2002. Thumbs up or thumbs down? Semantic orientation applied to unsupervised classification of reviews. In Proceedings of ACL 2002. find our card
  18. Universal Dictionary. 2007. Available at www.dicts.info/uddl.php. find
  19. Janyce Wiebe and Rada Mihalcea. 2006. Word sense and subjectivity. In Proceedings of COLING-ACL 2006. find
  20. Janyce Wiebe and Ellen Riloff. 2005. Creating subjective and objective sentence classifiers from unannotated texts. In Proceedings of CICLing 2005 (invited paper). Available at www.cs.pitt.edu/mpqarequest. find
  21. Janyce Wiebe, Theresa Wilson, and Claire Cardie. 2005. Annotating expressions of opinions and emotions in language. Language Resources and Evaluation, 39(2/3):164– 210. Available at www.cs.pitt.edu/mpqa. find
  22. Theresa Wilson, Janyce Wiebe, and Paul Hoffmann. 2005. Recognizing contextual polarity in phrase-level sentiment analysis. In Proceedings of HLT/EMNLP 2005. find our card
  23. Hong Yu and Vasileios Hatzivassiloglou. 2003. Towards answering opinion questions: Separating facts from opinions and identifying the polarity of opinion sentences. In Proceedings of EMNLP 2003. find our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mihalcea-2007-learning-multilingual-subjective-language-cross/