Elamur Research⌕
☰
home / papers / 2012 / Exploiting Discourse Relations for Sentiment Analysis

Exploiting Discourse Relations for Sentiment Analysis

Summary

The paper addresses document-level polarity classification of Chinese product reviews. Lexicon-style approaches count every clause equally, but discourse structure can shift the overall verdict: a concluding clause introduced by "in general" can outweigh earlier praise. The authors split each review into sub-sentences at punctuation marks. Each sub-sentence gets a polarity from the Basic SELC model of Qiu et al. (2009) and a discourse type (head or modifier of a relation). The types are assigned through explicit connectives drawn from a Chinese discourse relation hierarchy built on Rhetorical Structure Theory (RST), which the authors compiled themselves. A linear-kernel SVM learns one weight per discourse type plus an offset, and the document score is the weighted sum of sub-sentence polarities. The authors compare weighting schemes of increasing granularity and a greedy search that keeps only the most influential relations or connectives. On reviews collected from 360buy, every discourse-aware variant beats the baselines, and the greedy relation method performs best. The study suggests that shallow, connective-based discourse cues improve Chinese sentiment analysis even when no discourse parser is available.

Key points

  • The authors present what they describe as the first fairly complete Chinese discourse relation hierarchy with associated connectives. It includes multi-nucleus relations such as Coordinate and Temporal, and single-nucleus relations such as Contrast, Cause and Condition, each split into H (head) and M (modifier) types.
  • Data come from 360buy reviews, with labels assigned automatically: the "strong points" column counts as positive and the "shortcomings" column as negative. Word-level HowNet scoring reaches only 56.9% F, whereas the sub-sentence baseline reaches 83.55%, rising to 83.86% with the offset.
  • Five weighting variants were tested: SNSS, SNMS, MNMS, GDR and GCW. All outperform both baselines, with significance at p<0.01. MNMS does slightly worse than SNMS, possibly because it has too many weights to train.
  • GDR achieves the best overall F of 85.28%, which is 1.42 points above baseline2. Its greedy search selects Contrast, Cause, Condition and Generalization as the influential relations.
  • GCW relies on just six connectives (e.g., 但是|but, 虽然|although, 总的来说|in general) and still reaches 85.13% F.

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: Linguistics & languages

Topics: Discourse analysis

Cited by

6 papers of our library cite it (of the 140673 whose bibliographies we read); 2 of them have a page here, 4 not yet.

Cite

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

@inproceedings{wang2012exploiting,
  title = {{Exploiting Discourse Relations for Sentiment Analysis}},
  author = {Wang, Fei and Wu, Yunfang and Qiu, Likun},
  booktitle = {Proceedings of COLING 2012: Posters},
  year = {2012},
  url = {https://aclanthology.org/C12-2128/}
}

References

14 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Heerschop, B., Goossen, F., Hogenboom, A., Frasincar, F., Kaymak, U. and de Jong, F. (2011). Polarity analysis of texts using discourse structure. In Proceedings of CIMK-2011, pages 1061-1070. find
  2. Hernault, H., Bollegala, D. and Ishizuka, M. (2010). A semi-supervised approach to improve classification of infrequent discourse relations using feature vector extension. In Proceedings of the EMNLP-2010, pages 399-409. find our card
  3. Joty, S., Carenini, G. and Ng, R.T. (2012). A novel discriminative framework for sentence-level discourse analysis. In Proceedings of EMNLP-2012, page 904-915 find
  4. Mann, W.C. and Thompson, S.A. (1988). Rhetorial structure theory: towards a functional theory of text organization. Text, 8(3):243-281. find
  5. Narayanan, R., Liu, B. and Choudhary, A. (2009). Sentiment analysis of conditional sentences. In Proceedings of EMNLP-2009, pages 180-189. find
  6. Pang, B. and Lee, L. (2008). Opinion mining and sentiment analysis. Information Retrieval, 2(1-2): 1-135. find
  7. Polanyi, L. and Zaenen, A. (2006). Contextual valence shifters, Computing attitude and affect in text: Theory and applications, page 1-10. find
  8. Qiu, L., Zhang, W., Hu, C. and Zhao, K. (2009). SELC: a self-supervised model for sentiment classification. In Proceedings of CIKM-2009, pages 929-936. find
  9. Somasundaran, S., Namata, G. , Wiebe, J. and Getoor, L. (2009). Supervised and unsupervised methods in employing discourse relations for improving opinion polarity classification, In Proceedings of the ENMLP-2009 , pages 170-179. find our card
  10. Somasundaran, S., Wiebe, J. and Ruppenhofer, J. (2008). Discourse level opinion interpretation, In Proceedings of COLING-08, pages 801-808. find
  11. Taboada, M., Voll, K. and Brooke, J. (2008). Extracting sentiment as a function of discourse structure and topically. Technical Report No. 2008-20, Simon Fraser University. find
  12. Wang, W.T., Su, J. and Tan, C.L. (2010). Kernel based discourse relation recognition with temporal ordering information. In Proceedings of ACL-2010, pages 710-719. find our card
  13. Yang, Y. and Xue, N. (2012). Chinese comma disambiguation for discourse analysis. In Proceedings of ACL-2012, pages 786-794. find
  14. Zhou, L., Li, B., Gao, W., Wei, Z. and Wong, K.F. (2011). Unsupervised discovery of discourse relations for eliminating intra-sentence polarity ambiguities, In Proceedings of EMNLP-2011, pages 162-171. find our card

↑ back to top