Contextualized Embeddings for Enriching Linguistic Analyses on Politeness

What it is about

Politeness studies count words such as please, you or hedges regardless of how they are used. Here every occurrence of a word in the Stanford Politeness Corpus (Wikipedia and Stack Exchange requests, polite/impolite quartiles) gets a contextual BERT vector, and the occurrences are split by a decision-tree-like procedure: pick two centroids, split by distance, keep the split with the best information gain on the polite/impolite labels, recurse to a depth set by the word's frequency. The clusters are readable. Please at the start of a request leans impolite, please in the middle leans polite, but please in the middle after a blunt opening («Would you please stop?») is impolite again; can you after a greeting is polite and after a challenging question impolite; maybe can invite («Maybe we can meet») or dismiss. Cluster ids used as features beat plain politeness-word features in a linear classifier, and the procedure rediscovers known markers such as repeated question marks.

Ideas

• I1. Make Mova's politeness detectors in world/src/talk.rs interact with the opening of the request instead of adding up independently: the weight of please-in-the-middle or can/could you depends on whether the request opens with a greeting, thanks or first person (polite) or with a blunt second-person start or a challenge (impolite). Encode the few pairs the paper found as integer rules with the pair printed in the explanation; check on tale requests that the fired pairs read right.
• I2. Split a cue word's uses with Mova's own memory model rather than BERT: for each frequent cue in prag/data/cues.tsv (maybe, please, think, just), grow a small IGTree on the existing context slots (position, opening word, form, you, neg) with the silver labels (hedge, polarity, act) and keep only branches that are pure and well supported. The tree is the readable analysis the paper gets from clustering. Measure hedge and polarity macro-F1 with the split cues against the plain ones.
• I3. Add please position (initial, medial, final) as a feature value instead of the present yes/no please slot in prag/src/feats.rs; the start/middle contrast is one of the strongest politeness facts. Measure request and polarity scores.

For Mova

32. Politeness and hedging §2 (politeness in English grammar), §6 (computational politeness, Mova's detectors), §7 (hedging). Code: world/src/talk.rs, prag/src/feats.rs, prag/data/cues.tsv, prag/src/model.rs (IGTree). Measures: hedge/polarity/request macro-F1 in prag eval; checked examples of fired politeness pairs.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/2020.coling-main.198

@inproceedings{aljanaideh2020contextualized,
  title = {{Contextualized Embeddings for Enriching Linguistic Analyses on Politeness}},
  author = {Aljanaideh, Ahmad and Fosler-Lussier, Eric and Marneffe, Marie-Catherine de},
  booktitle = {Proceedings of the 28th International Conference on Computational Linguistics},
  year = {2020},
  doi = {10.18653/v1/2020.coling-main.198},
  url = {https://doi.org/10.18653/v1/2020.coling-main.198}
}

References

14 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Aubakirova, M., & Bansal, M. (2016). Interpreting Neural Networks to Improve Politeness Comprehension. In Empirical Methods in Natural Language Processing. find
2. Bamman, D., Eisenstein, J., & Schnoebelen, T. (2014). Gender Identity and Lexical Variation in Social Media. Journal of Sociolinguistics, 18, 135–160. find
3. Brown, P., & Levinson, S. C. (1987). Politeness: Some Universals in Language Usage volume 4. Cambridge University Press. find
4. Danescu-Niculescu-Mizil, C., Sudhof, M., Jurafsky, D., Leskovec, J., & Potts, C. (2013). A Computational Approach to Politeness with Application to Social Factors. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics. find
5. Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of The 2019 Conference of the North American Chapter of the Association for Computational Linguistics. find
6. Hyland, K. (2018). Metadiscourse: Exploring Interaction in Writing. Bloomsbury Publishing. find
7. Liu, B., Hu, M., & Cheng, J. (2005). Opinion Observer: Analyzing and Comparing Opinions on the Web. In Proceedings of the 14th international conference on World Wide Web (pp. 342–351). find
8. Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013a). Efficient estimation of word representations in vector space. In International Conference on Learning Representations Workshop: Workshops Track. find
9. Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013b). Distributed Representations of Words and Phrases and their Compositionality. In Advances in Neural Information Processing Systems (pp. 3111–3119). find
10. Niu, T., & Bansal, M. (2018). Polite Dialogue Generation without Parallel Data. Transactions of the Association for Computational Linguistics, 6, 373–389. find
11. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V. et al. (2011). Scikit-learn: Machine learning in Python. Journal of machine learning research, 12, 2825–2830. find
12. Pennington, J., Socher, R., & Manning, C. (2014). Glove: Global Vectors for Word Representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) (pp. 1532–1543). find
13. Peters, M. E., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., & Zettlemoyer, L. (2018). Deep Contextualized Word Representations. In Proceedings of The North American Chapter of the Association for Computational Linguistics. find
14. Quinlan, J. R. (1986). Induction of Decision Trees. Machine learning, 1, 81–106. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/aljanaideh-2020-contextualized-embeddings-enriching-linguistic-analyses/