Cross-Dialect Social Media Dependency Parsing for Social Scientific Entity Attribute Analysis

What it is about

The paper builds a dependency parser for English tweets, a graph-based biaffine parser with a BERTweet encoder fine-tuned on Tweebank v2. It reaches 87.2 UAS and 83.4 LAS on the test set, 3.4 UAS and 4.0 LAS above the previous best. It also measures dialect disparity: the relative error on African American English (AAE) against Mainstream American English (MAE), where a ratio of 1 means no disparity. BERTweet has a ratio of 1.07, and DeBERTa-v3 1.22. On top of the parser, a rule system (TweetIE) extracts how an entity is described: IS_A, HAS_A, AS_AGENT, AS_PATIENT. The case study covers 75,325 tweets about Anthony Fauci. TweetIE yields 74,572 extractions, against 43,097 for ClausIE and 16,980 for ReVerb. Its precision is 83.5% (187 of 222 sampled extractions), against 66.1% for ClausIE and 73.8% for ReVerb. Attributes are ranked by a one-sample t statistic on author political lean, with at least 20 unique users.

Ideas

I1. Negation as a frame flag: the paper adds a "not_" prefix to a relation when a negation token (no, not, never, n't) modifies it, because UD v2 has no neg relation. In Mova, carry Polarity=Neg from en/src/udreg.rs into the frame as a flag, and count negated frames apart.
I2. Port the seven relation patterns as rules over the trees (copula, apposition, possessive, adjective, active and passive). Measure precision on 250 sentences, with system outputs shuffled and hidden from the annotator, as the paper does.
I3. Rank attributes across groups with the same t-based threshold, with a minimum count of distinct speakers. Negative control: shuffle the group labels; few items should pass |t| > 2.
I4. Report the AAE/MAE error ratio for each release of the parser, using the same split. It is a cheap check against the disparity the paper measures.

For Mova

The parser is en/src/parse.rs, and the feature registry is en/src/udreg.rs. The frame and role rules are in prag/src/frames.rs and prag/src/roles.rs. The pronoun rule for a gendered target is a precision-first sieve like coref/src/sieve.rs, and the paper reports 33 of 39 correct pronouns. Measures: LAS on each dialect split (en parse-eval), the precision of the rule extractor on 250 sentences, and the shuffled-label control. Dialect data is covered by the dialect note in languages/en/08-national-varieties.md.

Topics: Dependency parsing

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.48448/sqtx-th07

@inproceedings{eggleston2022cross,
  title = {{Cross-Dialect Social Media Dependency Parsing for Social Scientific Entity Attribute Analysis}},
  author = {Eggleston, Chloe and O'Connor, Brendan},
  booktitle = {Proceedings of the Eighth Workshop on Noisy User-generated Text (W-NUT 2022)},
  year = {2022},
  doi = {10.48448/sqtx-th07},
  url = {https://doi.org/10.48448/sqtx-th07}
}

References

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

Elliott Ash, Germain Gauthier, and Philine Widmer. 2021. Relatio: Text semantics capture political and economic narratives. find
Nikolay Archak, Anindya Ghose, and Panagiotis G. Ipeirotis. 2007. Show me the money!: deriving the pricing power of product features by mining consumer reviews. In Proceedings of the 13th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 56–65. find