Overview

This is the repository for the software supporting the paper Cross-Dialect Social Media Dependency Parsing for Social Scientific Entity Attribute Analysis by Chloe Eggleston (@nu11us) and Brendan O'Connor (@brendano), Proceedings of the Workshop on Noisy User-generated Text (W-NUT) at COLING 2022.

A copy of the paper is included here; see also the ACL Anthology. Bibtex:

@inproceedings{eggleston-oconnor-2022-cross,
    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)",
    month = oct,
    year = "2022",
    address = "Gyeongju, Republic of Korea",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.wnut-1.4",
    pages = "38--50",
    abstract = "In this paper, we utilize recent advancements in social media natural language processing to obtain state-of-the-art syntactic dependency parsing results for social media English. We observe performance gains of 3.4 UAS and 4.0 LAS against the previous state-of-the-art as well as less disparity between African-American and Mainstream American English dialects. We demonstrate the computational social scientific utility of this parser for the task of socially embedded entity attribute analysis: for a specified entity, derive its semantic relationships from parses{'} rich syntax, and accumulate and compare them across social variables. We conduct a case study on politicized views of U.S. official Anthony Fauci during the COVID-19 pandemic.",
}

Pretrained Models

Pretrained SuPar models are available on the Huggingface Model Hub, with umassnlp/supar-bertweet being the full model and and umassnlp/super-bertweet-eval being the model without relational subtypes for use in the CoNLL evaluation scripts.

TweetIE Pipeline

In order to use TweetIE to analyze a corpus, you first need the following:

• Universal dependency parses, formatted in CoNLL-U format
• Named entity recognition tags, following CoNLL 2003 format
• Part-of-speech tags, either as UPOS tags in the CoNLL-U, or in a seperate file formatted in the same manner as the NER tags

The pipeline takes the following arguments:

• conll_file / ner_conll / pos_conll - UD/NER/POS parse file (POS optional if used in dependency parse file)
• lname - Name of target, can use last name if you want name expanded to first name
• assumed_gender - 'masc'/'fem'/'neuter' for personal pronoun coreference, use None if you wish to disable it
• filter_stops - Enable/disable stopword filtering
• stoplist - Stopword filter list (from NLTK, defaults to "english")
• lemmatize - Enable/disable lemmatization
• lemmatizer - NLTK lemmatizer (defaults to WordNetLemmatizer)
• assume_intersective_adj - Distribute adjectives intersectively (defaults to True)
• assume_intersective_title - Distributes fixed appositions intersectively (defaults to False)
• dep_prefixes / ner_prefixes / pos_prefixes - number of commented out line per entry in said files

The pipeline then can be used with the run(entry_num) on the specific entry in the corpus, and returns an array of tuples of IS_A/HAS_A/AS_AGENT/AS_PATIENT/AS_CONJUNCT and the token indices.

Paper errata

Paragraph 3 of Section 3.4 discusses the construction of the Tweebank v2 AAE and MAE splits of the test set. It should be changed to the following:

In order to measure disparity on the fine-tuning source, we measure the relative error of both the TwitterAAE dependencies and use the Twitter-AAE demographic dialect inference model to extract dialect-specific subsets of the Tweebank v2 test set based on whether the largest demographic proportion was MAE or AAE, yielding 672 and 163 tweets respectively (the remaining 366 tweets belong to neither subset). We also analyze the TwitterAAE dependencies in the same way, which provides 250 tweets of both MAE and AAE respectively.

All per-dialect accuracy results and disparities reported in the paper were evaluated on the above sets of messages.

The published text for this paragraph describes a different splitting method to partition the test set according to whether the AAE or MAE proportion is higher; this criterion results in supersets of the above selections, since the demographic model has four classes and one of the other two sometimes has the largest proportion.