📄 Towards a Crowd-Sourced WordNet for Colloquial English

Summary

Princeton WordNet is revised rarely, so it misses much of the slang, new coinages and vulgar vocabulary that spreads quickly on Twitter. Colloquial WordNet is an add-on to PWN for this kind of vocabulary. Its first release, built mostly by a single lexicographer, has 428 entries. This project report describes how the authors opened the resource to a wider group of volunteer contributors without giving up WordNet's structure.

They wrote annotation guidelines, added self-rated confidence levels so non-experts can take part, and simplified a Web editor that shows the guidelines inside the editing view. Work is handed out through personal queues of candidate terms. Candidates come from a filtered, year-long stream of English tweets and are ranked by how much more frequent they are in tweets than in a general Web corpus. Each candidate comes with automatically chosen example tweets.

The resource has not yet grown beyond its first version. The report's contribution is the platform and the method. The aim is wide coverage of social-media English for processing noisy user text, plus a feed of candidate entries for future PWN releases.

Key points

• Collection used the Twitter 1% "sample" streaming endpoint from August 2016. By August 2017 it held 435 million English tweets, and heuristic filtering reduced this to 34,776,298 tweets. The filters drop a tweet when:
  • over 30% of its words are 1–2 characters long,
  • it has more than two newlines,
  • over 30% of its words are not in a dictionary, or
  • over 30% of its words begin with # or @.
• Annotators rate how familiar they are with a term: Very Strong, Strong, Medium, Weak or Skip. Skipped terms go back to the pool for someone else.
• Each term gets a status:
  • Real lexemes are General, Novel or Vulgar.
  • Abbreviation, Misspelling, Inflected Form, Name, Not Lexical and Error are kept only as ancillary data.
  • Abbreviations, misspellings and inflected forms point to their canonical lemma.
• Every new word needs at least one link, usually to a PWN synset through the Interlingual Index (ILI). A link to another Colloquial WordNet entry also counts, as with "subtweet" linked to "tweet".
• The editor is built with Scalatra, SQLite, Bootstrap and Angular. Terms in a queue are locked to one annotator for 7 days. Planned specialised subtasks use two existing corpora: a gender-minority corpus covering 49 days in 2015 and a pro-anorexia hashtag corpus covering December 2012 to October 2015.

Interesting ideas

• I1. Annotators state how well they know a term, and anything they cannot handle goes back to the pool instead of receiving a forced guess, so low-confidence entries can be found and reviewed later.
• I2. Before any lexical work, each candidate string is sorted into a real word, a typo, an abbreviation, an inflected form, a name or a meaningless fragment, and the non-word cases are mapped to a canonical lemma.
• I3. Cheap ratio rules (share of very short words, non-dictionary words or hashtag/mention tokens, plus a newline count) removed most bot, spam and non-English material at almost no cost.
• I4. Example sentences are picked by a GDEX-style score: moderate length, proper capitalisation and final punctuation, mostly common words, and no banned tokens. Selection is greedy and forbids heavy word overlap with sentences already picked.
• I5. New vocabulary is found by comparing its frequency against a background corpus and keeping mostly-lowercase forms to exclude names. Each accepted entry must then link to at least one existing synset through the ILI.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): LLM teacher tiers could attach a similar familiarity or confidence tag to each silver label, sending "weak" or "skip" items to a stronger teacher or to human review instead of training on them directly.
• I2 (tokens): the en crate could use this same triage for unknown tokens in informal text, normalising misspellings and abbreviations like "IDK" to known lemmas before tagging, and recording which route was taken for explainability.
• I3 (learn): such filters could screen raw text before it reaches the witness parsers, so silver-data agreement is not wasted on fragments that no parser can analyse sensibly.
• I4 (answer): the same scoring could pick clean, varied illustrative sentences when building school-style test items or explaining a word's meaning to the user.
• I5 (world): when Mova meets a word missing from its WordNet-based ontology, it could require a link to an existing synset, so the new entity still inherits classes and relations for reasoning.