📄 Arabic Tweets Treebanking and Parsing: A Bootstrapping Approach

Summary

No dependency treebank existed for Arabic tweets, and parsers trained on edited text such as PATB handle noisy, dialect-tinged social media poorly. The authors build a tweet treebank without any human annotation by combining a slow rule-based chart parser (RBP) with the fast data-driven MaltParser. RBP has no lexicon and uses only coarse POS tags. It analyses a seed of 1K tweets. When it cannot cover a whole tweet, it returns the largest non-overlapping fragments. Tweets it cannot analyse at all are dropped as nonsense. MaltParser is trained on this seed and then parses 10K tweets. RBP serves as a grammar check: it may build only the links MaltParser proposed, so only rule-conforming trees or subtrees join the training pool. Retraining on this pool lifts accuracy from 64% to 71%. That beats an equally sized RBP-only treebank (68%) and takes about a quarter of the construction time. The work shows that a hand-written grammar paired with a learner can bootstrap a usable treebank for a noisy, low-resource domain.

Key points

• Data: tweets from the Arabian Peninsula, collected through the Twitter Stream API by geographic coordinates because those dialects are close to MSA. The tweets were POS-tagged with the Albogamy & Ramsay (2016) tweet tagger, and 10K were sampled. About 20% of sentences have 25+ words.
• RBP works without a lexicon because tweet vocabulary is so open: even after a million tweets, roughly one word in ten is still unseen.
• RBP-only strategy: a 162K-word treebank took 20K seconds to build, and MaltParser trained on it reached 68%.
• Bootstrapping strategy:
  • RBP built a seed of 15K words in 1K seconds, giving a 64% baseline.
  • MaltParser output, filtered by RBP, covered the remaining 147K words in 4K seconds.
  • Total time was about 5K seconds, and final accuracy was 71%.
• Evaluation:
  • Testing used a reserved 1K-tweet set with 5-fold cross validation, only on tweets RBP can analyse.
  • The metric is reported simply as "accuracy" (not specified as LAS).
  • For comparison, Kong et al. (2014) report 80% on tweets in English, while Seddah et al. (2012) report 67.8% on social media text in French.

Interesting ideas

• I1. The rule parser is run only over the arcs the statistical parser proposed, so a tree is accepted only when both parsers would build it. This yields precise silver data from two very different witnesses.
• I2. Instead of discarding a tweet that fails as a whole, the largest grammatical fragments are kept and added to the training pool. This salvages useful supervision from messy input.
• I3. Grammar failure doubles as an automatic nonsense detector: anything the rules cannot analyse at all never reaches training data.
• I4. Coarse POS tags with no lexicon or subcategorisation still let the grammar produce good analyses when it succeeds. Rough categories can be enough for structure in an open-vocabulary domain.
• I5. Returning maximal fragments naturally splits unpunctuated text that contains several sentences into separate subtrees.

For Mova

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

• I1 (learn): Restrict an explainable rule checker to the arcs a witness parser suggests, and admit to silver data only the subtrees both sides license.
• I2 (tree): Let the perceptron parser and gparse train on agreed partial trees, so noisy sentences still contribute their well-formed parts.
• I3 (learn): Use rejection by the rule layer as a garbage filter before training, and track the rejected share as a sanity signal.
• I4 (tokens): Provide a coarse-tag backoff path for unknown or misspelled words, so downstream parsing degrades gracefully rather than failing.
• I5 (tokens): Use parser fragment boundaries as evidence for sentence segmentation in informal text that lacks punctuation.