📄 An Unsupervised Approach for Bootstrapping Arabic Sense Tagging

Summary

Arabic had no word sense disambiguation (WSD) systems and almost no machine-readable lexical resources. Diab proposes SALAAM (Sense Annotations Leveraging Alignments And Multilinguality), an unsupervised method for bootstrapping Arabic sense tagging. It rests on two ideas. First, translation tends to split the senses of an ambiguous word. Second, words that share a translation share some meaning component. English words aligned to the same Arabic word are grouped into clusters. Their senses are chosen by Resnik's information-theoretic similarity in WordNet, and the chosen English tags are then projected onto the aligned Arabic tokens.

No Arabic gold standard existed, so the SENSEVAL2 English All Words test text was machine-translated by two commercial MT systems. It was then enlarged with genre-similar corpora and aligned with GIZA++. SALAAM reaches 56.9% absolute precision for Arabic. Native speakers judged roughly 90% of the projected tags correct on items whose English tag was correct. The paper also gives an early analysis of how well an English sense inventory fits Arabic.

Key points

• On the English side, SALAAM scores 64.5% precision and 53% recall on the SENSEVAL2 English All Words Task, the best result among unsupervised systems on that task.
• The Arabic gold set consists of the Arabic words aligned to correctly tagged English nouns (671 in the evaluation section, 673 in the conclusion), rated by 3 native speakers.
• The 242-line test text (SV2AW) was translated by both Tarjim and Almisbar, and the outputs were merged. It was padded with BC-SV1, SV2-LS and WSJ to about 226K lines and 5.56M tokens, transliterated with Buckwalter, tokenized, and aligned with GIZA++.
• How well WN17pre fits Arabic:
  • 63.9% of ambiguous Arabic word types are conceptually equivalent to their English translations.
  • 29.3% match only specific senses (138 instances, 87 types).
  • 6.7% are broader than the English word (40 instances, 20 types).
  • EuroWordNet reported 44–55% equivalence for its languages.
• Ambiguity shared by both languages is mostly regular polysemy or metonymy. Shared homonymy appears only in English borrowings. For example, the "tea party" sense of tea has no Arabic counterpart and needs a compound.

Interesting ideas

• I1. Treat a shared translation as evidence of shared meaning: cluster source words that map to one target word, then pick the senses that sit closest together in WordNet.
• I2. Merge outputs from two independent MT systems to get more varied translations, on the assumption that they rely on different dictionaries.
• I3. Pad a tiny test text with genre-matched corpora so a data-hungry statistical aligner works, then score only the original slice.
• I4. Score the projected labels only where the source label was already correct, which isolates projection error from upstream error.
• I5. Classify cross-lingual sense mismatches as equivalent, narrower or broader, and note which metonymy classes (animal/food, building/organization) carry across languages.

For Mova

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

• I1 (world): When anchoring lemmas to WordNet synsets in the ontology, translation clusters from parallel text or LLM translations could pick the mutually closest synsets and record that choice as an explainable reason.
• I2 (learn): This matches the witness-parser idea; several translators could act as independent witnesses whose agreement produces silver sense labels.
• I3 (learn): Small evaluation sets for statistical components such as aligners or clusterers could be embedded in larger same-genre data without contaminating the scored part.
• I4 (answer): Per-stage evaluation that conditions on correct upstream output would show whether a wrong answer comes from the tree, the frames, or the reader itself.
• I5 (back): When generating or translating from the meta-world, a sense with no direct lexical counterpart could be flagged and rendered as a compound or paraphrase, as Arabic does for the "tea party" sense.