📄 Arabic Metaphor Sentiment Classification Using Semantic Information

Summary

The paper tackles sentiment classification for Arabic text that contains metaphors. No dedicated tools exist for this task, and building annotated data for it is expensive. The author works with the Arabic Metaphor Corpus (AMC), 1,000 Arabic book reviews annotated by hand for metaphor terms, review-level sentiment and metaphor-level sentiment. The reviews are tagged with the Arabic Semantic Tagger (AraSAS).

Two rule-based Python tools are built on top of the emotion (E) tags:

• First tool: sums graded polarity weights taken from the + and − marks on the tags.
• Second tool: also adds or subtracts a fixed bonus according to the gold metaphor sentiment. It does not detect metaphors itself.

Both tools are scored against the gold standard with precision, recall and F-score. Results are broken down by sentiment class and by review length.

• The E-tag tool reaches an overall F-score of 0.526. It does far better on positive reviews than on negative or neutral ones.
• The metaphor-aware variant drops to 0.351 overall when compared with review-level gold labels.
• A later comparison table reports best per-category scores of up to about 0.96 for very short reviews.

The author presents the work as a first, annotation-light baseline for Arabic metaphor sentiment.

Key points

• AMC holds 1,000 book reviews: 702 positive, 171 negative and 127 neutral. Each has overall sentiment, metaphor-specific sentiment and AraSAS semantic tags, and the corpus is distributed as Excel.
• Practical workarounds were needed for AraSAS:
  • It has no built-in overall sentiment output.
  • It can take only 100 reviews per batch.
  • It treats full stops and exclamation marks as line breaks, so these were replaced with English letters before tagging.
• The scoring scheme works as follows:
  • E+ adds 0.5, E++ adds 1 and E+++ adds 1.5, with the mirror values for negative marks.
  • Neutral E-tags are ignored.
  • The final polarity is the sign of the total.
  • The metaphor tool adds ±2 for gold metaphor polarity.
• Table 1 (semantic tags only) shows:
  • an overall F-score of 0.526;
  • 0.640 for positive, 0.370 for negative and 0.107 for neutral reviews;
  • a best score of 0.661 on reviews of 1–4 tokens.
• Table 2 (with metaphor sentiment) shows:
  • an overall F-score of 0.351;
  • a best score of 0.526 for 500–999 tokens and a worst of 0.099 for 5–9 tokens.
  • The author blames the drop on comparing metaphor-level signals with review-level gold.
• The final comparison table gives its lowest best score, 0.33, for reviews of 1,000 tokens or more.

Interesting ideas

• I1. The tagger's intensity marks (+, ++, +++) are turned into additive, linearly graded weights, so each polarity verdict can be traced back to the individual tags that produced it. — For Mova (frames): Speech-act and irony analysis in prag could give evaluative lexical cues graded weights, making every stance judgement auditable down to specific words.
• I2. Scores are reported per review-length bucket rather than as one number, which shows that a simple tagger-based method copes well with short texts and fails on very long ones. — For Mova (learn): Evaluation could routinely split results by sentence or document length, exposing where parsing or frame extraction breaks down instead of hiding it in an average.
• I3. Metaphor sentiment is injected from gold annotation rather than detected, which isolates how much that one signal can contribute before any detector exists. — For Mova (learn): Oracle-component runs could be used, for example feeding gold coreference or gold UD trees, to measure each stage's ceiling before investing in improving it.
• I4. The low metaphor-tool scores are explained by evaluating a span-level signal against a document-level label, a reminder to match the evaluation unit to the unit being predicted. — For Mova (answer): When checking readers on fables and tests, the gold and the prediction should be compared at the same granularity, such as sentence-level frames against sentence-level gold.
• I5. Punctuation is pre-masked so a black-box tagger does not split text wrongly, and the restored text is used downstream. — For Mova (tokens): When external tools or witness parsers are wrapped, characters that confuse their segmentation could be protected reversibly, so token alignment back to the original text stays exact.