📄 Language-Specific versus Cross-Lingual Knowledge Graphs for Implicit Aspect Identification in Arabic: A Comparative Study of Reasoning and Adaptation Strategies

Summary

The paper asks how to find implicit aspects in Arabic reviews, meaning the evaluated property is implied by an opinion word but never named. The author compares two ways to build the knowledge graph that links opinion cues to Entity#Aspect categories. Strategy 1 reuses an existing English KG and reaches it through multilingual sentence embeddings. Strategy 2 builds a smaller native Arabic KG and matches words lexically after heavy normalization and morphological reduction.

Both strategies run inside one hybrid pipeline. That pipeline uses a QLoRA-tuned Llama-3-8B for explicit aspects, which abstains on sentences that have none. Evaluation covers M-ABSA, SemEval-2016 Arabic and HAAD. The native KG wins clearly on both precision and recall and roughly doubles micro-F1. A second comparison shows that zero-shot prompting of the 8B model is close to useless on Arabic, while task-specific fine-tuning makes the extractor workable.

The practical takeaway is that, for a morphologically rich and lower-resource language, native lexical resources and task adaptation pay off. The cheap cross-lingual shortcut does not.

Key points

• Implicit identification, micro-F1:

  • M-ABSA: 0.159 for the cross-lingual KG versus 0.358 for the Arabic KG.
  • SemEval-2016: 0.217 versus 0.468.
  • Strategy 1 recall is only about 0.10–0.14. Strategy 2 precision is 0.61–0.70, close to the English KG's precision on English data.

• Remaining recall gap: the Arabic KG detects no cue at all (STOPPED) on 52.1% of M-ABSA and 42.7% of SemEval-2016 implicit rows. A verb-phrase fallback for recommendation and affect expressions recovers 23.7% of M-ABSA implicit rows.

• Strategy 1 mechanics: it uses paraphrase-multilingual-MiniLM-L12-v2 and retrieves the top-5 English clue nodes.

  • Similarity thresholds are 0.75, or 0.85 for generic words.
  • The final score combines clue similarity (weight 0.6) with sentence-to-aspect similarity (weight 0.4), with a 0.65 cutoff.
  • Candidate adjectives come from a Stanza parse combined with CAMeL morphology.

• Explicit extraction with Llama-3-8B, micro-F1:

  Benchmark	Zero-shot	LoRA fine-tuned
  M-ABSA	0.048	0.663
  SemEval-2016	0.048	0.763
  HAAD	0.125	0.454

  Iterative mask-and-requery generation lifts M-ABSA to 0.766. Adding paraphrasing slightly hurts.

• Test data: M-ABSA has 1,356 explicit rows plus 718 implicit and 104 mixed. SemEval-2016 has 1,130 explicit rows and 96 implicit. HAAD has no implicit rows and is left out of the KG comparison. Significance is checked with paired bootstrap and McNemar's test.

Interesting ideas

• I1. The generative extractor is trained to output NULL when no aspect is stated on the surface, which hands those sentences to a separate symbolic reasoner; this cleanly splits explicit and implicit work.
• I2. A "lonely adjective" filter uses the dependency tree to find opinion adjectives that attach to no noun, treating them as signals of an unnamed target.
• I3. Lookup into the lexicon follows a fixed fallback order: exact form, then base form, then prefix, then a curated table, then substring. A part-of-speech guard blocks harmful suffix stripping.
• I4. Generic sentiment words get a stricter matching threshold and may only map to a GENERAL category, so vague words like "good" do not spray matches across specific aspects.
• I5. Results are reported as a breakdown of where inference stopped (no cue, inferred, cue found but no edge), not just as one F1. This shows whether the fix is better coverage or better graph edges.

For Mova

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

• I1 (world): a frame or entity extractor could emit an explicit "nothing stated" result that routes the sentence to knowledge-graph inference, instead of guessing a filler.
• I2 (frames): an adjective with no nominal head in the UD tree could trigger creation of an implicit evaluated entity or role, filled later from context.
• I3 (tokens): mapping lemmas onto WordNet/VerbNet entries could use the same ordered cascade, logging which tier fired so every match stays explainable.
• I4 (world): ontology linking could use a higher bar and a coarse fallback class for semantically bleached words, cutting spurious specific links.
• I5 (learn): evaluations could tag each failure by the pipeline stage where it stopped, pointing directly at the lexicon, the tree, or the knowledge graph as the place to repair.