📄 Arabic Keyphrase Extraction using Linguistic knowledge and Machine Learning Techniques

Summary

The paper tackles automatic keyphrase extraction for Arabic. Earlier Arabic extractors relied mostly on statistics such as term frequency and position, and they performed poorly. The authors build a supervised extractor that adds linguistic knowledge at three stages: preprocessing, candidate generation and feature computation. Words are looked up in a specially built annotated Arabic corpus, which supplies fine-grained word classes and an "abstract form" of each word. This abstract form hides most inflection better than a plain stem does. Candidate phrases are n-grams filtered by POS-based sequence rules. Each candidate gets statistical and linguistic features, including two new ones meant to catch titles and subtitles. The authors check the features with ANOVA and train the classifier with Linear Discriminant Analysis (LDA) on IT-domain documents. Compared with KP-Miner and the Sakhr extractor, the system is slightly better on IT texts. On longer non-scientific texts its precision and recall are roughly double KP-Miner's. This suggests that linguistic normalization makes frequency statistics much more reliable in a morphologically rich language.

Key points

• Pipeline:
  • split the text into sentences and words;
  • do corpus-based lexical analysis to get word class, gender, number, person and abstract form;
  • generate rule-filtered n-gram candidates;
  • compute a feature vector for each candidate;
  • classify with LDA, after an ANOVA check of the features.
• The features the paper names are NPL (Normalized Phrase Location), NSL (Normalized Sentence Location), NPLen (Normalized Phrase Length) and SCV (Sentence Contain Verb). NPLen and SCV are aimed at headings in long documents.
• Experiment 1 used 12 documents averaging 203 words, with 12 keyphrases extracted per document. Human judges accepted about 44.16% of them (std 6%).
• Experiment 2 used 20 IT documents averaging 376.1 words. At five keyphrases the scores (P/R) were:
  • this system: 0.53/0.36
  • KP-Miner: 0.48/0.30
  • Sakhr: 0.20/0.14
• Experiment 3 used 20 non-scientific newspaper and web documents averaging 675.2 words. At five keyphrases the scores (P/R) were:
  • this system: 0.65/0.40
  • KP-Miner: 0.34/0.20
  • Sakhr: 0.19/0.12

Interesting ideas

• I1. Collapse inflected variants into one normalized form before counting, so that frequency features measure concepts instead of surface strings; the paper credits this for the biggest gains on inflection-heavy texts.
• I2. Refine coarse word classes using affixes, for example marking a noun with a prepositional prefix as a distinct declined class, so that downstream rules see context-sensitive categories.
• I3. Let only linguistically plausible candidates reach the learner, using POS-sequence rules, so that the classifier is not swamped by junk n-grams and its output stays humanly sensible.
• I4. Detect titles and subtitles from structure: short segments without a verb tend to be headings, and the phrases inside them are strong keyphrase signals.
• I5. Test each feature's discriminative power with a simple ANOVA before training, and evaluate on a domain the model was not trained on (IT → social news).

For Mova

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

• I1 (world): When ranking which entities a text is "about", pool mentions by lemma plus morphological features (and coreference) before computing salience, rather than counting raw tokens.
• I2 (tokens): The English tagger could expose finer subclasses (proper, temporal, locative, count nouns) from morphology and lexicon cues, giving later stages richer, explainable categories than bare UPOS.
• I3 (tree): Generate key-concept candidates as UD noun-phrase subtrees (head plus amod/compound/nmod) instead of n-gram windows, keeping every candidate traceable to a parse.
• I4 (world): A verbless short root in the UD tree could mark a heading, and entities in it could get a topic boost when building the world of a long text.
• I5 (learn): Add per-feature significance checks and an out-of-domain test split to Mova's evaluation, alongside negative controls, to catch features that only fit the training genre.