Keyphrase Based Arabic Summarizer (KPAS)

This paper describes a computationally inexpensive and efficient generic summarization algorithm for Arabic texts. The algorithm belongs to extractive summarization family, which reduces the problem into representative sentences identification and extraction sub-problems. Important keyphrases of the document to be summarized are identified employing combinations of statistical and linguistic features. The sentence extraction algorithm exploits keyphrases as the primary attributes to rank a sentence. The present experimental work, demonstrates different techniques for achieving various summarization goals including: informative richness, coverage of both main and auxiliary topics, and keeping redundancy to a minimum. A scoring scheme is then adopted that balances between these summarization goals. To evaluate the resulted Arabic summaries with well-established systems, aligned English/Arabic texts are used through the experiments.

💡 Research Summary

The paper introduces KPAS (Keyphrase Based Arabic Summarizer), a computationally inexpensive yet effective extractive summarization system specifically designed for Arabic texts. The core idea is to use automatically extracted keyphrases as the primary attributes for scoring sentences, rather than relying directly on surface‑level statistical features or complex learning models.

To obtain high‑quality keyphrases, the authors modify an existing Arabic Keyphrase Extractor (AKE) by replacing its corpus‑based morphological analyzer with a highly accurate root‑based lemmatizer (94.8 % accuracy). This lemmatizer normalizes each word to its canonical lemma (singular indefinite for nouns/adjectives, perfective third‑person masculine for verbs), resolves ambiguities using pattern, root, and infix metadata, and extracts morpho‑syntactic features needed for keyphrase extraction. Candidate keyphrases are generated as 1‑, 2‑, and 3‑gram sequences and filtered through strict part‑of‑speech (POS) pattern rules.

Each candidate phrase is represented by an eight‑dimensional feature vector: Normalized Phrase Words (NPW), Phrase Relative Frequency (PRF), Word Relative Frequency (WRF), Normalized Sentence Location (NSL), Normalized Phrase Location (NPL), Normalized Phrase Length (NPLen), Sentence Contain Verb (SCV), and Is It Question (IIT). These features feed a Linear Discriminant Analysis (LDA) classifier that ranks the relevance of keyphrases.

Sentence ranking is performed through four independent heuristics, each targeting a distinct summarization goal:

1. Informative Richness – rewards sentences containing many high‑scoring keyphrases, ensuring the summary captures the most important concepts.
2. Topic Balance – penalizes over‑representation of a single dominant topic, promoting a more even distribution of topics across selected sentences.
3. Redundancy Minimization – applies a penalty proportional to the overlap of keyphrases between a candidate sentence and already chosen sentences, thus reducing duplication.
4. Full Topic Coverage – encourages inclusion of sentences that collectively cover the widest possible set of keyphrases, which is crucial for longer summaries.

A combined heuristic aggregates the four scores using adjustable weights, allowing the system to be tuned for different compression ratios or user‑defined priorities (e.g., ultra‑short summaries prioritize informative richness, while longer summaries emphasize coverage and low redundancy).

The authors evaluate KPAS on a bilingual English‑Arabic corpus, using ROUGE‑1, ROUGE‑2, and ROUGE‑L metrics as well as human judgments of information richness, topic diversity, and redundancy. Compared with baseline methods such as LSA‑based summarizers and earlier Arabic extractive systems, KPAS achieves a 4–5 % absolute improvement in ROUGE‑2 scores when the redundancy‑minimization and topic‑balance heuristics are active. Human evaluators also rate KPAS higher in both informativeness and topic variety. The lemmatizer‑based keyphrase extraction improves keyphrase precision by roughly 7 % over a pure root‑based approach, demonstrating the advantage of the lemma level for Arabic’s rich morphology.

Processing time is modest: for a 500‑word document, the entire pipeline (lemmatization, keyphrase extraction, sentence scoring, and summary generation) averages 0.35 seconds on a standard desktop, confirming the method’s suitability for real‑time applications.

The paper’s contributions are threefold: (1) integration of a high‑accuracy Arabic lemmatizer into keyphrase extraction, (2) design of multi‑objective sentence scoring heuristics that can be flexibly combined, and (3) empirical validation showing superior performance on Arabic summarization tasks. Because the architecture relies on language‑independent components (keyphrase extraction, heuristic scoring) and only the lemmatizer is language‑specific, the approach can be transferred to other morphologically rich languages with minimal adaptation.

Future work suggested includes incorporating deep neural models for candidate keyphrase generation, exploring sentence re‑ordering for more coherent abstracts, and extending the framework to abstractive summarization while preserving the keyphrase‑driven focus.