A Linking Persian Texts through Named Entity Recognition an‎d Topic Modeling Techniques

2/23/2027 12:00:00 AM

With the ever-increasing expansion of Persian content in sources such as Wikipedia, news agencies, an‎d social networks, the need for intelligent tools to analyze an‎d discover connections among these texts is felt more than ever. One suitable approach in this context is the combination of Named Entity Recognition methods with topic modeling. This research aims to present a novel an‎d integrated method for establishing links between Persian texts an‎d the automatic construction of a knowledge graph through entity extraction an‎d topic modeling. For this purpose, a complete version of the Persian Wikipedia corpus, comprising 1,011,252 documents, was used. In the first phase, named entities were extracted using the pre-trained Stanza model, which led to the identification of 19.5 million entity mentions an‎d ultimately 1.45 million unique entities. In the second phase, to discover the topical structure of the texts, four different models including LDA, NMF, CTM, an‎d BERTopic were compared. eva‎luations using the Coherence Score metric showed that the NMF model with TF-IDF weighting policy an‎d 240 topics, achieving a score of 0.7694, performed better than other models, including more advanced transformer-based models. Finally, by combining the results of the previous two phases an‎d using topical affinity as a criterion for inferring relationships, a massive an‎d native knowledge graph was constructed, containing 1,449,584 nodes (entities) an‎d 5,189,205 edges (semantic relations). Structural analysis of this graph revealed a very low density (near zero) an‎d an average degree of 5, indicating the successful elimination of ran‎dom relations an‎d the preservation of meaningful, topic-oriented connections. The results of this research are applicable in areas such as semantic search engines, recommender systems, an‎d Persian text content analysis

پردازش زبان طبيعي , شناسايي موجوديت نام‌گذاري شده , مدل‌سازي موضوعي , گراف دانش

Natural Language Processing , Named Entity Recognition , Topic Modeling , Knowledge Graph

Amir Radnia

Hassan Naderi