Query Expansion pada Retrieval Berbasis Corpus Content Base Retrieval (CBR) di Media Sosial untuk Meningkatkan Hasil Retrieval

Abstract

The rapid growth of social media data has generated a vast and diverse volume of data. Such data are typically unstructured, written in informal language, contain non-standard abbreviations, and exhibit high content dynamics. These characteristics pose significant challenges for Information Retrieval (IR) systems in producing relevant and accurate search results. This study focuses on improving IR performance in social media, specifically for e-government-related queries concerning Indonesia’s National Health Insurance (BPJS Kesehatan) collected from the Twitter (X) platform. Conventional IR models often struggle to handle unstructured content with informal language and abbreviations, leading to low retrieval accuracy. To address this issue, this research proposes a hybrid Query Expansion (QE) model called ROCBERT-QE, which integrates Corpus Content-Based Retrieval (CBR) with Bidirectional Encoder Representations from Transformers (BERT). The ROCBERT-QE model introduces a dual expansion mechanism in which corpus-based co-occurrence captures lexical relationships, while BERT embeddings preserve semantic meaning and contextual information. A domain-specific corpus comprising 5,017 preprocessed tweets related to Indonesia’s National Health Insurance (BPJS) was constructed, containing 6,215 unique terms that represent linguistic variations and informality within public discourse. Experimental results demonstrate that ROCBERT-QE outperforms baseline retrieval methods such as TF-IDF, BM25, and standard BERT. For single-word queries, the model achieved a Recall of 0.8574 and a Precision of 0.8807, while for sentence-based queries, Recall reached 0.8932 and Precision 0.9175. These improvements are attributed to the synergy between frequency-based expansion and deep contextual embeddings, which enable the model to effectively handle lexical noise and semantic ambiguity. The findings highlight the scientific potential of combining corpus-based and transformer-based approaches in IR tasks involving unstructured content. Practically, ROCBERT-QE can be applied to real-time analysis of public discourse in e-government contexts, such as service evaluation, policy feedback, and early detection of public issues. This framework is scalable and adaptable to other domains that feature informal or multilingual data characteristics.