Electronic Theses and Dissertation

Pencocokan antara resume dan deskripsi pekerjaan merupakan tantangan penting dalam sistem pencarian kerja otomatis, terutama ketika menangani dokumen dengan teks panjang. Penelitian ini bertujuan mengembangkan sistem pencocokan otomatis dokumen panjang dengan menerapkan pendekatan dual-encoder berbasis model BigBird. Dataset penelitian terdiri dari 4.145 resume dan 2.245 deskripsi pekerjaan yang telah melalui tahap pembersihan dan pelabelan otomatis menggunakan library g4f (GPT-4o). Representasi dokumen dihasilkan melalui embedding BigBird tanpa chunking, kemudian tingkat kesamaan dihitung menggunakan cosine similarity. Untuk meningkatkan ketepatan rekomendasi, digunakan model reranker berbasis XGBoost Classifier dengan tiga fitur utama, yaitu embedding cosine similarity, label cosine similarity, dan keyword overlap. Proses pelatihan dilakukan menggunakan GroupKFold untuk mencegah kebocoran data antar resume. Evaluasi dilakukan melalui classification report dan metrik pemeringkatan top-k. Hasil menunjukkan bahwa model mencapai akurasi klasifikasi rata-rata sebesar 91% pada skema 5-fold, dengan performa terbaik pada top-5 ranking, menghasilkan f1-score sebesar 82.43%. Temuan ini membuktikan bahwa kombinasi arsitektur dual-encoder BigBird dan reranker XGBoost mampu menangkap konteks global dokumen secara efektif dan menghasilkan rekomendasi pekerjaan yang relevan.

Matching resumes and job descriptions is an important challenge in automated job search systems, especially when dealing with documents containing long texts. This study aims to develop an automated matching system for long documents by applying a dual-encoder approach based on the BigBird model. The research dataset consists of 4,145 resumes and 2,245 job descriptions that have undergone automatic cleaning and labeling using the g4f (GPT-4o) library. Document representations are generated through BigBird embedding without chunking, and similarity is calculated using cosine similarity. To improve recommendation accuracy, an XGBoost Classifier-based reranker model is used with three main features: cosine similarity embedding, cosine similarity label, and keyword overlap. Training is conducted using GroupKFold to prevent data leakage between resumes. Evaluation is performed through classification reports and top-k ranking metrics. The results show that the model achieved an average classification accuracy of 91% on the 5-fold scheme, with the best performance on the top-5 ranking, yielding an F1-score of 82.43%. These findings demonstrate that the combination of the BigBird dual-encoder architecture and the XGBoost reranker effectively captures the global context of documents and generates relevant job recommendations.