Abstract
Traditional document classification methods in scientific research often overlook the semantic order of words, resulting in accuracy challenges. Although deep learning has addressed this by considering word order, it faces issues with overfitting in complex tasks. This paper aims to enhance document classification accuracy by utilizing features from article keywords and abstracts. The objective is to improve feature representation through weighted keyword extraction and refined abstract processing, followed by training a hierarchical deep learning model for superior performance in complex text multi-classification. Proposed method involves several key steps: 1) Feature Representation: Extracting keywords from article titles and abstracts with enhanced representation using TF-IDF weighting to handle overlapping keywords. 2) Abstract Refinement: Employing POS tagging to refine lengthy abstracts into concise and informative segments. 3) Hierarchical Deep Learning Model: Combining TextCNN and BiLSTM architectures to capture fine-grained features and semantic contexts effectively. 4) Training Strategy: Training the hierarchical model to classify scientific documents based on refined features from keywords and abstracts. The innovative approach, FFDLText, which combines TextCNN and BiLSTM models, achieves higher accuracy and training efficiency compared to baseline models on the WOS dataset. This research contributes a novel method for complex text multi-classification tasks. The paper introduces FFDLText, a novel approach to scientific document classification using fine-grained feature extraction and hierarchical deep learning. By refining keyword representation and abstract content with TF-IDF weighting and POS tagging, and leveraging TextCNN and BiLSTM models, this method significantly enhances accuracy and efficiency in scientific research document classification.