Research on a Chinese Text Information Density Evaluation Model Fusing Semantic and Statistical Features

Abstract

To address the issue in Chinese text information content evaluation where traditional methods primarily rely on statistical features and overlook semantic and structural complexity, this study proposes a Chinese text information density evaluation model that fuses semantic and statistical features. The model adopts a dual-channel fusion architecture: the semantic channel utilizes the pre-trained language model BERT to extract deep contextual embeddings of the text, combined with a Bidirectional Long Short-Term Memory network (BiLSTM) to capture long-range semantic dependencies; the statistical channel integrates Term Frequency-Inverse Document Frequency (TF-IDF), part-of-speech (POS) distribution, and dependency relations. These two types of heterogeneous features are concatenated and then fed into a fusion gating module for effective combination and non-linear interaction. Finally, a regression layer outputs a standardized information density score. For model training and evaluation, this study operationalized the definition of information density and constructed a manually annotated dataset comprising 210 diverse Chinese texts. Experimental results demonstrate that the proposed model significantly outperforms various baseline models across all evaluation metrics, validating the effectiveness and superiority of the fusion model for the task of Chinese text information density evaluation. This research provides a new analytical tool for applications such as text quality assessment and high-value information localization.