A Table Fact Verification Framework Based on Integrating Vocabulary and Deep Semantic Information

The table-based fact verification task problems inherently require informational reasoning about natural language as well as different numerical and logical reasoning about tables, Recent approaches have been to directly utilize natural language modeling methods for direct inference, which may not make good use of the contextual structure of the table, such as units, percentages, rankings, etc. If this contextual information is not taken into account, the implicit content may be overlooked, leading to missing information in the reasoning process. In this paper, we propose a new framework for table fact verification framework based on integrating vocabulary and deep semantic information. Specifically, we assign weights to different logical inference categories using Term Frequency-Inverse Document Frequency (tf-idf) for the benchmark dataset Tabfact, which contains weights for logical inference categories such as time, place, quantity, order and causality. To better identify logical types in the reasoning process, we employ the Mixture-of-Experts (MOE) framework. This framework is able to detect potential logic types in statements and utilizes multiple independent expert models to simulate different logical reasoning. Each of these expert models focuses on a specific reasoning task with a clear division of labor and is responsible for handling different types of logical reasoning. After fully utilizing the pre-trained model, the inputs are fed into different classifiers and then the outputs are weighted and summed according to the weights. Experiments show that our framework achieves 84.2% accuracy on the benchmark dataset Tabfact.