Specialty Text Inference Based On Structured Knowledge Extraction

Specialty text contains specialty knowledge,usually involving some specialty terms,concepts,and logical relationships between them.Automated grading subjective text is a kind of specialty text inference task.The automated grading process mitigates the effects of subjective factors that may arise during manual review,thereby enhancing the efficiency and equity of the grading.Thus,it has important practical value.The specialty knowledge in the texts is an important basis for inference.Existing methods mostly rely on semantic modeling for inference,lacking of explicit knowledge explanations for inference results This paper proposes a specialized text inference method based on structured knowledge extraction,the main contributions of this paper are as follows:(1)We propose a method for constructing knowledge graph for specialty text.Specialty text consists of specialty elements,such as terms,entities,and important general vocabulary,and their corresponding relationships.The concept of knowledge graph is proposed to show them.Using the glossary of terms to obtain a collection of candidate specialty elements Based on the labels of the specialty texts in training set,calculate the information gain of the candidate specialty elements,and construct a specialty element table correspondingly.For a given specialty text,extract a set of specialty elements based on the specialty element table to form the node set.Then the relationships between specialty elements are established based on multiple features,such as context features,dependency syntax features,and semantic similarity features,to construct the edge set and corresponding weights.Constructing the knowledge graph allows extracting the structured knowledge from specialty texts and visually displaying their knowledge structure.As an important basis for subsequent specialty text inference,the knowledge graph provides a certain degree of interpretability to inference results.(2)We propose a specialty text inference model based on the structured knowledge.Knowledge features are the core features of specialty texts,and the knowledge graph is used to represent the knowledge features of specialty texts.Use pre-training auto-encoder to encode the knowledge graph.The auto-encoder includes encoder and decoder.Multi-layer convolutional neural network is chosen as the encoder and multi-layer perceptron is used as the decoder.A large pre-training language model is used to encode the original sequential text to construct the vector representation of the semantic features of the specialty texts,which encodes the complete semantics of the text.We introduce a consistency loss function to train the model.This function combines the semantic features of the text and promotes the effectiveness of specialty text inference based on knowledge features.By utilizing interpretable knowledge representation and the semantic information,the model provides an interpretable form of knowledge while better understanding and inferring the specialty text.(3)We evaluate the proposed model using datasets that include subjective answers from multiple questions in national professional qualification exams.Accuracy is chosen as the evaluation metric.The comparison experiments with multiple text inference methods indicates that the proposed model outperforms the other methods.Analyze the impact of different sizes of the training set on the model’s performance and find that the proposed model exhibits better robustness.To verify the effectiveness of each module in the proposed model,we performed ablation analysis.The construction of the knowledge graph is an important component of subsequent specialized text inference methods.The experiment verifies the impact of components such as specialized element table construction,specialized element extraction,and relation extraction on the model’s performance.The results indicate that constructing the knowledge graph properly can improve the performance of subsequent text inference tasks.We analyze samples with different scores,demonstrate the knowledge graph of them.Through comparison with reference answers,we evaluate the accuracy,completeness,and effectiveness of the knowledge graph,illustrating its usefulness in text inference tasks.