SSMFRP: Semantic Similarity Model for Relation Prediction in KBQA Based on Pre-trained Models

Pre-trained Relation Extraction (RE) models are widely employed in relation prediction in Knowledge Base Question Answering (KBQA). However, pre-trained models are usually optimized and evaluated on datasets (e.g. GLUE) which contain various Natural Language Processing (NLP) tasks except a RE task. As a result, it is difficult to select a best pre-trained model for relation prediction unless we evaluate all available pre-trained models on a relation prediction dataset. As the Semantic Similarity (SS) task in GLUE is similar to a RE task, a Semantic Similarity Model for Relation Prediction (SSMFRP) is proposed in this paper to convert a RE task in relation prediction to a SS task. In our model, a relation candidate in a RE model is converted into the corresponding question which contains a relation candidate. Then a modified SS model is employed to find the best-matched relation. Experimental results show that the effectiveness of our proposed model in relation prediction is related to the effectiveness of the original pre-trained model in SS and GLUE. Our model achieves an average accuracy of 91.2% with various pre-trained models and outperforms original models by an average margin of 1.8% with the similar training cost. In addition, further experiments show that our model is robust to abnormal input and outperforms original models by an average margin of 1.0% on datasets of abnormal input.