A Post-Hoc Interpretability Study of Academic Literature Recommendations——Knowledge Graph Based on Joint Extraction of Entity Relationships
[Purpose/significance]Aiming at the current academic literature recommendation explainable research mostly focuses on model explainable,which not only has limited explaining strength but also involves certain performance sacrifices,a post hoc explain-able method based on knowledge graph is proposed.[Method/process]Based on deep learning,multi-task learning and attention mechanism,we construct a joint extraction model of entity relationship in medical literature with hybrid deep learning,and reveal the intrinsic semantic association between recommended literature and query topic through visual knowledge graph,so as to realize post hoc interpretability of recommendation results.[Result/conclusion]The Fl value on the manually labeled high-quality entity-relationship test dataset reaches 72.4％,which is 4.3％,5.2％,7.7％and 1.3％higher than the benchmark models BioBERT-BiLSTM-CRF and BioBERT-LSTM-CRF and LSTM-CRF and Joint-BiLSTM-RNN,respectively.The obtained knowledge graphs were mea-sured for interpretability in terms of seven metrics:transparency,trustworthiness,usefulness,effectiveness,persuasiveness,reviewabil-ity,and satisfaction,and received ratings of 4.3,3.6,4.1,3.8,3.7,1.0,and 2.7,respectively(with a rating interval of 0-5).[Innovation/limitation]The method of distant supervision is used to realize the automatic construction of large-scale training corpus,which can eliminate the tedious manual data labeling session,but the noise in the relation label reduces the performance of the model on relation extraction.The knowledge graph-based post hoc interpretable method can better explain the reasons for the recommendation results as well as assist users in making correct decisions,but there is still much room for improvement in terms of reviewability.
academic literature recommendationpost hoc interpretabilityknowledge graphjoint learningrelation extraction
NETL
NSTL
万方数据
