Background
Semantic textual similarity (STS) measures the degree of relatedness between sentence pairs. The Open Health Natural Language Processing (OHNLP) Consortium released an expertly annotated STS data set and called for the National Natural Language Processing Clinical Challenges. This work describes our entry, an ensemble model that leverages a range of deep learning (DL) models. Our team from the National Library of Medicine obtained a Pearson correlation of 0.8967 in an official test set during 2019 National Natural Language Processing Clinical Challenges/Open Health Natural Language Processing shared task and achieved a second rank.
Objective
Although our models strongly correlate with manual annotations, annotator-level correlation was only moderate (weighted Cohen κ=0.60). We are cautious of the potential use of DL models in production systems and argue that it is more critical to evaluate the models in-depth, especially those with extremely high correlations. In this study, we benchmark the effectiveness and efficiency of top-ranked DL models. We quantify their robustness and inference times to validate their usefulness in real-time applications.
Methods
We benchmarked five DL models, which are the top-ranked systems for STS tasks: Convolutional Neural Network, BioSentVec, BioBERT, BlueBERT, and ClinicalBERT. We evaluated a random forest model as an additional baseline. For each model, we repeated the experiment 10 times, using the official training and testing sets. We reported 95% CI of the Wilcoxon rank-sum test on the average Pearson correlation (official evaluation metric) and running time. We further evaluated Spearman correlation, R², and mean squared error as additional measures.
Results
Using only the official training set, all models obtained highly effective results. BioSentVec and BioBERT achieved the highest average Pearson correlations (0.8497 and 0.8481, respectively). BioSentVec also had the highest results in 3 of 4 effectiveness measures, followed by BioBERT. However, their robustness to sentence pairs of different similarity levels varies significantly. A particular observation is that BERT models made the most errors (a mean squared error of over 2.5) on highly similar sentence pairs. They cannot capture highly similar sentence pairs effectively when they have different negation terms or word orders. In addition, time efficiency is dramatically different from the effectiveness results. On average, the BERT models were approximately 20 times and 50 times slower than the Convolutional Neural Network and BioSentVec models, respectively. This results in challenges for real-time applications.
Conclusions
Despite the excitement of further improving Pearson correlations in this data set, our results highlight that evaluations of the effectiveness and efficiency of STS models are critical. In future, we suggest more evaluations on the generalization capability and user-level testing of the models. We call for community efforts to create more biomedical and clinical STS data sets from different perspectives to reflect the multifaceted notion of sentence-relatedness.
Comparison of statistical indices for the evaluation of crop models performance
