Understanding the relationship between two sentences through tasks including natural language inference, paraphrase detection, semantic textual similarity, and semantic relatedness is a fundamental step to natural language understanding. We propose an approach to infer the relationship between two sentences using a multitask framework to generate a universal representation of the relationship. Our model consists of a universal layer shared for all tasks with several task-specific layers on top for each task. To generate universal representation, we employ the enhanced sequential inference model based on a deep learning and soft alignment techniques. The task-specific layers are composed of multilayer perceptrons. The main feature of the proposed approach is that a single encoder can model various relationship of sentences at same time on multiple tasks. When we evaluated our approach on four public datasets for four different tasks regarding the relationship between two sentences, it outperformed state-of-the-art methods for two datasets and performed significantly well for the other two datasets. Further investigation of our proposed model showed that it captures comprehensive information together with specific knowledge regarding each task to infer semantic similarity. The detailed analysis supports that the proposed approach is robust over all semantic inference tasks using a single model.
HongSeok Choi
The extraction of interactions between chemicals and proteins from several biomedical articles is important in many fields of biomedical research such as drug development and prediction of drug side effects. Several natural language processing methods, including deep neural network (DNN) models, have been applied to address this problem. However, these methods were trained with hard-labeled data, which tend to become over-confident, leading to degradation of the model reliability. To estimate the data uncertainty and improve the reliability, “calibration” techniques have been applied to deep learning models. In this study, to extract chemical–protein interactions, we propose a DNN-based approach incorporating uncertainty information and calibration techniques. Our model first encodes the input sequence using a pre-trained languageunderstanding model, following which it is trained using two calibration methods: mixup training and addition of a confidence penalty loss. Finally, the model is re-trained with augmented data that are extracted using the estimated uncertainties. Our approach has achieved state-of-the-art performance with regard to the Biocreative VI ChemProt task, while preserving higher calibration abilities than those of previous approaches. Furthermore, our approach also presents the possibilities of using uncertainty estimation for performance improvement.
Dongha Choi
