Semantic Textual Similarity in Japanese Clinical Domain Texts Using BERT

Background Semantic textual similarity (STS) captures the degree of semantic similarity between texts. It plays an important role in many natural language processing applications such as text summarization, question answering, machine translation, information retrieval, dialog systems, plagiarism detection, and query ranking. STS has been widely studied in the general English domain. However, there exists few resources for STS tasks in the clinical domain and in languages other than English, such as Japanese. Objective The objective of this study is to capture semantic similarity between Japanese clinical texts (Japanese clinical STS) by creating a Japanese dataset that is publicly available. Materials We created two datasets for Japanese clinical STS: (1) Japanese case reports (CR dataset) and (2) Japanese electronic medical records (EMR dataset). The CR dataset was created from publicly available case reports extracted from the CiNii database. The EMR dataset was created from Japanese electronic medical records. Methods We used an approach based on bidirectional encoder representations from transformers (BERT) to capture the semantic similarity between the clinical domain texts. BERT is a popular approach for transfer learning and has been proven to be effective in achieving high accuracy for small datasets. We implemented two Japanese pretrained BERT models: a general Japanese BERT and a clinical Japanese BERT. The general Japanese BERT is pretrained on Japanese Wikipedia texts while the clinical Japanese BERT is pretrained on Japanese clinical texts. Results The BERT models performed well in capturing semantic similarity in our datasets. The general Japanese BERT outperformed the clinical Japanese BERT and achieved a high correlation with human score (0.904 in the CR dataset and 0.875 in the EMR dataset). It was unexpected that the general Japanese BERT outperformed the clinical Japanese BERT on clinical domain dataset. This could be due to the fact that the general Japanese BERT is pretrained on a wide range of texts compared with the clinical Japanese BERT.

Technological troubleshooting based on sentence embedding with deep transformers

In nowadays manufacturing, each technical assistance operation is digitally tracked. This results in a huge amount of textual data that can be exploited as a knowledge base to improve these operations. For instance, an ongoing problem can be addressed by retrieving potential solutions among the ones used to cope with similar problems during past operations. To be effective, most of the approaches for semantic textual similarity need to be supported by a structured semantic context (e.g. industry-specific ontology), resulting in high development and management costs. We overcome this limitation with a textual similarity approach featuring three functional modules. The data preparation module provides punctuation and stop-words removal, and word lemmatization. The pre-processed sentences undergo the sentence embedding module, based on Sentence-BERT (Bidirectional Encoder Representations from Transformers) and aimed at transforming the sentences into fixed-length vectors. Their cosine similarity is processed by the scoring module to match the expected similarity between the two original sentences. Finally, this similarity measure is employed to retrieve the most suitable recorded solutions for the ongoing problem. The effectiveness of the proposed approach is tested (i) against a state-of-the-art competitor and two well-known textual similarity approaches, and (ii) with two case studies, i.e. private company technical assistance reports and a benchmark dataset for semantic textual similarity. With respect to the state-of-the-art, the proposed approach results in comparable retrieval performance and significantly lower management cost: 30-min questionnaires are sufficient to obtain the semantic context knowledge to be injected into our textual search engine.

Predicting Semantic Similarity Between Clinical Sentence Pairs Using Transformer Models: Evaluation and Representational Analysis

Background Semantic textual similarity (STS) is a natural language processing (NLP) task that involves assigning a similarity score to 2 snippets of text based on their meaning. This task is particularly difficult in the domain of clinical text, which often features specialized language and the frequent use of abbreviations. Objective We created an NLP system to predict similarity scores for sentence pairs as part of the Clinical Semantic Textual Similarity track in the 2019 n2c2/OHNLP Shared Task on Challenges in Natural Language Processing for Clinical Data. We subsequently sought to analyze the intermediary token vectors extracted from our models while processing a pair of clinical sentences to identify where and how representations of semantic similarity are built in transformer models. Methods Given a clinical sentence pair, we take the average predicted similarity score across several independently fine-tuned transformers. In our model analysis we investigated the relationship between the final model’s loss and surface features of the sentence pairs and assessed the decodability and representational similarity of the token vectors generated by each model. Results Our model achieved a correlation of 0.87 with the ground-truth similarity score, reaching 6th place out of 33 teams (with a first-place score of 0.90). In detailed qualitative and quantitative analyses of the model’s loss, we identified the system’s failure to correctly model semantic similarity when both sentence pairs contain details of medical prescriptions, as well as its general tendency to overpredict semantic similarity given significant token overlap. The token vector analysis revealed divergent representational strategies for predicting textual similarity between bidirectional encoder representations from transformers (BERT)–style models and XLNet. We also found that a large amount information relevant to predicting STS can be captured using a combination of a classification token and the cosine distance between sentence-pair representations in the first layer of a transformer model that did not produce the best predictions on the test set. Conclusions We designed and trained a system that uses state-of-the-art NLP models to achieve very competitive results on a new clinical STS data set. As our approach uses no hand-crafted rules, it serves as a strong deep learning baseline for this task. Our key contribution is a detailed analysis of the model’s outputs and an investigation of the heuristic biases learned by transformer models. We suggest future improvements based on these findings. In our representational analysis we explore how different transformer models converge or diverge in their representation of semantic signals as the tokens of the sentences are augmented by successive layers. This analysis sheds light on how these “black box” models integrate semantic similarity information in intermediate layers, and points to new research directions in model distillation and sentence embedding extraction for applications in clinical NLP.

Mutual Implication as a Measure of Textual Equivalence

Semantic Textual Similarity (STS) and paraphrase de- tection are two NLP tasks that have a high focus on the meaning of sentences, and current research in both re- lies heavily on comparing fragments of text. Little to no work has been done in studying inference-centric ap- proaches to solve these tasks. We study the relation be- tween existing work and what we call mutual implica- tion (MI), a binary relationship between two sentences that holds when they textually entail each other. MI thus shifts the focus of STS and paraphrase detection to un- derstanding the meaning of a sentence in terms of its in- ferential properties. We study the comparison between MI, paraphrasing, and STS work. We then argue that MI should be considered a complementary evaluation met- ric for advancing work in areas as diverse as machine translation, natural language inference, etc. Finally, we study the limitations of MI and discuss possibilities for overcoming them.