Search

Transformer-based “behemoths” have grown in popularity, as well as structurally, shattering multiple NLP benchmarks along the way. However, their real-world usability remains a question. In this work, we empirically assess the feasibility of applying transformer-based models in real-world ad-hoc retrieval applications by comparison to a “greener and more sustainable” alternative, comprising only 620 trainable parameters. We present an analysis of their efficacy and efficiency and show that considering limited computational resources, the lighter model running on the CPU achieves a 3 to 20 times speedup in training and 7 to 47 times in inference while maintaining a comparable retrieval performance. Code to reproduce the efficiency experiments is available on “https://github.com/bioinformatics-ua/EACL2021-reproducibility/“.

Natural language processing and text analysis methods offer the potential of uncovering hidden associations from large amounts of unprocessed texts. The SemEval-2015 Analysis of Clinical Text task aimed at fostering research on the application of these methods in the clinical domain. The proposed task consisted of disorder identification with normalization to SNOMED-CT concepts, and disorder attribute identification, or template filling. We participated in both sub-tasks, using a combination of machine-learning and rules for recognizing and normalizing disease mentions, and rule-based methods for template filling. We achieved an F-score of 71.2% in the entity recognition and normalization task, and a slot weighted accuracy of 69.5% in the template filling task.

Clinical texts, such as discharge summaries or test reports, contain a valuable amount of information that, if efficiently and effectively mined, could be used to infer new knowledge, possibly leading to better diagnosis and therapeutics. With this in mind, the SemEval-2014 Analysis of Clinical Text task aimed at assessing and improving current methods for identification and normalization of concepts occurring in clinical narrative. This paper describes our approach in this task, which was based on a fully modular architecture for text mining. We followed a pure dictionary-based approach, after performing error analysis to refine our dictionaries. We obtained an F-measure of 69.4% in the entity recognition task, achieving the second best precision over all submitted runs (81.3%), with above average recall (60.5%). In the normalization task, we achieved a strict accuracy of 53.1% and a relaxed accuracy of 87.0%.