Comparing deep learning methods for concept recognition in geo-analytic questions

Abstract

Named Entity Recognition (NER) is an important process of NLP systems for relation extraction, information retrieval and machine translation. Although various NER systems researched and improved for many decades, more accurate and advanced NER systems, which exploit deep learning techniques have emerged in the NLP domain, only the last few years. These newly emerged NER systems, due to the word embeddings and the non-linear transformations of data, lead to significantly improved performance. They are capable of tagging and classifying semantic entities such as person, location, organization, time, quantities, etc. more easily and accurately. For interpretation of geo-analytical questions, these NER systems should detect GIS-related semantics such as geographic phenomena, place names and temporal information. The last two pieces of information can be recognized by the current NER models, but none of them can identify and categorize geographic phenomena. To this end, this study presents two deep learning-based NER systems to extract geographic phenomena from geo-analytical questions and classify them into core concepts of spatial information that conceptually model and distinguish spatial information. The NER systems are trained by BERT and Bi-LSTM models on 278 geo-analytical questions and tested on 31 validation questions, from a corpus that contains 309 questions in total. The evaluation and comparison results showed that the BERT model had higher accuracy, precision, recall and F1-score on recognizing core concepts in geo-analytical questions, compared to Bi-LSTM.