SciNER: Extracting Named Entities from Scientific Literature

Abstract

The automated extraction of claims from scientific papers via computer is difficult due to the ambiguity and variability inherent in natural language. Even apparently simple tasks, such as isolating reported values for physical quantities (e.g., "the melting point of X is Y") can be complicated by such factors as domain-specific conventions about how named entities (the X in the example) are referenced. Although there are domain-specific toolkits that can handle such complications in certain areas, a generalizable, adaptable model for scientific texts is still lacking. As a first step towards automating this process, we present a generalizable neural network model, SciNER, for recognizing scientific entities in free text. Based on bidirectional LSTM networks, our model combines word embeddings, subword embeddings, and external knowledge (from DBpedia) to boost its accuracy. Experiments show that our model outperforms a leading domain-specific extraction toolkit by up to 50%, as measured by F1 score, while also being easily adapted to new domains.

Keyword extraction and summarization from unstructured text: A case study with open data from legal domain

“…In another study, using a domain-specific ontology for the Hotels domain, the authors have shown accurate extraction of named entities [2]. Hong et al [5] discussed the implementation of scientific named entity extraction. This is solely for entities that are scientific names in a given text; this can be further improved to a generalized entity extraction which does not require using a restricted entity.…”

Models and Processes to Extract Drug-like Molecules From Natural Language Text

“…While SpaCy is easy to use, it lacks flexibility: its end-to-end encapsulation does not expose many tunable parameters. Thus we also explore the use of a Keras-LSTM model that we developed in previous work for identification of polymers in materials science literature (Hong et al, 2020b). This model is based on the Bidirectional LSTM network with a conditional random field (CRF) layer added on top.…”

“…The Keras LSTM model requires external word vectors since, unlike SpaCy, it does not include a word embedding model. To explore the affect of different word embedding models we trained both BERT (Devlin et al, 2018), a top-performing language model developed by Google, and FastText (Bojanowski et al, 2016), a model shown to have outperformed traditional Word2Vec models such as CBOW and Skipgram in our previous work (Hong et al, 2020b). While Google has released pre-trained BERT models, and researchers often build upon these models by "fine-tuning" them with additional training on small external datasets, it is not suitable to our problem as the vocabulary used in the CORD-19 is very different than the datasets used to train these models.…”

“…Having collected adequate training data via this model-guided human annotation process, we then use the resulting labeled data to re-train a NER model originally developed to identify polymer names in materials science publications (Hong et al, 2020b) and apply this trained model to We show that the labeled data produced by our approach are of sufficiently high quality than when used to train NER models, which achieves a best F-1 score of 80.5%-roughly equivalent to that achieved by nonexpert humans.…”

Leveraging Theory for Enhanced Machine Learning

“…The lack of large data sets is currently being tackled by several efforts to compile data, ,,,, including through natural language processing − and the generation of large computational data sets. − It is also being tackled by transfer learning, a growing technique in ML where knowledge is transferred between tasks (e.g., prediction of properties), domains (e.g., scientific literature or English literature), or both, as detailed in an excellent review . Most commonly knowledge is transferred from a task or domain where data is plentiful to a task or domain where data is limited.…”