CHEMDNER system with mixed conditional random fields and multi-scale word clustering

Abstract

BackgroundThe chemical compound and drug name recognition plays an important role in chemical text mining, and it is the basis for automatic relation extraction and event identification in chemical information processing. So a high-performance named entity recognition system for chemical compound and drug names is necessary.MethodsWe developed a CHEMDNER system based on mixed conditional random fields (CRF) with word clustering for chemical compound and drug name recognition. For the word clustering, we used Brown's hierarchical algorithm and Skip-gram model based on deep learning with massive PubMed articles including titles and abstracts.ResultsThis system achieved the highest F-score of 88.20% for the CDI task and the second highest F-score of 87.11% for the CEM task in BioCreative IV. The performance was further improved by multi-scale clustering based on deep learning, achieving the F-score of 88.71% for CDI and 88.06% for CEM.ConclusionsThe mixed CRF model represents both the internal complexity and external contexts of the entities, and the model is integrated with word clustering to capture domain knowledge with PubMed articles including titles and abstracts. The domain knowledge helps to ensure the performance of the entity recognition, even without fine-grained linguistic features and manually designed rules.

Putting hands to rest: efficient deep CNN-RNN architecture for chemical named entity recognition with no hand-crafted rules

“…Due to many difficulties inherent to chemical entity detection and normalisation [1], even manual annotation yields the interannotator agreement score of 91%, which can be regarded as the theoretical limit for any automatic system trained on this corpus. Twenty six teams have submitted their NER systems for the challenge, best of which have reached the F1 score of ∼ 72 − 88% [2,3,4,5,6,7,8,9] on two subtasks: chemical entity mention (CEM) and chemical document indexing (CDI).…”

“…Although tokenisation typically reduces the number of time-steps in the sequence, thus reducing the input complexity, it can introduce severe artefacts, e.g. merged/overlapping entities [5,9]. It makes it essential to use an adequate tokeniser with rules finely adjusted for the task at hand.…”

“…, t T ) of orthogonal unit vectors (also known as one-hot encodings), both sparse and purely categorical: their pair-wise distances carry no underlying information about semantical similarity. In their chemical NER system, Lu et al [9] successfully used the skip-gram embedding model to overcome these limitations. The model uses context information and a shallow neural network to embed high-dimensional one-hot encoded vectors in a lower-dimensional vector space, wherein pair-wise distances represent semantical similarity [10,11].…”

“…LeadMine (179) [8] is exclusively commercial. While there is a GitHub repository for the model devised by Lu et al (team 231) [9,30], it literally contains nothing but a link to an archive (which supposedly contains the model), uploaded to a file-hosting service that requires a proprietary application to download the archive. Since both the file-hosting and the application are only available in Chinese, we have been unable to download the archive and thus consider it inaccessible.…”

“…Apart from the models submitted for the CHEMDNER challenge we have also considered ChemDataExtractor [22], a recently introduced general purpose package for chemical text analyses, because its NER model is very much akin to [9], which is unavailable. Both utilise unsupervised word-clustering and CRFs, though ChemDataExtractor uses a hierarchical detection system with a built-in database updated in an online fashion to help it extract abbreviations and identifiers.…”