Deep learning with word embeddings improves biomedical named entity recognition.

MOTIVATION: Text mining has become an important tool for biomedical research. The most fundamental text-mining task is the recognition of biomedical named entities (NER), such as genes, chemicals and diseases. Current NER methods rely on pre-defined features which try to capture the specific surface properties of entity types, properties of the typical local context, background knowledge, and linguistic information. State-of-the-art tools are entity-specific, as dictionaries and empirically optimal feature sets differ between entity types, which makes their development costly. Furthermore, features are often optimized for a specific gold standard corpus, which makes extrapolation of quality measures difficult.
RESULTS: We show that a completely generic method based on deep learning and statistical word embeddings [called long short-term memory network-conditional random field (LSTM-CRF)] outperforms state-of-the-art entity-specific NER tools, and often by a large margin. To this end, we compared the performance of LSTM-CRF on 33 data sets covering five different entity classes with that of best-of-class NER tools and an entity-agnostic CRF implementation. On average, F1-score of LSTM-CRF is 5% above that of the baselines, mostly due to a sharp increase in recall.
AVAILABILITY AND IMPLEMENTATION: The source code for LSTM-CRF is available at https://github.com/glample/tagger and the links to the corpora are available at https://corposaurus.github.io/corpora/ . CONTACT: habibima@informatik.hu-berlin.de.

1 Introduction

Text mining is an important tool for many types of large-scale biomedical data analysis, such as network biology (Zhou ), gene prioritization (Aerts ), drug repositioning (Wang and Zhang, 2013) or creation of curated databases (Li ). The most fundamental task in biomedical text mining is the recognition of named entities (called NER), such as proteins, species, diseases, chemicals or mutations. To date, the best performing NER tools rely on specific features to capture the characteristics of the different entity classes. For instance, the suffix ‘-ase’ is more frequent in protein names than in diseases; species names often consist of two tokens and have latin suffixes; chemicals often contain specific syllabi like ‘methyl’ or ‘carboxyl’, and mutations usually are sequences of letters and digits encoding genomic position, type of mutation, and base changes. Feature engineering, i.e. finding the set of features that best helps to discern entities of a specific type from other tokens (or other entity classes) currently is more of an art than a science, incurring extensive trial-and-error experiments. On top of this costly process, high-quality NER tools today need further entity-specific modules, such as whitelist and blacklist dictionaries, which again are difficult to build and maintain. Defining these steps currently takes the majority of time and cost when developing NER tools (Leser and Hakenberg, 2005) and leads to highly specialized solutions that cannot be used for other entity types than the ones they were designed for. On the other hand, the method used to identify entities in a given text based on the defined features nowadays is fairly homogeneous: conditional random fields (CRFs) (Lafferty ), a statistical sequential classification method, is the de-facto standard method. Here, we show that an entirely generic NER method based on deep learning and distributional word semantics outperforms such specific high-quality NER methods across different entity types and across different evaluation corpora.

In the general field of information extraction, two recent developments lead to substantial improvements. First, word embeddings have been introduced to represent a single word by a low-dimensional vector capturing—in some way—the frequencies of co-occurring adjacent words. When compared with the bag-of-words approach underlying the conventional methods outlined earlier, word embeddings capture semantic similarities between words (as mathematical similarities between their vectors) that are not visible from their surface; for instance, the words ‘enables’ and ‘allows’ are syntactically very different, yet their meaning is somewhat related, which leads to similar sets of co-occurring words, whereas the co-occurrences of the word ‘swim’ would be completely different. The underlying idea of representing words ‘by the company they keep’ (Mackin, 1978) is an old concept in linguistics, usually called distributional semantics; its recent popularity is based on the novel idea that the embeddings are automatically adjusted such that information extraction tools benefit the most. Second, it has been shown that the application of artificial neural networks (ANNs), which automatically learn non-linear combinations of features, leads to better recognition results than the usage of CRFs, which can only learn (log-)linear combinations of features. Deep neural networks, and especially long short-term memory networks (LSTM), perform this task particularly efficiently and effectively. As with word embeddings, this idea is not new (Hochreiter and Schmidhuber, 1997), but only recent progress in the available data volumes and machine capabilities made it applicable to practically relevant problems (Pascanu ).

Following a suggestion from Lample , we combined the power of word embeddings, LSTMs and CRFs into a single method for biomedical NER, called LSTM-CRF. This method is completely agnostic to the type of the entity; all it requires is an entity-annotated gold standard and word embeddings pre-computed on a large, entity-independent corpus (typically all PubMed abstracts). We assessed the performance of LSTM-CRF by performing 33 evaluations on 24 different gold standard corpora (some with annotations for more than one entity type) covering five different entity types, namely chemical names, disease names, species names, genes/protein names, and names of cell lines. These corpora encompass patents and scientific articles and partly consist of abstracts and partly of full texts. We compared the performance of LSTM-CRF with that of best-of-class, entity-specific NER tools and with another generic NER method using a CRF with a typical NER feature set plus the word embeddings as input. LSTM-CRF turned out to have the best F1-score on 28 of the 33 cases; on average, it is 5% better than the entity-specific NER tools and 3% better than the CRF method with word embeddings.

2 Materials and methods

In the following sections, we give a technical explication of the LSTM-CRF approach and describe the competitor NER systems. Furthermore, we describe the corpora we used for evaluation, the different embeddings evaluated, and details regarding text pre-processing and evaluation metrics.

2.1 LSTM-CRF

LSTM-CRF (Lample ) is a domain-independent NER method which does not rely on any kind of background knowledge. We first describe LSTM, a specific kind of ANN, and then discuss the architecture of LSTM-CRF in detail.

An LSTM is a special kind of ANN which processes sequences of arbitrary length and is able to model dependencies between sequence elements even if they are far apart (Hochreiter and Schmidhuber, 1997). The input to an LSTM unit is a sequence of vectors of length T, for which it produces an output sequence of vectors of equal length by applying a non-linear transformation learned during the training phase. Each is called the activation of the LSTM at token t. The exact formula to compute one activation of an LSTM unit in the LSTM-CRF model is provided below (Lample ): where all s and s are trainable parameters, σ denotes the element-wise sigmoid function and is the element-wise product.

Such an LSTM-layer processes the input only in one direction and thus can only encode dependencies on elements that came earlier in the sequence. As remedy for this problem, we use another LSTM-layer processing in the reversed direction, which allows detecting dependencies on elements later in the text. The resulting neural network is called a bi-directional LSTM (Graves and Schmidhuber, 2005).

The architecture of LSTM-CRF is shown in Figure 1. It is comprised of three main layers. The first layer is the embedding layer. It receives the raw sentence S made of the sequence of words as its input and produces an embedding (i.e. a dense vector representation) for each word in S. The embedding vector of word w is a concatenation of a word- and a character-level embedding. The word-level embedding is simply retrieved from a lookup-table (see the example in Fig. 1) of word embeddings trained on a large corpus (see Section 2.3). The character-level embedding is obtained by applying a bi-directional LSTM to the character sequence of each word and then concatenating the last activations of both directions, as exemplified for the word ‘SH3’ in the left side of Figure 1. The resulting sequence of embeddings is fed into another bi-directional LSTM-layer that produces a refined representation of the input, which is the input to a final CRF-layer. The classical Viterbi algorithm is used to obtain the final output from this layer. All components together form a single fully differentiable neural network that can be trained by backpropagation.

Fig. 1

CRF-LSTM architecture. For instance, for the word ‘SH3’ from the input sentence S, the character-based representation is computed by applying a bi-directional LSTM onto the sequence of its characters ‘S’, ‘H’, ‘3’. The resulting embedding is concatenated with the corresponding word embedding, trained on a huge corpus. This word representation is then processed by another bi-directional LSTM and finally by a CRF layer. The output is the most probable tag sequence, as estimated by the CRF

In our experiments, we used a learning rate of 0.005 for all corpora except for BioSemantics where we set it to 0.0005, because the model did not converge with the default one. Regarding the other hyperparameters, we used the default values from (Lample ) except for (i) the tag scheme which we set to IOB instead of IOBES, and (ii) the dropout rate which we set to 0.3 instead of 0.5 because this value was optimal for most corpora evaluated by Lample .

2.2 Competitor systems

We compare the performance of LSTM-CRF against two types of competitors: a CRF using a generic feature set for NER tasks plus word embeddings, and entity-specific NER tools for each class. The former should help to separate the impact of using word embeddings from the impact of using the LSTM-CRF architecture. For the latter we selected the current best-in-class entity-specific NER tools. All trainable baseline systems were retrained using the same corpora and folds.

2.2.1 Baseline CRF

We used CRFSuite (http://www.chokkan.org/software/crfsuite/) (Okazaki, 2007) with default settings to train a first-order linear-chain CRF model utilizing identical features for all entity types. These features were defined by the NER feature extractor shipped with CRFsuite and were designed for domain-independent NER. Additionally, we provided as features (in turn) one of the word embeddings described in Section 2.4.

2.2.2 Baseline methods

For each of the five entity classes considered in our evaluation, we chose the presumably best performing and publicly available current NER tool:

Chemicals: we employed tmChem (https://www.ncbi.nlm.nih.gov/CBBresearch/Lu/Demo/tmTools/) (Leaman ) Model I, considered as the state-of-the-art chemical NER tool also in other recent studies (Habibi ; Leaman and Lu, 2016;). tmChem Model I trains a first-order CRF using features from BANNER (Leaman and Gonzalez, 2008) plus further ones obtained through a trial-and-error procedure, including character n-grams, chemical specific identifiers, and the output of ChemSpot (Rocktäschel ) (another high-quality chemical NER tool). The output of the model is filtered by several type-specific post-processing steps for abbreviation resolution, enforcing of tagging consistency and balancing of parentheses.

Genes/proteins: we utilized Gimli (http://bioinformatics.ua.pt/software/gimli/) (Campos ), an open source retrainable gene/protein NER tool with competitive performance (Campos ). Gimli trains both first- and second-order CRF models using a set of specific features, such as orthographic, morphological, linguistic-based, dictionary-based and a conjunction of features from adjacent tokens. Post-processing steps like parentheses correction, abbreviation resolution, and name extension using a domain dictionary are also applied.

Species: we used SPECIES (http://species.jensenlab.org/) (Pafilis ), a freely available dictionary-based tool with state-of-the-art tagging quality. Species is the only NER tool in our set which does not build on a CRF, and it is also the only tool which does not train a corpus-specific model.

Diseases: we used DNorm (https://www.ncbi.nlm.nih.gov/CBBresearch/Lu/Demo/tmTools/) (Leaman ), a freely available tool showing excellent performance in several studies (Leaman and Lu, 2016; Wei ). It uses the rich feature set provided by BANNER and a dictionary of disease names created by the authors to train a first-order CRF model. The output of the model is also filtered by manual rules like abbreviation resolution.

Cell Lines: we used a model presented in (Kaewphan ), reported to outperform many other available systems in a cross-corpus setting. It is based on the domain-independent CRF-based tool NERsuite (http://nersuite.nlplab.org/). In addition to the pre-defined NERsuite features, it uses a comprehensive dictionary of cell line names from multiple sources.

2.3 Gold standard corpora

We performed our evaluations on five entity types: genes/proteins, chemicals, diseases, cell lines and species. We relied on a total of 24 corpora, each containing manual annotations for one or more of these entity types, such as CHEMDNER patent (Krallinger ,b) for chemicals and genes/proteins, NCBI Disease (Doğan ) for disease names, Gellus (Kaewphan ) for cell lines and S800 (Pafilis ) for species. The corpora encompass two different genres of texts: (i) patent texts from the European Patent Office (EPO) (http://www.epo.org/), World Intellectual Property Organization (WIPO) (http://www.wipo.int/), and United States Patent and Trademark Office (USPTO) (http://www.uspto.gov/) and (ii) scientific articles from PubMed Central (PMC) (http://www.ncbi.nlm.nih.gov/pmc) and PubMed (https://www.ncbi.nlm.nih.gov/pubmed/).Table 1 lists all corpora together with important characteristics like the number of sentences, tokens, and annotated entities per entity class (measured after text pre-processing as described in Section 2.5).

Table 1.

The details of the gold standard corpora

Corpora	Text genre	Text type	Entity type	No. sentences	No. tokens	No. unique tokens	No. annotations	No. unique annotations
CHEMDNER patent (Krallinger et al., 2015a,b)	Patent	Abstract	Chemicals	35 584	1 465 471	62 344	64 026	20 893
			Genes/proteins	35 584	1 465 471	62 344	12 597	5835
CHEBI (http://chebi.cvs.sourceforge.net/viewvc/chebi/chapati/)	Patent	Full-text	Chemicals	10 638	314 105	24 424	17 724	5109
BioSemantics (Akhondi et al., 2014)	Patent	Full-text	Chemicals	173 808	5 690 518	208 326	368 091	72 756
CHEMDNER (Krallinger et al., 2015a)	Scientific article	Abstract	Chemicals	89 679	2 235 435	114 837	79 842	24 321
CDR (Li et al., 2016)	Scientific article	Abstract	Chemicals	14 228	323 281	23 068	15 411	3629
			Diseases	14 228	323 281	23 068	12 694	3459
BioCreative II GM (Smith et al., 2008)	Scientific article	Abstract	Genes/proteins	20 510	508 257	50 864	20 703	14 906
JNLPBA (Kim et al., 2004)	Scientific article	Abstract	Genes/proteins	22 562	597 333	25 046	35 460	10 732
			Cell Lines	22 562	597 333	25 046	4332	2528
CellFinder (Neves et al., 2012)	Scientific article	Full-text	Genes/proteins	2177	65 031	7977	1340	600
			Species	2177	65 031	7977	435	51
			Cell Lines	2177	65 031	7977	350	69
OSIRIS (Furlong et al., 2008)	Scientific article	Abstract	Genes/proteins	1043	28 697	4669	768	275
DECA (Wang et al., 2010)	Scientific article	Abstract	Genes/proteins	5468	138 034	14 515	6048	2360
Variome (Verspoor et al., 2013)	Scientific article	Full-text	Genes/proteins	6471	172 409	12 659	4309	596
			Diseases	6471	172 409	12 659	6025	629
			Species	6471	172 409	12 659	182	8
PennBioIE (Kulick et al., 2004)	Scientific article	Abstract	Genes/proteins	14 305	357 313	21 089	17 427	4348
FSU-PRGE (Hahn et al., 2010)	Scientific article	Abstract	Genes/proteins	35 465	899 426	57 498	56 087	19 470
IEPA (Ding et al., 2002)	Scientific article	Abstract	Genes/proteins	243	15 174	2923	1109	209
BioInfer (Pyysalo et al., 2007)	Scientific article	Abstract	Genes/proteins	1141	33 832	5301	4162	1150
miRNA (Bagewadi et al., 2014)	Scientific article	Abstract	Genes/proteins	2704	65 998	7821	1006	409
			Diseases	2704	65 998	7821	2123	671
			Species	2704	65 998	7821	726	47
NCBI Disease (Doğan et al., 2014)	Scientific article	Abstract	Diseases	7639	174 487	12 128	6881	2192
Arizona Disease (Leaman et al., 2009)	Scientific article	Abstract	Diseases	2884	76 489	7358	3206	1188
SCAI (Gurulingappa et al., 2010)	Scientific article	Abstract	Diseases	4332	104 015	12 558	2226	1048
S800 (Pafilis et al., 2013)	Scientific article	Abstract	Species	7933	195 197	20 526	3646	1564
LocText (Goldberg et al., 2015)	Scientific article	Abstract	Species	949	22 550	4371	276	39
Linneaus (Gerner et al., 2010)	Scientific article	Full-text	Species	17 788	473 148	34 396	4077	419
CLL (Kaewphan et al., 2016)	Scientific article	Abstract, Full-text	Cell Lines	215	6547	2468	341	311
Gellus (Kaewphan et al., 2016)	Scientific article	Abstract, Full-text	Cell Lines	11 221	278 910	25 628	640	237

2.4 Word embeddings

We utilized word embedding techniques to capture the semantics of words (and their similarities) based on their surrounding words. We evaluated three different embedding models. The first model, denoted PubMed-PMC, was trained on a combination of PubMed abstracts (nearly 23 million abstracts) and PMC articles (nearly 700, 000 full texts). The second model, denoted Wiki-PubMed-PMC, was trained using these two collections plus approximately four million English Wikipedia articles. The second model thus mixes domain-specific texts with domain-independent ones. Both models were created by Pyysalo using Google's word2vec (http://bio.nlplab.org/); we use 200D vectors. To also be able to study the impact of text genre, we trained a third model with 50 dimensions, denoted Patent, on a set of roughly 20 000 European patents with biomedical topics using the Gensim toolkit. We optimized the hyper-parameters of Gensim for the F1-score of the CRF model (see Section 2.2) on the development set of the CHEMDNER patent corpus.

2.5 Text pre-processing

All corpora first were converted into a simple uniform format to ease further processing. In this format, all texts and all annotations are stored in one single file. Each document is represented by a document identifier, a tab separator and the entire document text in one line. Annotations are given in the following lines, one annotation per line, with start character position end character position, entity mention text, and entity type. An empty line indicates the end of a document, and the next document starts after this empty line. We converted this file into the specific input format defined by each baseline NER tool. Moreover, we used Apache OpenNLP (https://opennlp.apache.org/) to split documents into sentences, sentences into tokens, and to assign part-of-speech tags. Finally, we created a file in CoNLL2003 (Tjong Kim Sang and De Meulder, 2003) format as input for the LSTM-CRF model and the CRFSuite tool.

2.6 Evaluation metrics

We randomly divided each corpus into three disjoint subsets. 60% of the samples were used for training, 10%, as the development set for the training of methods, and 30% for the final evaluation. We compared all methods in terms of precision, recall and F1-score on the test sets. We performed exact matching to compute these performance values. We also performed an error analysis by comparing the sets of false positives (FPs) and false negatives (FNs) of the different NER methods. To this end, we measured the number of FP and FN counts for each mention by each method and then calculated the overlap between sets of FP or FN using fuzzy set operations that take into account the frequency of mistakes per entity mention (Thole ).

3 Results

We assessed the performance of a novel method for entity-type independent NER, namely LSTM-CRF, on 33 different evaluation sets covering five different types of biomedical entities. LSTM-CRF uses as features only low-dimensional representations of the words in the vicinity of the to-be-classified tokens, created by mixing word-level embeddings created in an unsupervised fashion with character-level embeddings trained on the respective corpus. Results were compared with a traditional CRF using typical NER features and the same word embeddings, and to type-specific baselines representing the state-of-the-art in biomedical NER.

3.1 Impact of different word embeddings

We first studied the impact of using different word embeddings. We compared the results of three models, differing only in the unsupervised part, i.e. the text collections used for computing word-level embeddings. For evaluation, we considered the LSTM-CRF and the pure CRF approach. The macro averaged performance values over all corpora in terms of precision, recall and F1-score are provided in Table 2; detailed performance values are given in Appendix A. In five out of six cases, Wiki-PubMed-PMC achieves the best performance, and this model is also very close to the best one in the sixth case. Based on this data, we used the Wiki-PubMed-PMC embeddings in all further evaluations.

Table 2.

Macro averaged performance values in terms of precision, recall and F1-score for CRF and LSTM-CRF methods with word embedding features: (i) Patent, (ii) PubMed-PMC and (iii) Wiki-PubMed-PMC

	Precision (%)			Recall (%)			F1-score (%)
	(i)	(ii)	(iii)	(i)	(ii)	(iii)	(i)	(ii)	(iii)
CRF	82.71	84.55	84.49	71.98	73.07	73.26	76.36	77.98	78.04
LSTM-CRF	80.10	81.39	81.77	81.04	80.72	81.08	80.26	80.79	81.11

The highest values for each method are represented in bold.

3.2 Performance of LSTM-CRF, CRF, and baselines

We compared the performance of five baseline NER tools, the CRF, and the LSTM-CRF using Wiki-PubMed-PMC embeddings on 33 evaluation sets. Results in terms of F1-score for each entity type and each corpus are shown Figure 2; exact precision, recall, and F1 values are given in Appendix A. LSTM-CRF achieves the best performance for 28 out of 33 evaluations and is very close to the best method in the remaining five cases. On average (macro average), F1-scores are 81.11% for the generic LSTM-CRF method, 78.04% for the generic CRF method and 76.61% for the baselines. The improvements are mainly due to a strong increase in recall (mean 81.08%, 73.26%, 75.13%) at the cost of decrease in precision (mean 81.77%, 84.49%, 80.38%).

Fig. 2

F1-scores of baseline (B), generic CRF (C) and generic LSTM-CRF (L) for five entity types, each measured within 4–12 corpora. The score for each corpus per entity type is depicted by a specific colored circle

We also computed performance metrics aggregated per entity type to see if methods are more suitable for some types than for the others. Both macro and micro averaged performance values in Table 3 reveal that this does not seem to be the case; LSTM-CRF achieves the best average F1-score and recall for all entity types.

Table 3.

Macro and micro averaged performance values in terms of precision, recall and F1-score for the baselines (B), the generic CRF method (C) and the generic LSTM-CRF method (L) over the corpora per each entity type

a) Macro averaged performance
	Precision (%)			Recall (%)			F1-score (%)
	B	C	L	B	C	L	B	C	L
Chemicals	79.8	85.81	82.82	80.73	79.26	84.77	80.22	82.32	83.71
Diseases	78.54	82.48	81.66	73.67	73.47	78.70	75.89	77.56	80.11
Species	72.75	89.59	80.84	79.63	76.15	87.64	73.03	82.09	83.60
Gene/protein	83.77	82.54	81.57	75.87	74.67	80.06	79.35	78.15	80.58
Cell lines	85.14	84.04	82.65	61.32	56.91	73.22	70.35	66.96	77.2
Average	80.38	84.49	81.77	75.13	73.26	81.08	76.61	78.04	81.11
b) Micro averaged performance
	Precision (%)			Recall (%)			F1-score (%)
	B	C	L	B	C	L	B	C	L
Chemicals	75.16	83.45	81.81	79.7	79.12	83.86	77.36	81.23	82.82
Diseases	79.29	83.62	82.56	75.98	77.17	81.68	77.60	80.26	82.12
Species	77.33	88.92	83.84	67.74	76.68	83.33	72.22	82.35	83.59
Gene/protein	81.34	81.79	81.50	77.45	78.14	82.71	79.35	79.92	82.10
Cell lines	71.82	70.20	68.59	57.20	57.60	66.84	63.68	63.28	67.70
Average	76.4	83.11	81.72	78.87	78.71	83.45	77.62	80.85	82.58

The highest values for each entity class are highlighted in bold.

An interesting observation is that the pure CRF method often performs better than the entity-specific baseline algorithms. There are two explanations for this apparent contradiction to our introductory words, claiming that the best methods to date are CRF with type-specific feature sets. First, NER methods are often developed for a very specific sub-problem, and often only have excellent performance on a particular evaluation corpus. Our evaluation on a much larger set of corpora reveals that such corpus-specific advantages cannot be simply extrapolated to other settings. Second, by using word embeddings our CRF model represents words in a context-encoding space, which enables it to recognize semantic similarity between words. This feature is missing in the word-based baseline tools.

3.3 Error analysis

We compared the errors made by the three methods by computing intersections of the sets of FPs and FNs for each method and each entity type. Results per entity class are shown in Figure 3.

Fig. 3

Venn diagrams demonstrating the area of overlap among the FP sets or the FN sets of the three methods: the baseline (B), the generic CRF method (C) and the generic LSTM-CRF method (L) per entity type

The Venn diagrams of FP sets or FN sets for the different entity types follow a similar pattern. Generally, error sets of CRF and LSTM-CRF are more similar to each other than to errors of the baseline methods, probably due to the strong reliance of all baselines on entity type-specific dictionaries, creating their own specific errors. Relying on dictionaries carries two types of dangers: first, they are notoriously incomplete, leading to FNs; second, dictionary matching disregards context which leads to FP matches in case of ambiguous terms. This is particularly visible for species, where the baseline only uses dictionary matching, leading to the worst precision among all entity types.

We also speculated that the length of an entity might have an impact on the performance of approaches. Very short entities are often abbreviations which are notoriously ambiguous; entities consisting of multiple tokens often produce errors at the border tokens, which are severely punished in any exact matching evaluation (one FP and one FN). To this end, we measured the average length of a FP or FN mention in terms of non-space characters for the three methods per entity class. Furthermore, we computed the F1-scores of single- and multi-token mentions separately. Results are shown in Table 4. First, LSTM-CRF is the best method regardless whether an entity name consists of only one or of multiple tokens. Only for species, the dictionary-based baseline method has a slightly better F1-score for multi-token entities, which is, however, outweighed by a much worse F1-score for single entity names in the overall evaluation. This result shows that the LSTM-CRF method with word embeddings manages to tag precisely also multi-token entities, without relying on any post-processing rules. However, we also observe that the performance of CRF and LSTM-CRF on single-token mentions is considerably better than on multi-token entities, showing that there is still room for improvement regarding such entity names. Second, there is an interesting tendency that FPs tagged by LSTM-CRF are slightly shorter than those found by the CRF, while FNs are slightly longer, indicating that LSTM-CRF seems to be biased towards shorter entity names.

Table 4.

The average length of errors from the lists of FPs and FNs, and the F1-scores of single-token and multi-token entities measured for baselines (B), generic CRF methods (C) and generic LSTM-CRF methods (L) per entity type

		F1-score (%)		Mention length
		Single-Token	Multi-Token	FP	FN
Chemicals	L	84.02	79.90	19.07	20.69
	C	82.53	78.07	25.1	20.11
	B	76.77	76.84	19.13	16.49
Diseases	L	86.13	76.54	14.51	15.58
	C	84.20	74.92	15.09	14.66
	B	80.42	73.80	13.90	14.36
Species	L	86.11	79.10	10.68	12.34
	C	85.57	76.75	11.55	11.52
	B	67.72	79.80	8.00	9.15
Genes/Proteins	L	86.49	72.64	12.85	14.13
	C	84.77	69.32	13.50	13.52
	B	83.21	70.99	11.73	11.98
Cell Lines	L	72.94	64.93	20.02	15.65
	C	64.77	62.52	19.01	15.24
	B	62.55	64.24	18.41	13.91

The highest F1-scores are emphasized in bold.

3.4 Precision-recall trade-off

The results in Table 3 show that, on average, LSTM-CRF significantly outperforms CRF and baselines in terms of recall at the expense of a less strong decrease in precision. Since, for some applications, obtaining high precision is more important than high recall, we also implemented and evaluated a method which assigns confidence scores to the predictions made by LSTM-CRF, followed by an entity-type independent filter for low confidence predictions with the aim to reduce the number of FP entities. The filter removes a percentage of entities with the lowest confidence values from the output of the generic LSTM-CRF method. First, the labels for all the tokens of a sentence are predicted by Viterbi decoding using the trained model (as described in Section 2.1). Then, all tokens labeled as entity are assigned a confidence score which estimates the probability (given the LSTM-CRF model) of the predicted entity label being part of the true labels of the sentence. The score for a given entity is obtained by summing the scores of all possible label sequences of the sentence which produce this entity label normalized by the sum of scores of all the possible label sequences. These scores can be obtained by the constrained forward-backward algorithm (Culotta and McCallum, 2004) (The implementation is available online at “https://github.com/leonweber/tagger”.).

We removed 5, 10 and 15% of the entities tagged by the generic LSTM-CRF model with lower confidence values and, on average, obtained precision (recall) scores of 83.62% (78.97%), 85.16% (76.22%) and 86.45% (73.05%), respectively; detailed performance values are given in Appendix A. Across all evaluations, the generic LSTM-CRF model without the 10% lowest confidence predictions obtains, on average, higher precision and higher recall values than the generic CRF model and the baselines. In Figure 4, we compare precision and recall values of the LSTM-CRF method before and after applying filters together with CRF and baselines, macro-averaged by entity type. For four out of five entity types, precision and recall values of this configuration is very close to or higher than those of the generic CRF method and of the baselines. The only exception is species, where even the most stringent filtering that we applied cannot make the LSTM-CRF method reach a precision higher than that of the CRF method. We inspected these evaluations in more detail. The lower precision values on average are mostly due to the two corpora CellFinder and Variome. However, both corpora miss some of the species annotations. For instance, the 41 FP predictions of LSTM-CRF for Variome are 21 times the term ‘mouse’, 4 times ‘mice’ and 16 times ‘human’—all of which are valid species names, and all of which achieve high prediction confidence. The situation is similar for CellFinder, with tokens like ‘goat’, ‘human’ or ‘EColi’ being not annotated as species.

Fig. 4

Aggregated precision and recall of the generic LSTM-CRF method before (L) and after applying filters, removing 5 (L—5%), 10 (L—10%) and 15 (L—15%) of entities, per entity type. The highest averaged precision and recall per entity type obtained by baselines or the generic CRF model are represented by the green and the orange dash line, respectively

4 Discussion

4.1 Genre and word embeddings

The word embeddings we used for our study were either derived from patents, from scientific articles or from a mixture of scientific articles and Wikipedia entries. Since also some of our corpora are based on patents and others on scientific articles, we wondered if the use of patent-derived embeddings is advantageous for patent corpora yet less adequate for scientific articles and vice versa. Indeed, our patent-derived embeddings slightly outperform the others on patent-derived corpora in terms of F1-score (recall) (Patent: 78.52% (82.75%), PubMed-PMC: 78.48% (81.85%), Wiki-PubMed-PMC: 77.84% (81.00%)) and the model derived from Wiki-PubMed-PMC achieves a better F1-score (recall) on corpora consisting of scientific articles (Patent: 80.50% (80.81%), PubMed-PMC: 81.11% (80.57%) and Wiki-PubMed-PMC: 81.56% (81.09%)). This observation again shows that patents are different from scientific articles (Habibi ) and that their analysis calls for specific resources.

4.2 Corpus size and word embeddings

The performance values obtained in Section 3.1 show that both CRF and LSTM-CRF achieve the best performance using Wiki-PubMed-PMC embeddings in most of the cases. Notably, this model is derived from a collection of domain-specific texts (PubMed, PMC) mixed with domain-unspecific texts (Wikipedia). A possible explanation for its superiority is that it uses the largest text base among the three models; previous works have shown that word embeddings tend to be the more effective the larger the text base is (Stenetorp ). Furthermore, the use of general domain corpora in addition to the domain-specific ones may add more out-of-domain information to the embeddings. However, more investigations are required to define optimal corpora for derivation of word embeddings for concrete tasks.

4.3 Related work

Traditional biomedical NER methods have relied on rule- or dictionary-based approaches. The rule-based techniques recognize biomedical entities using several rules manually defined based on the textual patterns of entities (Narayanaswamy ; Eltyeb and Salim, 2014). These patterns vary depending on the specific textual properties of an entity class. The definition of such entity-specific patterns is time consuming and requires domain-expert knowledge. Dictionary-based methods extract named entities by searching them in dictionaries constructed for each entity type. For instance, Hettne employ a dictionary-based approach for the extraction of drugs, and Gerner use a dictionary-based approach for species names extraction. Again, building such dictionaries is time consuming and challenging (Liu ). Moreover, the recall obtained using these methods is generally low due to the inherent difficulty of the methods in capturing new entities; a strong advantage of dictionary-based methods is that they directly solve the named entity normalization (NEN) problem, i.e. they can output a database identifier for each recognized entity.

Over the last years, pattern- and dictionary- based methods have been superseded by approaches relying on supervised machine learning, especially sequential classification algorithms, such as Hidden Markov Models (Rabiner, 1989) and CRFs (Lafferty ). CRFs have become the de-facto standard model, being the method of choice for essentially all tools winning recent NER competitions, such as BioCreative IV (Krallinger ) or i2b2 (Uzuner ). Popular biomedical NER tools using CRFs are, for instance, ABNER (A Biomedical Named Entity Recognizer) (Settles, 2005) and BANNER (Leaman and Gonzalez, 2008). Hybrid methods combine machine learning methods with dictionary- or rule-based techniques. For instance, ChemSpot (Rocktäschel ) integrates results of a CRF model with a dictionary-matching module for chemical NER, and Gimli (Campos ) applies post-processing steps like parentheses balancing to the output of the CRF models.

All these methods build upon pre-defined sets of features, whose correlations with other features and the target class of tokens are learned from the gold standard annotations. In contrast, methods based on deep ANNs also consider non-linear combinations of feature values (Hastie ). This drastically increases the search space, leading to the fact that ANNs for long were considered impractical for many applications. This situation changed only recently due to the steep increase in the compute power of machines. When combined with specifically trained word embeddings, deep learning with ANNs has been shown to outperform other methods in many areas, such as sentiment analysis (Dai and Le, 2015) or language modeling (Jozefowicz ). These methods are also gradually entering the field of biomedical information extraction, yet results so far have been mixed. Segura-Bedmar used word embeddings as input to a CRF and reported only marginal effects in chemical NER. Liu ) found word embeddings to only marginally improve the performance of a CRF-based method using comprehensive dictionaries as features. Tang compared several ways of obtaining word embeddings and reported up to 2% increase in recall. The method we use in this paper, LSTM-CRF, was proposed as a general NER algorithm by Lample . A few previous works have applied LSTM-CRF for biomedical NER (e.g. Chalapathy ,b). However, all these evaluations considered at most a handful of corpora. In contrast, our evaluation of biomedial NER methods based on 33 evaluations using 24 corpora to our knowledge is the most comprehensive one ever performed in this field. Larger previous evaluations we are aware of were performed in Campos , who measured the performance of different machine learning-based NER methods on 14 corpora, and Batista-Navarro , who reported on the performance of a CRF-based NER tool using nine different corpora.

5 Conclusion

In summary, our results indicate that LSTM-CRF improves considerably upon current biomedical NER methods. We find this exciting particularly because the method is completely agnostic to entity types; thus, the costly development of specific tools using specific dictionaries could become superfluous. However, further research is necessary to turn this observation into useful applications. In particular, LSTM-CRF only helps with the recognition of entities in texts; the next step in most text-mining applications, which is mapping recognized entities to standard nomenclature (e.g. Entrez-IDs for genes, ChEBI-IDs for chemicals etc.), is not addressed. Thus, LSTM-CRF should be combined with generic NEN tools, as for instance presented in (Leaman ). In the future, we plan to study the inclusion of background knowledge into the LSTM-approach, for instance in the form of post-processing rules to deal with long multi-token entities; the hope is to achieve further improvements especially in terms of precision, though the price will be to lose the entity-independence. Nevertheless, our results imply that the advances in statistical NLP and machine learning can help to make biomedical text mining (i) more accurate, (ii) less laborious to develop and (iii) more robust with respect to the specific texts being mined.

Table A1.

Precision, recall and F1-scores obtained for each corpus by the baselines and eight variants of the CRF and the LSTM-CRF methods

		Baseline	CRF	CRF (i)	CRF (ii)	CRF (iii)	LSTM- CRF	LSTM- CRF (i)	LSTM- CRF (ii)	LSTM- CRF (iii)	LSTM-CRF (iii) 5%	LSTM-CRF (iii) 10%	LSTM-CRF (iii) 15%
a) Precision scores
Chemicals	CHEMDNER patent	82.56	84.52	84.29	84.42	84.29	81.09	82.25	83.35	83.33	84.66	86.14	87.44
	CHEBI	71.46	80.59	78.56	78.68	79.67	71.58	71.1	72.95	69.5	71.3	72.85	74.57
	BioSemantics	72.56	82.17	82.1	81.93	81.96	80.64	82.07	79.72	80.88	82.4	83.81	85.11
	CHEMDNER	83.23	90.4	90.52	90.66	90.79	87.31	87.27	88.25	87.83	89.97	91.77	93.47
	CDR	89.19	91.04	91.47	91.81	92.37	88.19	91.12	92.18	92.57	94.61	95.98	97.16
Genes/Proteins	CHEMDNER patent	67.02	67.01	67.33	67.66	67.13	61.05	64.02	65.74	66.23	67.3	68.32	69.29
	BioCreative II GM	79.92	76.45	77.07	77.27	77.22	73.79	77.22	78.99	77.5	79.6	81.5	83.81
	JNLPBA	72.74	73.28	73.21	73.39	73.65	71.86	73.07	74.77	74.83	76.59	78.16	79.71
	CellFinder	81.98	74.39	77.27	84.05	83.77	78.49	82.01	83.56	88.73	91.19	93.12	95.28
	OSIRIS	83.33	77.92	76.88	81.38	78.61	74.07	76.52	76.73	78.02	80	81.73	82.74
	DECA	81.89	76.89	76.6	78.1	77.66	64.11	67.22	72.91	75.27	76.76	78.67	80.05
	Variome	90.32	90.08	89.56	89.92	90.12	87.63	87.64	86.59	87.47	89.41	90.72	91.33
	PennBioIE	88.12	88.14	88.32	89.5	89.58	83.31	86.51	88.02	86.97	89.34	91.21	93.11
	FSU-PRGE	87.79	87.04	86.71	87.33	87.2	81.69	85.09	86.03	87.26	89.7	91.51	92.88
	IEPA	90.68	88.28	87.08	90.11	88.89	91.01	84.67	86.6	88.01	89.21	89.35	90.76
	BioInfer	94.16	91.57	91.05	92.38	92.01	87.73	90.87	91.71	92.64	94.72	95.86	96.76
	miRNA	87.35	83.45	80.31	85.11	84.69	70.15	69.94	75	75.95	78.47	80.38	80.82
Diseases	NCBI Disease	80.11	84.81	84.25	84.87	85.18	81.7	86.16	86.43	85.31	87.49	89.1	90.4
	CDR	80.61	83.54	83.26	83.36	83.72	79.24	81.06	83.3	84.19	84.63	86.4	88.17
	Variome	78.42	85.42	84.68	85.17	85.28	83.13	83.42	85.12	84.51	86.21	87.62	88.89
	Arizona Disease	76.11	79.54	78.41	80.91	79.89	72.64	74.52	79.15	76.64	78.62	81.17	82.89
	SCAI	76.99	82.85	81.47	81.51	81.3	67.4	74.97	77.21	78.46	80.1	81.61	83.49
	miRNA	79.01	81.56	77.9	80.27	79.54	80.39	77.44	80.84	80.86	82.81	83.92	85.59
Species	S800	77.66	77.33	76.55	76.8	76.8	69.67	73.39	72.67	74.55	75.95	77.27	78.99
	CellFinder	82	93.18	89.9	90.22	91.3	81.98	83.76	81.67	79.03	80.51	83.04	83.02
	Variome	32.17	82.98	81.25	83.33	83.33	50.46	66.27	62.92	59	61.05	62.22	64.71
	LocText	84.78	95.45	92.5	95.65	94.37	94.74	93.48	90.36	91.01	92.94	92.59	93.42
	Linneaus	87.4	98.26	96.21	97.73	97.63	93.23	91.54	92.49	93.57	95.27	96.38	96.99
	miRNA	72.54	92.67	85.06	94.19	94.16	91.67	89.29	86.13	87.88	92.95	94.59	94.29
Cell lines	CLL	83.33	71.88	73.97	77.78	81.97	76.47	82.43	77.63	86.84	87.67	91.3	90.77
	Gellus	93.18	93.6	86.25	93.55	93.42	70.89	85.64	90.23	89.53	91.46	91.61	94.56
	CellFinder	98.72	94.12	95.45	97.01	96.67	89.47	87.34	97.06	92.52	93.14	95.88	95.6
	JNLPBA	65.34	64.8	64.19	64.11	64.11	51.48	54.2	59.83	61.73	63.54	64.75	66.95
b) Recall scores
Chemicals	CHEMDNER patent	84.53	83.09	83.99	84.14	84.14	86.23	90.41	87.72	87.53	85.28	82.19	78.8
	CHEBI	66.87	60.61	64.81	64.06	65.58	71.44	80.27	78.52	79	76.99	74.52	72.05
	BioSemantics	78.37	78.12	78.29	78.38	78.34	79.29	82.56	84.42	83.14	80.47	77.54	74.37
	CHEMDNER	85.08	78.27	80.51	81.58	81.8	79.8	84.96	83.17	85.45	83.36	80.54	77.48
	CDR	88.83	81.25	84.63	86.03	86.46	85.58	89.86	89.94	88.77	86.18	82.84	79.18
Genes/Proteins	CHEMDNER patent	69.85	64.19	67.69	68.88	68.17	69.9	77.76	76.74	74.34	72.84	70.04	67.1
	BioCreative II GM	75.21	67.97	69.69	70.83	70.89	72.46	77.72	78.16	78.13	76.19	73.9	71.78
	JNLPBA	72.77	73.31	73.93	74.96	75.14	75.32	79.11	79.22	79.82	77.62	75.04	72.27
	CellFinder	63.17	56.53	63.47	67.47	68.8	36	63.2	65.07	65.07	63.47	61.33	59.2
	OSIRIS	63.49	48.58	57.89	61.94	63.97	56.68	76.52	76.11	73.28	71.26	68.83	65.99
	DECA	58.24	55.12	57.7	61.25	61.09	67.68	73.45	66.82	66.16	64.08	62.21	59.78
	Variome	93.51	90.2	92.24	93.47	93.06	91.56	94.56	96.6	96.87	94.15	90.48	85.99
	PennBioIE	82.61	80.54	81.82	82.35	82.76	82.98	87.09	85.44	86.54	84.47	81.7	78.77
	FSU-PRGE	84.14	81.25	82.27	83.42	83.44	84.25	87.55	88.23	87.24	85.79	82.91	79.48
	IEPA	85.18	75.33	78.67	79	80	81	84.67	88.33	85.67	82.67	78.33	75.33
	BioInfer	89.44	88.42	89.98	90.06	90.14	90.06	92.72	91.78	89.59	87.01	83.41	79.5
	miRNA	72.93	42.76	54.77	56.54	58.66	49.82	80.57	79.51	78.09	75.97	73.85	69.96
Diseases	NCBI Disease	79.69	76.62	79.74	80.69	81.59	75.86	83.06	82.92	83.58	81.45	78.56	75.3
	CDR	78.36	73.21	76.47	79.15	79.17	76.98	83.15	83.38	82.79	81.54	78.87	76.01
	Variome	74.66	81.2	81.47	81.64	81.82	84.38	84.82	86.32	87.64	85	81.82	78.38
	Arizona Disease	59.3	57.64	62.63	64.75	64.3	65.05	70.8	70.65	72.47	70.65	69.14	66.72
	SCAI	70.6	48.45	59.6	61.02	62.01	56.07	76.55	70.34	70.48	68.22	65.82	63.56
	miRNA	79.41	65.05	73.18	71.11	71.97	70.93	78.37	76.64	75.26	73.36	70.42	67.82
Species	S800	69.08	55.79	60.09	59.72	60.93	62.9	64.95	69.35	69.81	67.57	65.14	62.9
	CellFinder	87.85	77.36	83.96	78.3	79.25	85.85	92.45	92.45	92.45	89.62	87.74	83.02
	Variome	98.48	59.09	59.09	60.61	60.61	83.33	83.33	84.85	89.39	87.88	84.85	83.33
	LocText	86.66	68.48	80.43	71.74	72.83	78.26	93.48	81.52	88.04	85.87	81.52	77.17
	Linneaus	64.56	79.37	84.64	90.78	90.34	76.21	94.03	87.53	93.24	90.25	86.39	82.09
	miRNA	71.15	89.1	94.87	93.59	92.95	91.67	96.15	95.51	92.95	92.95	89.74	84.62
Cell lines	CLL	77.92	59.74	70.13	63.64	64.94	67.53	79.22	76.62	85.71	83.12	81.82	76.62
	Gellus	49.8	47.37	55.87	58.7	57.49	61.13	67.61	63.56	62.35	60.73	57.49	56.28
	CellFinder	60.63	25.81	33.87	52.42	46.77	41.13	55.65	79.84	79.84	76.61	75	70.16
	JNLPBA	56.95	55.31	57.13	59.14	58.47	63.44	67.94	66.7	64.98	63.54	61.34	59.9
c) F1-scores
Chemicals	CHEMDNER patent	83.53	83.8	84.14	84.28	84.22	83.58	86.14	85.48	85.38	84.97	84.12	82.9
	CHEBI	69.09	69.18	71.03	70.63	71.94	71.51	75.41	75.63	73.95	74.03	73.67	73.29
	BioSemantics	75.35	80.1	80.15	80.12	80.11	79.96	82.32	82.01	81.99	81.42	80.55	79.27
	CHEMDNER	84.15	83.9	85.22	85.88	86.06	83.38	86.1	85.63	86.62	86.54	85.79	84.73
	CDR	89.01	85.87	87.92	88.83	89.31	86.87	90.48	91.05	90.63	90.2	88.93	87.25
Genes/Proteins	CHEMDNER patent	68.4	65.57	67.51	68.26	67.65	65.17	70.22	70.81	70.05	69.96	69.17	68.18
	BioCreative II GM	77.49	71.96	73.2	73.91	73.92	73.12	77.47	78.57	77.82	77.86	77.52	77.33
	JNLPBA	72.75	73.3	73.57	74.17	74.39	73.55	75.97	76.93	77.25	77.1	76.57	75.81
	CellFinder	71.35	64.24	69.69	74.85	75.55	49.36	71.39	73.16	75.08	74.84	73.95	73.03
	OSIRIS	72.06	59.85	66.05	70.34	70.54	64.22	76.52	76.42	75.57	75.37	74.73	73.42
	DECA	68.06	64.21	65.82	68.65	68.39	65.85	70.2	69.73	70.42	69.85	69.48	68.45
	Variome	91.88	90.14	90.88	91.66	91.57	89.55	90.97	91.32	91.93	91.72	90.6	88.58
	PennBioIE	85.27	84.17	84.94	85.78	86.04	83.14	86.8	86.71	86.75	86.84	86.19	85.34
	FSU-PRGE	85.92	84.05	84.43	85.33	85.27	82.95	86.3	87.12	87.25	87.7	87	85.66
	IEPA	87.84	81.29	82.66	84.19	84.21	85.71	84.67	87.46	86.82	85.81	83.48	82.33
	BioInfer	91.73	89.97	90.52	91.2	91.07	88.88	91.79	91.75	91.09	90.7	89.21	87.29
	miRNA	79.49	56.54	65.13	67.94	69.31	58.26	74.88	77.19	77	77.2	76.98	75
Diseases	NCBI Disease	79.89	80.51	81.94	82.73	83.35	78.67	84.58	84.64	84.44	84.36	83.5	82.16
	CDR	79.46	78.03	79.72	81.2	81.38	78.1	82.09	83.34	83.49	83.06	82.46	81.64
	Variome	76.49	83.26	83.04	83.37	83.51	83.75	84.11	85.71	86.05	85.6	84.62	83.3
	Arizona Disease	66.66	66.84	69.64	71.93	71.25	68.64	72.61	74.66	74.49	74.42	74.67	73.93
	SCAI	73.65	61.14	68.84	69.79	70.35	61.22	75.75	73.61	74.26	73.68	72.87	72.17
	miRNA	79.2	72.38	75.47	75.41	75.57	75.37	77.9	78.69	77.96	77.8	76.58	75.68
Species	S800	73.11	64.82	67.33	67.19	67.95	66.11	68.91	70.97	72.1	71.51	70.69	70.03
	CellFinder	84.82	84.54	86.83	83.84	84.85	83.87	87.89	86.73	85.22	84.82	85.32	83.02
	Variome	48.49	69.03	68.42	70.18	70.18	62.86	73.83	72.26	71.08	72.05	71.79	72.85
	LocText	85.7	79.75	86.05	81.99	82.21	85.71	93.48	85.71	89.5	89.27	86.71	84.52
	Linneaus	74.26	87.81	90.05	94.13	93.84	83.86	92.77	89.94	93.4	92.7	91.11	88.92
	miRNA	71.83	90.85	89.7	93.89	93.55	91.67	92.59	90.58	90.34	92.95	92.11	89.19
Cell lines	CLL	80.54	65.25	72	70	72.46	71.72	80.79	77.12	86.27	85.33	86.3	83.1
	Gellus	64.91	62.9	67.81	72.14	71.18	65.65	75.57	74.58	73.51	72.99	70.65	70.56
	CellFinder	75.12	40.51	50	68.06	63.04	56.35	67.98	87.61	85.71	84.07	84.16	80.93
	JNLPBA	60.86	59.68	60.46	61.52	61.16	56.84	60.3	63.08	63.31	63.54	63	63.23

CRF and LSTM-CRF stand for the CRF and the LSTM-CRF methods without word embedding features. The other ones represented by numbers are using one of the three word embeddings: (i) Patent, (ii) PubMed-PMC and (iii) Wiki-PubMed-PMC. The ones represented by percentage numbers indicate the performance of the method after removing a percentage of predicted entities with lower confidence values.