Proceedings of the Second Workshop on Scholarly Document Processing, pages 36–48June 10, 2021. ©2021 Association for Computational Linguistics

36

The Biomaterials Annotator: a system for ontology-based conceptannotation of biomaterials text

Javier Corvi1, Carla V. Fuenteslópez2, José M. Fernández1, Josep Lluis Gelpí1,3,Maria Pau Ginebra4, Salvador Capella-Gutierrez1 and Osnat Hakimi1,5

1Barcelona Supercomputing Center (BSC), Barcelona, Spain2Institute of Biomedical Engineering, Botnar Research Centre, University of Oxford, UK

3Dept. of Biochemistry and Molecular Biology, University of Barcelona, Spain4Dept. of Material Science and Engineering, Universitat Politècnica de Catalunya, Spain5Faculty of Medicine and Health Sciences, Universitat Internacional de Catalunya, Spain

javier.corvi@bsc.esosnat.hakimi@gmail.com

Abstract

Biomaterials are synthetic or natural materialsused for constructing artificial organs, fabricat-ing prostheses, or replacing tissues. The lastcentury saw the development of thousands ofnovel biomaterials and, as a result, an expo-nential increase in scientific publications in thefield. Large-scale analysis of biomaterials andtheir performance could enable data-drivenmaterial selection and implant design. How-ever, such analysis requires identification andorganization of concepts, such as materials andstructures, from published texts. To facilitatefuture information extraction and the applica-tion of machine-learning techniques, we de-veloped a semantic annotator specifically tai-lored for the biomaterials literature. The Bio-materials Annotator has been implemented fol-lowing a modular organization using softwarecontainers for the different components and or-chestrated using Nextflow as workflow man-ager. Natural language processing (NLP) com-ponents are mainly developed in Java. Thisset-up has allowed named entity recognitionof seventeen classes relevant to the biomateri-als domain. Here we detail the development,evaluation and performance of the system, aswell as the release of the first collection of an-notated biomaterials abstracts. We make boththe corpus and system available to the commu-nity to promote future efforts in the field andcontribute towards its sustainability.

1 Introduction

The last two decades saw the field of biomate-rials and tissue engineering grow from a smallniche of biomedical research to an extensive do-main, covering topics such as functional materi-als, cell-material interaction, nanomaterials andmedical devices. The expanding scientific data

generated by the field is primarily available intext documents, such as peer-reviewed researchpapers, patents and conference abstracts. This ever-growing knowledge is increasingly harder for re-searchers to efficiently discover, organize and use.For example, systematically reviewing the appli-cations and scaffolds made of a commonly usedpolymer such as poly-lactic-glycolic-acid (PLGA),requires skimming through >12,000+ abstracts(MEDLINE search on October 2020). Among thedifferent alternatives for the automated processingof available texts, Natural Language Processing(NLP) workflows for information retrieval and in-dexing offer a much needed automated solution.Such computational workflows facilitate informa-tion discovery, information extraction and orga-nization, saving researchers time and minimizingmanual tasks.

Central to information retrieval and indexing isthe extraction of concepts of interest, also knownas Named Entity Recognition (NER). NER is anintegral part of NLP workflows as it allows the au-tomated identification of concepts in unstructuredtext and its assignment to a pre-defined categoryor class. For example, in the field of biomaterials,categories may include ‘Biomaterials’ (‘PLGA’),‘Structures’ (such as ‘fibre’ or ‘sponge’) and ‘Tis-sues’ (such as ‘tendon’ or ‘bone’). The use of NERto automatically recognize entities enables severaldownstream applications, including machine trans-lation, information retrieval and indexing as wellas automated question-answering mechanisms.

The recognition of concepts in the biomaterialsdomain is complicated by language and terminol-ogy originating from multiple scientific disciplines(chemistry, engineering, biology, medicine). A sig-nificant challenge lies in identifying and combining

37

Figure 1: Overview of the workflow used in the development and validation of the Biomaterials Annotator.

lexical and semantic resources across domains, andthus to date there are no automatic biomaterials-specific NER systems to detect relevant entities ofinterest.

Here, we report the development of the firstbiomaterials-specific annotation system, designedto recognize named entities from seventeen differ-ent categories, reflecting the complexity and diver-sity of contemporary biomaterials research. Whenconsidering approaches for the design of the Bio-materials Annotator, i.e. lexical versus machinelearning-based NER, such as CRF or RNN, it wasessential to consider the number of desired annota-tion categories in the system (17) and the absenceof an annotated corpora for text mining efforts forthe majority of them. Based on these premises, itwas concluded that training a model for each cat-egory was impractical. Thus, the system relies onmanually curated and validated lexical resources.

To cover entities from different domains, mul-tiple nomenclatures, vocabularies, and especiallyontologies were identified and combined. To com-bine these resources into a single instrument, theDevices, Experimental scaffolds and Biomateri-als Ontology (DEB) was used providing the log-ical schema and the definition of key categories(Hakimi et al., 2020).

The resulting open source-system, the Bio-

materials Annotator, along with an annotatedcollection of biomaterials literature, are publiclyavailable for use and further development athttps://github.com/ProjectDebbie/Biomaterials_annotator.

2 Previous relevant work

Unlike general purpose NLP systems, biomedicaldomain-specific tools require advanced approachesto detect classes of interest such as diseases andgene names. In this area, there are several well-known and widely used systems and tools, suchas Metamap (Aronson, 2001) and Pubtator (Weiet al., 2013), which were developed using dif-ferent NER methodologies and approaches, e.g.gazetteers and hand-made rule-based NER; ma-chine learning-based NER that includes HiddenMarkov Model, Conditional Random Fields (CRF)and recurrent neural network (RNN); and HybridNER (Lee et al., 2003; McCallum and Li, 2003;Song et al., 2004; GuoDong and Jian, 2004; Zhao,2004; Yeh et al., 2005; Campos et al., 2013; Songet al., 2018; Dang et al., 2018; Kaewphan et al.,2018; Cho and Lee, 2019). In the context of thiswork, generic text mining tools previously devel-oped for the eTRANSAFE project (Pognan et al.,2021) have been adapted and further developed forthe biomaterials domain.

38

Whilst there are a handful of ontologies in thebiomaterials domain (such the nanoparticle ontol-ogy NPO (Thomas et al., 2011) and the Bone andCartilage Tissue Engineering Ontology BCTEO(Viti et al., 2014)), to the best of the our knowledge,the DEB ontology (Hakimi et al., 2020) is the onlyone that is tailored to link and curate concepts forBiomaterials NER. Therefore, it specifically cov-ers different categories related to the biomaterialsdomain.

3 Methodology overview

To develop the annotation tool, the workflowin Figure 1 was followed. Various corpora ofabstracts were used during the development,covering the general biomaterials literature.These corpora included a collection of man-ually curated abstracts of the biomedicalpolymer polydioxanone (Fuenteslópez et al 2021,manuscript in preparation, GitHub repository:https://github.com/ProjectDebbie/polydioxanone_project) and a previouslypublished biomaterials gold standard collection(Hakimi et al., 2020), comprising a total of 1222abstracts. Corpora were passed through four steps,each described in detail below. The first step was atext preprocessing component (section 3.1). Thiswas followed by concept recognition (section 3.2),initially using the MeSH controlled vocabularyand the DEB ontology. Then, the annotationswere evaluated by two domain experts, errorswere flagged up and additional lexical resourceswere added through keyword searches. Conceptrecognition, manual evaluation and curation oflexical resources were performed in an iterativemanner during the development phase (section 3.3)over 1000 abstracts. Once the development phasewas completed, validation by domain experts wasperformed on 199 independent abstracts whichwere not used during the development process(section 4.1). The resulting annotated collectionof biomaterials abstracts was published as opensource.

3.1 Text preprocessing

To prepare the text for concept recognition,several Natural Language Processing (NLP) stepswere performed, namely: tokenization, sentencesplitting, part-of-speech tagging and morpho-logical analysis (Figure 2.A). We developed theStandard NLP preprocessing component which

Figure 2: Overview of the components of the Bioma-terials Annotator; including the standard preprocessingsteps (A) and the biomaterials named entity recognitionsteps (B).

includes the steps previously outlined. Thiscomponent is written in JAVA and it uses theStanford CoreNLP Natural Language Processingopen source toolkit. The use of the StanfordCoreNLP API benefits greatly from the provisionof a set of stable, robust, high quality linguisticanalysis components, which can be easily invokedfor common scenarios (Manning et al., 2014).The Standard NLP preprocessing componentis available at https://gitlab.bsc.es/inb/text-mining/generic-tools/nlp-standard-preprocessing.

3.2 Concept recognition

Here, we developed NER components to detectrelevant entities related to the biomaterials domainbased on the DEB ontology in conjunction withother open relevant resources, such as the NationalCancer Institute Thesaurus (NCIT) and (CHEBI).A comprehensive description of the resourcesincluded in this work is described in section 3.3 andAppendix B. Lexical resources were transformedinto gazetteers to be used in the NER process(Figure 2.B). Internally, the NER process wasdivided into three main steps; the MSH Annotator,which annotates relevant categories from theMeSH terminology; the Dictionary Annotator,which annotated predefined categories from therelevant dictionaries; and the Post-processing stepin which specific rules were executed. These in-clude entity recognition based on lexical rules andthe removal of false positives, among other tasks.

39

The MSH Annotator is available at https://github.com/ProjectDebbie/debbie_umls_annotations; and the Dictionary An-notator and Post-processing rules are available athttps://github.com/ProjectDebbie/DEBBIE_dictionaries_annotations.These components are instances of the nlp-gate-generic-component (https://gitlab.bsc.es/inb/text-mining/generic-tools/nlp-gate-generic-component), ageneric component developed in JAVA by ourteam that uses the General Architecture of TextEngineering (GATE) software (Cunningham et al.,2013) and can be parametrized with gazetteers andspecific handmade JAPE (Java Annotation PatternsEngine) rules. Using the Biomaterials Annotator,every recognised entity is labelled with one of thecategories (Figure 3.A-B).

The nlp-gate-generic-component was configuredto use the GATE Flexible gazetteer, allowing tocapture the words present in the text as well astheir morphological root value (lemma). This en-sures that inflected forms of a word (i.e. plural,singular, -ing forms, tense) can be recognised andanalysed as a single item. In addition, the dictio-naries used in the Biomaterials Annotator includepreferred synonyms, providing the possibility tomap terms semantically to a specific primary con-cept. Thus, the Biomaterials Annotator performssemantic mapping of the annotations by, not onlyrecognizing the category of an entity, but also link-ing it to the appropriate entry in a well-establishedresource (Jovanovic and Bagheri, 2017). For exam-ple, the terms: “canine”, “dogs” and “dog” wereall annotated under the ‘Species’ category; and in-side the features of each annotation the preferredterm is “dog”. This enables the retrieval of all thecorresponding terms using the single search term

‘dog’.

To complete the annotation process, the annota-tor executes JAPE rules for post-processing func-tions, such as the removal of false positives and theaddition of information to each annotation. Addedinformation includes the ontology source, the ontol-ogy term id, the lemma and the preferred synonym(Figure 3.C-D). In addition, JAPE rules were runto identify entities using lexical constraints andaddress the concept recognition of abbreviations.Rule-based entities recognition can use part-of-speech of concepts, as an example; in the caseof ‘Cell’ category, there is a lexical rule defined to

detect concepts:(Token.pos == "JJ" | Token.pos == "NN") To-ken.root == "cell"The inclusion of this rule enables the detection ofCell-type concepts that are not present in the dic-tionaries; e.g. “neuronal cells”, “cancer cell” and

“osteogenic cells”. The discovery of such rules isa continuous work; future Biomaterials Annotatorversions will improve the lexical rules included todetect relevant concepts.

Another key problem to address is the recogni-tion of abbreviation concepts; to achieve this prob-lem we developed a post-processing rule basedon a modified version of Schwartz’s algorithm(Schwartz and Hearst, 2003). First, we detectan abbreviation candidate given a text pattern(regex=”(?:[a-z]*[A-Z][a-z]*)2,”); subsequently,the Schwartz’s algorithm is applied to detectwhether there is a definition that matches the abbre-viation candidate in the sentence; in such case, ifthe definition has an entity class assigned to it, weannotate the abbreviation with the same class. Asan example, in the following sentence: “We investi-gated the potential of human bone marrow derivedMesenchymal stem cells (MSCs) for neuronal differ-entiation in vitro....”; the expression ‘Mesenchymalstem cells’ is annotated under the ’Cell’ category.But the ‘MSCs’ abbreviation is not; moreover in therest of the text the abbreviation is used instead ofits long form. The abbreviation-rule detects ‘MSCs’as an abbreviation of ‘Mesenchymal stem cells’ andassigns the ’Cell’ category to all the ‘MSCs’ men-tions in the text.

3.3 Terminologies and ontologies curationand manual evaluation

One of the main hurdles to biomaterials conceptrecognition is the interdisciplinary nature of thedomain, with scientific texts containing conceptsfrom various fields such as biology, chemistry,engineering and medicine. A key objective ofthe Biomaterials Annotator was to identify andcombine lexical resources from the differentdomains in order to cover as many relevantbiomaterials concepts as possible. Resources wereidentified using a manual, bottom-up approach,with cyclic re-iteration, as shown in Figure 1. As astarting point, abstracts were annotated with theautomated NER approach described in section 3.2using the DEB ontology. After each annotationround, manual evaluation was performed by

40

Figure 3: The appearance of an annotated abstract on GATE’s user interface. A) Shows the annotated text and in B)colored labels used to tag annotations by their respective category. C) Information regarding each annotation (type,position, features), and in D) a specific example: “polymers”: "BiomaterialType" entity with their correspondingfeatures.

B I O M A T E R I A L

A N N O T A T I O N C A T E G O R I E S

B I O L O G I C A L L YA C T I V E

S U B S T A N C E T I S S U E

R E S E A R C HT E C H N I Q U E

C E L L T Y P E

S I L K F I B R O I N

R G D P A N C R E A S S C A N N I N GE L E C T R O N

M I C R O S C O P Y E X A M P L E : F I B R O B L A S T

M A I NS E M A N T I C

R E S O U R C E S :

D E B

M E S H

C H E B I

D E B

M E S H

N C I T

M E S H

U B E R O N

P R E M E D O N T O

M E S H

N C I T

N P O

M E S H

N C I T

U B E R O N

Figure 4: An illustration of part of the annotation schema (showing five out of seventeen categories), which relieson multiple semantic resources for each annotation category. Full details of all categories and resources are in theAppendix A.

41

two domain experts. The evaluation entailedreviewing samples of 10-20 abstracts in orderto flag annotation errors and highlight relevantconcepts which were missed by the system. Theflagged terms were used for keyword search inthe Bioportal (Martínez-Romero et al., 2017) andthe UMLS metathesaurus browser (Bodenreider,2004). Through these searches, specific classes(or ‘parent concepts’) within relevant ontologiesand UMLS ‘semantic types’ were identifiedand added to the annotation schema (part ofwhich is shown for illustration in Figure 4). Theresources identified belonged to three categories:ontologies, controlled vocabularies and nomen-clatures. All the ontologies were open access anddownloaded in .owl format from NCBO Bioportal(http://bioportal.bioontology.org)(Martínez-Romero et al., 2017). The controlledvocabulary (MeSH) was downloaded for useunder license from the UMLS TerminologyServices. The GMDN nomenclature was kindlyprovided in .xml format by the GMDN agencyunder a license. A summary of all the usedresources is in Appendix B. For the resourcesto be used in the annotation system, relevantclasses were imported into a dictionary (gazetteer)containing the following fields: the term, itslabel (annotation category), the ID and wheneveravailable, a preferred synonym. The extraction ofdesired classes from the ontologies to dictionaryformat was done using an implementation ofowlready2 (Lamy, 2017) and the code (namedowl2dict_light) is available in an open githubrepository as part of the (https://github.com/ProjectDebbie/OWL2DICT). Theresulting dictionaries are also available(https://github.com/ProjectDebbie/DEBBIE_dictionaries_annotations).These were in turn used by the DictionaryAnnotator component for concept recognition asdescribed above in section 3.2.

4 Results

4.1 Expert validation

To measure the efficiency of a text mining systemsuch as the Biomaterials Annotator, it is fundamen-tal to organize and plan a validation stage aimed atindicating the performance of the system. The Bio-materials Annotator was validated through manualverification of the validation set, an independentcollection of 199 abstracts. The annotated valida-

tion set, resulting from the execution of the Bio-materials Annotator, was manually verified by 9biomaterials experts. The validation process wasperformed using the GATE user interface, whereannotations made by the system were presentedto the biomaterials experts with the possibility ofadding missing annotations, removing false annota-tions and editing annotations. Once the expert hadfinished the validation of a document, it was savedas a different validated copy.

Two strategies to indicate if two annotationsagree or not were considered; a strict approach,in which the annotations agree if they have thesame origin and end offset, and a more relaxed or“lenient” approach, where the annotations agreeif they overlap at some point. For example, inthe partial approach the biomaterials expressions“polyvinyl alcohol” and “polyvinyl” are consideredto agree, which does not happen in the strict agree-ment.

To measure the performance of the NER system,the set validated by the experts was taken as thegold standard and the system’s output as the setto be validated. Table 1 shows the recall, preci-sion and F-score, including the strict and lenientapproaches, as well as an average between them.The global scores calculated for the system are alsopresented, obtaining an 0.75 strict F-score, 0.79lenient F-Score and 0.77 average F-score.

Figure 5 shows the average F-scores calculatedfor the different categories. Categories with an av-erage F-score above 0.8 are considered categoriesin which the concepts are satisfactorily covered bythe resources used (e.g. Structure, BiomaterialTypeand Tissue). On the other hand, there are categorieswith lower scores, and specifically: ’Biomaterial’,’Biologically active substance’ and ’Cell’. The cat-egories Biomaterials and Biologically active sub-stance had significantly reduced accuracy becausethey include many ambiguous concepts. Some ma-terials may act as a biomaterial in one set-up, butcan also be measured in terms of cell expressionor non-biomaterial use in another set-up (e.g. col-lagen). In the latter case, the human validator willdelete the ‘Biomaterial’ annotation. Solving thiskind of ambiguities will require other strategies,such as specific lexical rules or machine learningapproaches. Another factor impeding good qualityannotations of Biomaterials is the lack of good qual-ity vocabulary of medical polymers. Polymer andco-polymer naming is notoriously variable, with

42

Category Precision- strict

Recall- strict

F-score- strict

Precision- lenient

Recall- lenient

F-score- lenient

Precision- average

Recall- average

F-score- average

Adverse Effects 0.94 0.75 0.82 1 0.8 0.87 0.97 0.77 0.85Associated Biological Process 0.88 0.68 0.77 0.94 0.73 0.82 0.91 0.71 0.79Biologically Active Substance 0.58 0.43 0.49 0.7 0.52 0.59 0.64 0.48 0.54Biomaterial 0.76 0.47 0.57 0.83 0.52 0.63 0.79 0.49 0.6Biomaterial Type 0.92 0.88 0.9 0.98 0.93 0.95 0.95 0.9 0.92Cell 0.76 0.59 0.66 0.84 0.65 0.73 0.8 0.62 0.69Effect On Biological System 0.96 0.69 0.79 1 0.72 0.82 0.98 0.71 0.8Manufactured Object 0.96 0.86 0.9 0.96 0.86 0.9 0.96 0.86 0.9Manufactured Object Component 0.91 0.84 0.86 0.91 0.84 0.87 0.91 0.84 0.87Manufactured Object Features 0.68 0.59 0.62 0.71 0.61 0.65 0.69 0.6 0.64Material Processing 0.78 0.6 0.67 0.83 0.63 0.71 0.81 0.61 0.69Medical Application 0.68 0.49 0.57 0.82 0.6 0.69 0.75 0.54 0.63Research Technique 0.81 0.63 0.71 0.87 0.68 0.76 0.84 0.66 0.73Species 0.97 0.79 0.87 0.99 0.81 0.89 0.98 0.8 0.88Structure 0.93 0.77 0.84 0.95 0.79 0.86 0.94 0.78 0.85Study Type 0.96 0.95 0.96 0.99 0.97 0.98 0.98 0.96 0.97Tissue 0.8 0.77 0.78 0.85 0.82 0.83 0.82 0.8 0.81Global 0.84 0.69 0.75 0.89 0.73 0.79 0.86 0.71 0.77

Table 1: The performance of the Biomaterials Annotator in a test set of 199 abstracts validated manually by 9experts.

some named by their commercial name or abbrevi-ation. To address these inaccuracies, future workwill involve expanding relevant ontologies usingtools such as Spike (Taub-Tabib et al., 2020), in-cluding additional lexical rules, and adding ma-chine learning components.

4.2 Full system implementation andavailability

A significant challenge for scientific software ap-plications is providing facilities to share, distributeand run such systems in a simple and convenientway. Furthermore, an important concern is thepossibility of replicating the results obtainedduring the research. In order to accomplish theserequirements and follow good practices, we de-veloped the Biomaterials Annotator using Dockeras software container technology and Nextflowas the workflow manager. Through the use ofDocker, all the subcomponents of the BiomaterialsAnnotator were individually compartmentalized;hosting their own dependencies and programs thatwork only inside the isolated container. In addition,the Nextflow workflow manager was used forthe automated orchestration and execution of thepipeline. By using this architecture, the entire tool,or any of its individual components, can be easilyinstalled and run in heterogeneous environments.The Biomaterials Annotator is available athttps://github.com/ProjectDebbie/Biomaterials_annotator.

The Biomaterials Annotator is part of DEB-BIE (Database of biomedical materials), awider system that retrieves abstracts frompubmed, annotates using the Biomaterials An-

notator and deposits them in an open accessdatabase. DEBBIE is under development andcan be accessed at https://github.com/ProjectDebbie/DEBBIE_pipeline.

Category CountAdverse Effects 657Associated Biological Process 6231Biologically Active Substance 7709Biomaterial 5726Biomaterial Type 1543Cell 6839Effect On Biological System 972Manufactured Object 5967Manufactured Object Component 2307Manufactured Object Features 4200Material Processing 2728Medical Application 3868Research Technique 3701Species 2089Structure 4136Study Type 1806Tissue 9997Entities 70476Tokens 392605Sentences 15979Abstracts 1222

Table 2: Annotated biomaterials corpus statistics.

4.3 Annotated corpus release

Another key objective was to generate the first an-notated corpus with entities related to the bioma-terials domain. Such a corpus will facilitate thedevelopment and evaluation of text mining mod-els for automated extraction of biomaterials-relateddata from text.

The biomaterials annotated dataset consists of1222 biomaterials abstracts describing the evalu-ation of biomaterials in either a laboratory or a

43

Figure 5: Average F-score of the automated annotations across categories.

clinical setting. Each abstract is individually con-tained as a separate file under the GATE format.Table 2 shows statistics concerning the number ofconcepts corresponding to the different categories,as well as the number of total entities, sentencesand tokens.

The annotated biomaterials corpus isavailable and free for use; informationto access the corpus can be found athttps://github.com/ProjectDebbie/Biomaterials_annotator.

5 Conclusions and future directions

In this work we present the Biomaterials Annota-tor, an ontology-based NER system that identifies17 domain specific types of concepts and deliversan annotated biomaterials corpus of 1222 MED-LINE articles available for future text mining andmachine learning efforts. We have carried out a val-idation activity to measure the performance of theNER system, with the participation of nine bioma-terials experts, obtaining a global average F-scoreof 0.77.

Future work in the development of the systemcould involve annotating relations and linking iden-tified concepts to manufactured biomaterials ob-jects. It may also include incorporating additionalcategories using controlled resources. Improve-ments to the system will continue in an iterative

manner aiming to enhanced performance in keycategories such as Biomaterials and Cells. In addi-tion, future versions of the Biomaterials Annotatorwill be closely related to the DEBBIE system andinclude additional functionalities and features de-veloped to achieve its main objectives.

The Biomaterials Annotator and the annotatedcorpus are open source and available to the com-munity to promote future efforts in the field andcontribute towards its sustainability.

6 Acknowledgements

This project has received funding from the Euro-pean Union Horizon 2020 programme under theMarie Skodowska- Curie grant agreement DEB-BIE, project number: 751277. O. H. is fundedthrough a Bosch-Aymerich fellowship. J-M.F, S.C-G and J-L.P are partly supported by INB Grant(PT17/0009/0001 - ISCIII-SGEFI / ERDF). M-P. G.acknowledges the ICREA Academia Award fromGeneralitat de Catalunya. J.C. is partly supportedby eTRANSAFE (received funding from the Inno-vative Medicines Initiative 2 Joint Undertaking un-der grant agreement No 777365 and support fromthe European Union’s Horizon 2020 research andinnovation programme and EFPIA).

44

ReferencesA. R. Aronson. 2001. Effective mapping of biomedical

text to the UMLS Metathesaurus: the MetaMap pro-gram. Proceedings / AMIA ... Annual Symposium.AMIA Symposium, pages 17–21.

Olivier Bodenreider. 2004. The Unified Medical Lan-guage System (UMLS): Integrating biomedical ter-minology. Nucleic Acids Research.

David Campos, Sérgio Matos, and JoséLuís Oliveira.2013. Current Methodologies for BiomedicalNamed Entity Recognition. In Biological Knowl-edge Discovery Handbook, pages 839–868. John Wi-ley Sons, Inc.

Hyejin Cho and Hyunju Lee. 2019. Biomedicalnamed entity recognition using deep neural net-works with contextual information. BMC Bioinfor-matics, 20(1):735.

Hamish Cunningham, Valentin Tablan, Angus Roberts,and Kalina Bontcheva. 2013. Getting More Out ofBiomedical Documents with GATE’s Full LifecycleOpen Source Text Analytics. PLoS ComputationalBiology.

Thanh Hai Dang, Hoang Quynh Le, Trang M. Nguyen,and Sinh T. Vu. 2018. D3NER: Biomedical namedentity recognition using CRF-biLSTM improvedwith fine-tuned embeddings of various linguistic in-formation. Bioinformatics, 34(20):3539–3546.

Zhou GuoDong and Su Jian. 2004. Exploring deepknowledge resources in biomedical name recogni-tion. page 96.

Osnat Hakimi, Josep Luis Gelpi, Martin Krallinger,Fabio Curi, Dmitry Repchevsky, and Maria-PauGinebra. 2020. The Devices, Experimental Scaf-folds, and Biomaterials Ontology (DEB): A Toolfor Mapping, Annotation, and Analysis of Bio-materials Data. Advanced Functional Materials,30(16):1909910.

Jelena Jovanovic and Ebrahim Bagheri. 2017. Seman-tic annotation in biomedicine: The current land-scape.

Suwisa Kaewphan, Kai Hakala, Niko Miekka, TapioSalakoski, and Filip Ginter. 2018. Wide-scopebiomedical named entity recognition and normaliza-tion with CRFs, fuzzy matching and character levelmodeling. Database, 2018(2018):96.

Jean Baptiste Lamy. 2017. Owlready: Ontology-oriented programming in Python with automaticclassification and high level constructs for biomed-ical ontologies. Artificial Intelligence in Medicine,80:11–28.

Ki-Joong Lee, Young-Sook Hwang, and Hae-ChangRim. 2003. Two-phase biomedical NE recognitionbased on SVMs.

Christopher Manning, Mihai Surdeanu, John Bauer,Jenny Finkel, Steven Bethard, and David McClosky.2014. The Stanford CoreNLP Natural LanguageProcessing Toolkit. In Proceedings of 52nd An-nual Meeting of the Association for ComputationalLinguistics: System Demonstrations, pages 55–60,Stroudsburg, PA, USA. Association for Computa-tional Linguistics.

Marcos Martínez-Romero, Clement Jonquet, Martin J.O’Connor, John Graybeal, Alejandro Pazos, andMark A. Musen. 2017. NCBO Ontology Recom-mender 2.0: An enhanced approach for biomedicalontology recommendation. Journal of BiomedicalSemantics.

Andrew McCallum and Wei Li. 2003. Early results fornamed entity recognition with conditional randomfields, feature induction and web-enhanced lexicons.In Proceedings of the Seventh Conference on Natu-ral Language Learning at HLT-NAACL 2003, pages188–191.

François Pognan, Thomas Steger-Hartmann, Car-los Díaz, Niklas Blomberg, Frank Bringezu,Katharine Briggs, Giulia Callegaro, SalvadorCapella-Gutierrez, Emilio Centeno, Javier Corvi,Philip Drew, William C. Drewe, José M. Fernán-dez, Laura I. Furlong, Emre Guney, Jan A. Kors,Miguel Angel Mayer, Manuel Pastor, Janet Piñero,Juan Manuel Ramírez-Anguita, Francesco Ronzano,Philip Rowell, Josep Saüch-Pitarch, Alfonso Valen-cia, Bob van de Water, Johan van der Lei, Erik vanMulligen, and Ferran Sanz. 2021. The etransafeproject on translational safety assessment throughintegrative knowledge management: Achievementsand perspectives. Pharmaceuticals, 14(3).

Ariel S. Schwartz and Marti A. Hearst. 2003. A simplealgorithm for identifying abbreviation definitions inbiomedical text. Pacific Symposium on Biocomput-ing. Pacific Symposium on Biocomputing.

Hye Jeong Song, Byeong Cheol Jo, Chan Young Park,Jong Dae Kim, and Yu Seop Kim. 2018. Com-parison of named entity recognition methodologiesin biomedical documents. BioMedical EngineeringOnline, 17(Suppl 2).

Yu Song, Eunju Kim, Gary Geunbae Lee, and Byoung-kee Yi. 2004. POSBIOTM-NER in the shared taskof BioNLP/NLPBA 2004.

Hillel Taub-Tabib, Micah Shlain, Shoval Sadde, DanLahav, Matan Eyal, Yaara Cohen, and Y. Goldberg.2020. Interactive extractive search over biomedicalcorpora. ArXiv, abs/2006.04148.

Dennis G. Thomas, Rohit V. Pappu, and Nathan A.Baker. 2011. NanoParticle Ontology for cancer nan-otechnology research. Journal of Biomedical Infor-matics.

Federica Viti, Silvia Scaglione, Alessandro Orro, andLuciano Milanesi. 2014. Guidelines for managing

45

data and processes in bone and cartilage tissue engi-neering. BMC Bioinformatics, 15(Suppl 1):S14.

Chih Hsuan Wei, Hung Yu Kao, and Zhiyong Lu. 2013.PubTator: a web-based text mining tool for assistingbiocuration. Nucleic acids research, 41(Web Serverissue).

Alexander Yeh, Alexander Morgan, Marc Colosimo,and Lynette Hirschman. 2005. BioCreAtIvE task1A: Gene mention finding evaluation. BMC Bioin-formatics, 6(SUPPL.1):1–10.

Shaojun Zhao. 2004. Named entity recognition inbiomedical texts using an HMM model. (Grefen-stette 1994):84.

46

A Semantic resources

Table 3: List of semantic resources used by the Biomaterials Annotator

Semantic Resource Name Acronym Scope and relevance Type

1 Chemical MethodsOntology CHMO Methods used to collect chemical

experiments data. Ontology

2 Chemical Entities ofBiological Interest CHEBI Compounds of biological relevance,

macromolecules. Ontology

3The Devices, ExperimentalScaffolds and BiomaterialsOntology

DEBBiomaterials-related concepts,materials, structures,material processing.

Ontology

4 EDAM BioimagingOntology

EDAM-BIOIMAGING

Imaging and sample preparationtechniques. Ontology

5 Global Medical DeviceNomenclature GMDN Full names of medical devices. Nomenclature

6 Interlinking ontology ofbiological concepts IOBC

Biological concepts includingbiological phenomena, diseases,molecular functions,research imaging techniques.

Ontology

7 Medical Subject Headings MeSHA hierarchically organizedvocabulary produced bythe NLM.

Controlledvocabulary

8 National Cancer InstituteThesaurus NCIT Vocabulary for clinical care,

translational and basic research.Controlledvocabulary

9 Nanoparticle ontology NPO The description, preparation, andcharacterization of nanomaterials. Ontology

10 Ontology for BiomedicalInvestigations OBI

Biomedical protocols, instruments,data generated, materials, analysisperformed.

Ontology

11 Ontology of NuclearToxicity ONTOTOXNUC Models, chemicals, tools, research

techniques and models. Ontology

12 Precision MedicineOntology PREMEDONTO

Human disease terms, genomic,molecular, phenotype, relatedmedical vocabulary.

Ontology

13 Uber Anatomy Ontology UBERON An integrated cross-speciesanatomy ontology Ontology

47

B Annotation categories

Table 4: Annotation categories, their respective semantic resources and imported classes

Annotationcategory Definition Resource and imported classes

1 BiomaterialA non-drug raw materialor substance suitable forinclusion in systems whichaugment or replace thefunction of bodily tissuesor organs.

DEB: BiomaterialsCHEBI: MacromoleculeMeSH: Biomedical or Dental Material

2 BiomaterialTypes

Classification or natureof biomaterials. DEB: Biomaterial Type

3BiologicallyActiveSubstance

Substance included in amanufactured object inorder to impart a biologicalactivity.

DEB: Biologically Active SubstanceMeSH: amino acid, peptide, proteinBiologically Active SubstancePharmacologic SubstanceNCIT: Protein Domain

4 ManufacturedObject

A physical object createdby hand or machine.

DEB: Manufactured ObjectMeSH: Medical deviceGMDN: Full nomenclature

5ManufacturedObjectComponent

A part, region orcomponent referred toas a distinct unit, suchas a surface or a layer.

DEB: Manufactured Object Component

6 MedicalApplication,Diseaseor condition

Intended use, context,function or outcome ofthe manufactured object.

DEB: Medical ApplicationMeSH: Disease or SyndromeTherapeutic or Preventive ProcedureAnatomical Abnormality

7 ManufacturedObjectFeatures

Characteristics inherentor given during processingto a manufactured object orits components.

DEB: Manufactured Object FeaturesMeSH: Chemical Viewed Structurally

8 Structure

The configuration, formor texture associated witha manufactured objector its components.

DEB: Structure

9AssociatedBiologicalProcess

A cellular or biologicalprocess that themanufactured object isdesigned to causeor support, or is measuredto affect.

DEB: Associated Biological ProcessMeSH: Organ or Tissue FunctionMolecular FunctionCell FunctionBiological functionNCIT: Cellular Process

10 MaterialProcessing

A planned process whichresults in physical changesin a specified inputmaterial.

DEB: Material ProcessingCHMO: Material Processing

11 CellThe reported cell lineor primary celltype.

MeSH: CellNCIT: CellUBERON: Bone cell, cardiocytecirculating cellconnective tissue cellepithelial cell

12 SpeciesThe species and /orbreed used in thestudy.

MeSH: Mammal

13 TissueA tissue or an organmentioned in thestudy as the targetor test system forthe biomaterial objector medical device.

MeSH: Tissue,Body Location or organBody part, organ ororgan componentUBERON: TissuePREMEDONTO:Body Part,Organ, Organ System

48

Table: Continued

Annotationcategory Definition Resource and imported classes

14 AdverseEffects

An unfavourable orunintended disease, sign,or symptom (including anabnormal laboratoryfinding) that is temporallyassociated with the useof a medical device orbiomaterial.

DEB: Adverse EffectsMeSH: Pathologic Function

15 ResearchTechnique

The reported laboratorytechnique or instrumentused in an experimentalstudy.

MeSH: Laboratory Procedure,Molecular Biology Research TechniqueDEB: Research TechniqueNCIT: Research TechniqueNPO: InstrumentIOBC: MicroscopeOBI: AssayEDAM: Imaging,Sample preparationONTOTOXNUC: Outil

16EffectOn BiologicalSystem

The effect associated withmanufactured object ina specific test system(cells, tissue or organism).

DEB: Effect On Biological System

17 StudyType

The study set up,such as in vitro,in vivo, or clinical.

DEB: Study Type