Information retrieval (IR) implements the representation, storage and organisation of textual data to enable a user to access relevant pieces of information 33. Biomedical experts regularly exploit IR to locate relevant information (most often in the form of scientific publications) on the Internet. Apart from general-purpose search engines such as Google™ 34, many IR systems have been designed specifically to query databases of biomedical publications (e.g. 3536373839) such as Medical Literature Analysis and Retrieval System Online (MEDLINE) 40 and PubMed Central (PMC) 41 (henceforth referred to together as PubMed), which provide peer-reviewed literature and make it freely accessible in a uniform format. MEDLINE distributes abstracts only, while PMC provides full-text articles. PubMed is accessible through Entrez 42, an integrated retrieval system that provides access to a family of related biomedical databases maintained by the National Center for Biotechnology Information (NCBI).

Documents available in PubMed are indexed by Medical Subject Headings (MeSH) 43 terms (index terms are pre-selected to refer to the content of a document 33). MeSH is a CV consisting of hierarchically organised terms that serve as descriptors to index and annotate documents. This permits direct access to relevant documents at various levels of specificity, thus improving the performance of IR in terms of speed as well as precision and recall. Entrez uses automatic term mapping to match terms against the MeSH hierarchy and to expand a query with (near-)synonyms and subsumed terms. For example, all of the following terms are explicitly listed as terms matching Magnetic Resonance Spectroscopy in MeSH:

• In Vivo NMR Spectroscopy

• Magnetic Resonance

• MR Spectroscopy

• NMR Spectroscopy

• NMR Spectroscopy, In Vivo

• Nuclear Magnetic Resonance

• Spectroscopy, Magnetic Resonance

• Spectroscopy, NMR

• Spectroscopy, Nuclear Magnetic Resonance

Similarly, a query searching for information on Gas Chromatography can be expanded automatically to include Gas Chromatography-Mass Spectrometry as a more specific term (see figure 2).

While the use of the MeSH for indexing and query expansion in Entrez is undoubtedly useful, these benefits cannot be fully exploited for the particular problem of accessing articles describing research that utilizes some analytical technology. In particular, an analytical technique employed in metabolomics is unlikely to be the main focus of the reported studies. Consequently, the corresponding documents may not necessarily be indexed with technology-related MeSH terms. Further, the abstracts of such articles are more likely to report the actual findings rather than the technology-specific experimental conditions applied. These parameters are usually described in the Materials and methods section or as part of the supplementary material. Hence, two points arise when retrieving documents containing information pertinent for analytical techniques deployed in metabolomics studies. First, it is important to search full-text articles as opposed to abstracts only. For this reason we used PMC, which provides access to full-text articles, in addition to MEDLINE, which offers only abstracts. Second, it is necessary to go beyond MeSH terms in query formulation. This problem is alleviated using the following assumption: terms denoting related concepts tend to co-occur within textual documents 4445. On this basis, terms from an initially compiled CV can be combined in a search query to retrieve additional documents that describe research that utilises a technology, i.e. the ones that do not necessarily deal with the technology per se and thus may not be indexed by technology-related MeSH terms. To achieve this, we index the literature with the CV terms. Each CV term is used to search the literature via Entrez. As a result, each term is mapped to a set of documents it matches. This information is stored in a local database using the following structure described in SQL:

CREATE TABLE index

(

term VARCHAR(200) NOT NULL,

document VARCHAR(50) NOT NULL

);

A cut-off point (this is a configurable parameter; the specific values used in our case studies are reported in the Results & Discussion section) is set to remove the non-discriminatory terms, i.e. the ones that return too many documents. These are likely to be broad terms not limited to a specific analytical technique, and consequently introducing unwanted noise in the context of the domain-specific corpus. For example, in the case of the NMR CV, the mean number of abstracts returned was 2,772 with the median being just 0, which is due to the fact that the NMR CV was constructed using a considerable number of terms coming from database schemata. These terms are semi-formal in the sense that they do not necessarily reflect the terminology used in the literature, e.g. AMIX VIEWER & AMIX-TOOLS and JEOL NMR instrument. On the other extreme, terms returning the maximal number of abstracts (set to 50,000) were: analysis, characteristic, concentration, Delta, instrument, method, reference, software, states and tube. The following SQL query can be used to identify such terms:

SELECT term, COUNT(document) AS matching_documents

FROM index

GROUP BY term

WHERE matching_documents >= D;

where D is chosen a cut-off point. Having removed such terms from further consideration from the IR point of view, a cut-off point (as before, this is a configurable parameter, and the specific values used in our case studies are reported in the Results & Discussion section) is set to remove the documents that do not contain a sufficient number of the CV terms. The following SQL query can be used to identify such documents:

SELECT document, COUNT(term) AS matching_terms

FROM index

GROUP BY document

WHERE matching_terms <= T;

where T is chosen a cut-off point. For example, some of the documents with the highest number of matching terms from the NMR CV were 464748.

The IR module based on the methods described above is encoded in Java. The Java application takes advantage of E-Utilities 42, a web service which enables the users to run Entrez queries and download data using their own applications. The information gathered about terms, documents and their relations is stored in a local database (DB) hosted on a PostgreSQL 49 system. By storing the mappings between terms and documents, the querying ability of the DB management system can be combined with that of Entrez. The local DB is also accessible via Java applications (using the JDBC protocol – a standard SQL DB access interface). Hence, all our implemented IR modules can be incorporated into customised workflows 50.

In the literature dealing with terminology issues, a term is intuitively defined as a phrase (typically a noun phrase 751): (1) frequently occurring in texts restricted to a specific domain, and (2) having a special meaning in the given domain 52. Bearing in mind the potentially unlimited number of different domains and the dynamic nature of newly emerging ones (many of which expand rapidly together with the corresponding terminologies, as is the case in metabolomics), the need for efficient term recognition becomes apparent. Manual term recognition approaches are time-consuming, labour-intensive and prone to error due to subjective judgement. These shortcomings can be addressed by automatic term recognition (ATR), the process of annotating an electronic document with a set of terms extracted from the document 53. Here, we emphasise that ATR refers to the computer-based extraction of terms from a domain-specific corpus as opposed to merely matching the corpus against a dictionary of terms 54. It has been suggested that scientific corpora can be used as reliable sources for terminology construction exploiting 8:

• the growing number of electronic corpora,

• efficient NLP tools (such as part-of-speech taggers, parsers, etc.),

• linguistically and/or statistically based ATR procedures, and

• the fact that domain experts often use terms that have not been standardised, and as such are not included into standardised dictionaries.

The lack of terminological standards is especially apparent in the rapidly expanding domain of metabolomics, where there is no exact consensus on what constitutes a metabolite name although naming conventions do exist for some entities, e.g. the Chemical Entities of Biological Interest (ChEBI) dictionary that is emerging for small molecules 55. Still, these are only guidelines and as such do not impose restrictions on domain experts.

Manual term recognition is performed by relying on conceptual knowledge, i.e. humans identify terms by relating them to the corresponding concepts. It is currently not feasible to implement an ATR approach following such a paradigm due to the lack of appropriate knowledge representation systems and the difficulty of automatically performing “intelligent” tasks. For these reasons, ATR approaches resort to other types of knowledge that can provide clues about the terminological status of a given natural language clause 56. Generally, the knowledge used for ATR may involve two types of information:

• internal: morphological, syntactic, semantic and/or statistical knowledge about terms and/or their constituents (nested terms, words, morphemes), and

• external: linguistic and/or statistical knowledge regarding the term context, together with the knowledge contained in external resources, such as electronic dictionaries, ontologies, corpora, etc.

ATR methods typically combine two approaches: linguistic (or symbolic) and statistical (or numeric) 51. Linguistic approaches to ATR usually involve pattern matching to recognise candidate terms by checking if their internal structure conforms to a predefined set of morpho-syntactic rules. Statistical methods rely on at least one of the following hypotheses regarding the term usage 7:

• specificity: terms are likely to be confined to a single or few domains,

• absolute frequency: terms tend to appear frequently in their domain, and

• relative frequency: terms tend to appear more frequently in their domain than in general.

Statistical approaches are prone to extracting not only terms, but also other types of collocations (sequences of words co-occurring more frequently than would be expected by chance) 57: functional, semantic, thematic and others, e.g. “…to play an important role in…”. This problem is typically remedied by employing linguistic filters to extract candidate terms from a corpus, which are then ranked using statistical methods.

In this work, we utilised the C-value method 58, publicly accessible at 59 to the TM community via a web service. It first applies syntactic pattern matching to select term candidates, e.g. noun phrases having the structure described by the following regular expression:

( A D J | N ) + |   ( ( A D J | N ) *   [ N   P R E P ]   ( A D J | N ) * )   N MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXafv3ySLgzGmvETj2BSbqeeuuDJXwAKbsr4rNCHbGeaGqipu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=xfr=xb9adbaqaaeGaciGaaiaabeqaaeaabaWaaaGcbaWaaeWaaeaacqWGbbqqcqWGebarcqWGkbGscqGG8baFcqWGobGtaiaawIcacaGLPaaadaahaaWcbeqaaiabgUcaRaaakiabcYha8jabbccaGmaabmaabaWaaeWaaeaacqWGbbqqcqWGebarcqWGkbGscqGG8baFcqWGobGtaiaawIcacaGLPaaadaahaaWcbeqaaiabcQcaQaaakiabbccaGmaadmaabaGaemOta4KaeeiiaaIaemiuaaLaemOuaiLaemyrauKaemiuaafacaGLBbGaayzxaaGaeeiiaaYaaeWaaeaacqWGbbqqcqWGebarcqWGkbGscqGG8baFcqWGobGtaiaawIcacaGLPaaadaahaaWcbeqaaiabcQcaQaaaaOGaayjkaiaawMcaaiabbccaGiabd6eaobaa@592C@

where ADJ, N and PREP denote adjective, noun and preposition respectively. The C-value of each candidate term t is then calculated as:

C − v a l u e ( t ) = { ln ⁡ | t | ⋅ f ( t ) , i f   S ( t ) = ∅ ln ⁡ | t | ⋅ ( f ( t ) − 1 | S ( t ) | ∑ s ∈ S ( t ) f ( s ) ) , i f   S ( t ) ≠ ∅ MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXafv3ySLgzGmvETj2BSbqeeuuDJXwAKbsr4rNCHbGeaGqipu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=xfr=xb9adbaqaaeGaciGaaiaabeqaaeaabaWaaaGcbaGaem4qamKaeyOeI0IaemODayNaemyyaeMaemiBaWMaemyDauNaemyzau2aaeWaaeaacqWG0baDaiaawIcacaGLPaaacqGH9aqpdaGabaqaauaabaqaciaaaeaacyGGSbaBcqGGUbGBcqGG8baFcqWG0baDcqGG8baFcqGHflY1cqWGMbGzdaqadaqaaiabdsha0bGaayjkaiaawMcaaaqaaiabcYcaSGqaaiab=LgaPjab=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@8897@

where |t| is the length of t in words, f(t) is t's frequency of occurrence and S(t) is the set of other term candidates containing t as a sub-phrase. All candidates whose C-value exceeds a certain threshold are proposed as domain-specific terms by this method. The threshold chosen will affect the performance of ATR in terms of precision and recall, which are calculated as P = A / (A + B) and R = A / (A + C), where A is the number of true positives (correctly recognised terms), B is the number of false positives (phrases incorrectly recognised as terms) and C is the number of false negatives (non-recognised terms). Higher thresholds will typically result in higher precision and lower recall, and vice versa, lower thresholds will increase the recall at the expense of precision. In general, a threshold used should be corpus-specific (e.g. the average C-value found in the given corpus), as the C-value of each term candidate also depends on the corpus.

By its definition, the C-value method favours longer and more frequent phrases that are not typically nested within a relatively small set of other phrases. Obviously, the C-value method relies primarily on the frequency of term usage and their general syntactic properties rather than exploiting orthographic, morphological and lexical features of specific named entities. For example, while protein names may vary significantly between authors, some general characteristics still apply 6061:

• distinctive orthographic characteristics of protein names such as capital letters, digits, special characters (e.g. p

54

SAP

• keywords (e.g. protein, receptor, etc.) describing the protein function in multi-word protein names (e.g. Ras GTPase-activating

protein

• morphological principles for naming proteins, such as highly abundant affixes -ase, -in, etc. (e.g. hexokin

ase

Opting for a similar named entity recognition approach would significantly increase the time and cost of developing CV term acquisition methods, as these would have to be re-implemented for specific domains. Moreover, the type of terms sought may not necessarily exhibit sufficiently discriminatory textual properties 32.

On the other hand, a generic ATR approach (such as the C-value method) can be manipulated to extract terms that are more likely to be of the required type by targeting only relevant documents, and within them specific sections potentially dense with terms of the given type. This can be followed by additional filtering of terms, known to be of different and not directly relevant semantic types to the ones needed, by using lexical resources of these terms where such resources exist. This issue of ATR targeting only relevant documents has been addressed by the IR module described in the previous section. A domain-specific corpora is produced as a result of IR by using either MeSH or CV terms in the search queries over collections of either abstracts or full-text articles in PubMed.

Further, it is particularly important to target only sections that are likely to contain terms relevant for an analytical technology as a preparation step for ATR in order to increase its precision. Therefore, when using full-text documents we reduce them to the Materials and Methods sections, which are recognised automatically utilising PMC's XML format in which articles are distributed. Once a domain-specific corpus is obtained, the C-value terms are extracted and further inspected to see if they include any terms known to belong to other sub-domains not directly related to the analytical technology under investigation, in which case they can be safely filtered out.

Given the initially compiled CVs for NMR and GC, we automatically obtained terms loosely related to these two analytical techniques by applying IR to compile a technology-specific corpus, followed by ATR to extract a list of terms from the corpus in a way described in the preceding sub-sections. Manual inspection of the extracted terms revealed typical types of terms frequently co-occurring with the NMR- and GC-specific terms, namely those denoting substances, organisms, organs, conditions/diseases, etc., which are not of direct interest for the analytical technology per se. Examples of such terms automatically extracted by the C-value method are: amino acid, linseed oil, pancreatic juice, blood glucose, cell wall, Halophilic bacterium, Streptomyces antibioticus, systemic hypertension, cervical dislocation, etc. Unlike analytical techniques, many of which are relatively recent, some of these terminologies are relatively stable with respect to the number of new terms being introduced, e.g. Linnaean taxonomy 62 classifies living organisms in a systematic manner.

The Unified Medical Language System 63 is a multi-purpose resource merging information from over 100 biomedical source vocabularies developed for different purposes. By providing uniform access (including a web service) to terms belonging to various sub-domains of interest, UMLS aims to facilitate the development of information systems for text processing in biomedicine via a semi-formal representation of domain-specific knowledge in order to process, retrieve, integrate, and aggregate biomedical data and information contained in the relevant literature 64. It currently contains 1.4 million concepts named by 7.2 million terms, organised into a hierarchy of 135 semantic types and interconnected by 54 different relations.

The following semantic types in the UMLS proved relevant to our problem of detecting technique-specific terms in a subtractive approach: Organism, Anatomical Structure, Substance, Biological Function and Injury or Poisoning. Given these semantic types as part of the input to the term filtering module (implemented as a Java application), the subsumed terms are automatically selected from the latest version of the UMLS thesaurus. Then, a simple pattern matching approach is applied to filter out these terms and their variations. For example, the filtering approach helped identify the following “outliers” amongst terms extracted by the C-value method: experimental

rat

We have described an integrative approach combining relatively generic software (e.g. Entrez for IR, C-value for ATR) and data resources (e.g. UMLS as a semantic network of biomedical terms) for the rapid development of a TM tool for automatic expansion of CVs as a practical alternative to tailor-made named entity recognition methods (see discussion above). An HTML report is generated as a result of the automated CV expansion (see Figure 3 for an example report generated for the NMR CV). The report summarises the output of each module described earlier, i.e.:

• the number of documents collected by the IR module with a link to the list of their citation details (see Figure 4) and cross-references to the actual documents in PubMed (see Figure 5)

• the size of the final text corpus with a link to the corresponding ASCII file (see Figure 6), and

• the number of new terms extracted by ATR with a link to the list of terms sorted by their C-values.

Terms extracted from four different corpora are also amalgamated into a single, alphabetically ordered list (see Figure 7, left-hand side window). To aid the curation of automatically extracted terms and their incorporation into the CV, the context of a term can be obtained on-the-fly. The context should help the curator interpret the intended meaning of a term and provide clues useful for generating its textual definition. The context of a term rather than its definition may be more crucial for the association of a term with its correct meaning 65. Terms sharing the same context are likely to have similar (or even the same) meaning 66. Conversely, different contexts of the same term may point to the problem of term ambiguity (the same term denoting different concepts). Less drastically, the context may “deviate” the meaning of a term by emphasising only certain aspects of a term (e.g. insulin can be interpreted as both hormone and pharmacological substance). Bearing in mind the importance of contextual information in determining the correct meaning of a term and hence its position in a CV, we deployed a practical solution: all new terms reported are linked to MedEvi 67, a service providing local context (extracted from MEDLINE) for query terms 68. Clicking on a term launches a query to MedEvi, which in turn returns the aligned concordance (words used in a context) lines together with some handy features such as lists of co-occurring keywords and terms (see Figure 7, right-hand side window).