There are multiple ontology matching approaches that use domain-speci c background knowledge to match labels in domain ontologies or classi cations. However, they tend to rely on lexical knowledge and do not consider the speci cities of domain grammar. In this paper, we demonstrate the usefulness of both lexical and grammatical linguistic domain knowledge for ontology matching through examples from multiple domains. We also provide an evaluation of the impact of such knowledge on a real-world problem of matching classi cations of mental illnesses from the health domain. Our experimentation with two matcher tools that use very di erent matching mechanisms|LogMap and SMATCH|shows that both lexical and grammatical knowledge improve matching results.
Ontology Matching (OM) aims at nding correspondences between the classes
and instances of multiple ontologies [
A large number of matchers analyse natural language labels, on di erent
levels of complexity. A common approach is to incorporate linguistic background
knowledge (BK) into the matcher [
SMATCH [
All of these matchers, however, are limited to using lexical BK. While some of them do address grammar through basic domain-independent methods (tokenisation, lemmatisation, stop word elimination), they do not cope with cases where the grammar depends on the domain.
In this paper, we investigate the impact of both domain lexicon and domain grammar in ontology matching, mainly focusing on label-based matching.
The paper is organised as follows. In Section 2 we describe how domain knowledge appears in ontology labels. Section 3 focusses on di erent approaches that matchers may use to take advantage of domain knowledge. A case study on the health domain is presented in section 4, being evaluated in section 5. The paper nishes with some concluding remarks and future works included in section 6. 2
The use of specialised linguistic constructs is common in most domains of knowledge. The most obvious case is the use of specialised terminology, consisting of words and expressions that either are used exclusively within the context of a domain (such as to deglaze in cooking meaning `to loosen bits of food which stuck on the bottom of a pan by adding liquid' ), or that gain a new meaning within a domain (such as to clarify which in cooking refers speci cally to butter). Domain-speci c meaning can, however, also be vehicled by non-lexical means, a phenomenon that we globally call domain grammar. Domain grammar can be found even within the short labels typical of ontologies and classi cations. Below we provide examples of domain language from specialised text, including ontology labels.
Domain terms. The UK Civil & Protection Lexicon (UKCP) de nes the term medevac that means medical evacuation, itself considered a specialised term. In order to align these two terms, a matcher would either need lexical background knowledge that states their synonymy, or|in this speci c case|word-level analysis in order to detect that one term is the abbreviated form of the other. Domain acronyms. The acronym REM has many meanings; in the domain of neurology it means rapid eye movement. Again, in a matching task the acronym can be matched either through the use of domain lexical knowledge or through acronym detection.
Word derivation. Derivation rules allow the creation of words through the use of a xes, such as voyeur 7! voyeurism or anorexia 7! anorexic. While, as in these cases, domain language often relies on the derivational rules of general grammar, domain-speci c derivational a xes and rules also exist, such as candida 7! candidiasis in the medical domain. Even though the common approach in lexicography is to enumerate derived words as separate lexical entries, lexicons are often incomplete in practice due to the high productivity of a xes. Thus, grammar-based approaches to detecting the relatedness of derived terms can be useful, as when matching the label fetishism with fetishistic disorder.
Word in ection. In ection rules are de ned by general language; yet, particular in ected forms can be more or less speci c to domains. A well-known example are cooking recipes where sentences tend to begin with verbs either in in nitive or imperative form (e.g. `Peel the onions', which in French may be expressed either as `Peler les ognons' or as `Pelez les ognons' ).
Speci c uses of punctuation. In labels of the International Classi cation of Diseases (ICD), such as `Hallucinogen use, unspeci ed with hallucinogen persisting perception disorder ( ashbacks)', parentheses are used to provide clues for the interpretation of the label. Square brackets, commas, or parentheses are also widely used in ontologies, classi cations, and data schemas, such as to provide units of measure for numerical values: speed (km/h). The precise interpretation (e.g. relevance or not with respect to the matching task) of such punctuation and the text they delimit depends on the domain and the particular application at hand.
Domain syntax. The same `Hallucinogen use. . . ' example from above shows that labels can use non-standard syntax. This is sometimes motivated by the context of use, such as the need to sort the labels alphabetically motivates the use of the adjective unspeci ed in a postpositive form. The phrase hallucinogen persisting perception disorder, on the other hand, includes syntax that is not considered as standard in general language but is common in medical text. While syntax may play a minor role in matching very short labels, for longer classi cation entries it may be taken into account by the matcher tool, as in the case of SMATCH that performs syntactic parsing. 3
The hypothesis veri ed in this paper is that \matching performance can be improved by relying on knowledge that is speci c to domain language ". However, as domain language also incorporates elements of general language, our study also considers this aspect. Accordingly, we classify linguistic background knowledge with respect to being general or domain-speci c, as well as with respect to being lexical or grammatical. This delineates the following four categories of knowledge: (1) general lexicon; (2) general grammar; (3) domain lexicon; and (4) domain grammar. Furthermore, we consider three di erent forms of grammatical knowledge with respect to the linguistic elements to which they apply: (a) phrase-level (syntax, dealing with the way words are organised within labels); (b) word-level (morphology, i.e. grammar that deals with the structure of words); and (c) character-level (e.g. orthography and use of punctuation). Due to the shortness of ontology and classi cation labels, we deem it su cient to consider only these three levels of granularity of grammar.
General Lexicon A domain-independent resource that is commonly used is
Princeton WordNet [
Domain Lexicon There are multiple domain-speci c resources such as lexicons
or domain terminologies that contain the technical terms of an speci c domain.
In the literature, we can nd di erent approaches to integrating these resources
within WordNet [
The main advantage of using domain lexical knowledge is that matchers have
an enriched BK and are able to nd mappings of labels that include some of the
added new terms. Moreover, when matching ontologies from multiple or partially
di erent domains (such as reference health knowledge involving subdomains of
healthcare), domain information can be leveraged for word sense disambiguation
within the matching process, resulting in improved precision [
Tokenisation. Labels are segmented into tokens (e.g. \medium-scale evacuation" becomes <medium, scale, evacuation>).
Acronym extraction. Characters of tokens are used to extract/discover acronyms (e.g. \Non Governmental Organisation " becomes \NGO "). String similarity. Compare string labels considering di erent measures and return a value according to their similarity degree (e.g. \Level of emergency " and \Level 1 emergency " have a high similarity degree). Stopword elimination. Tokens that are recognised as articles, prepositions, conjunctions are removed (e.g. \level of emergency " becomes \level emergency "). { Word Level Grammar .
Lemmatisation. Tokens are reduced to basic forms (e.g. \disasters " becomes \disaster "). { Character Level Grammar .
Normalisation. This task includes several subtasks such as: case normalisation, diacritics suppression, blank normalisation, digit suppression or punctuation elimination.
Domain Grammar There are cases in which applying the previous domain-independent tasks to domain-speci c resources is counter-productive. For example, if we apply digit suppression and stopword elimination to the following labels: \Level of emergency ", \Level 1 emergency ", \Level 2 emergency ", \Level 3 emergency "; the matcher might output that all labels represent the same knowledge. Another example appears when the case normalisation task is just limited to transform all characters within the label into lower case letters. In this case, if the label contains Roman numerals they might pass unnoticed after the case normalisation. For these reasons, it is necessary to consider domain-speci c grammar and address it conscientiously. Below there are described the approaches that we have implemented in our research: { Phrase Level Grammar . Finding clues or postscripts that recurrently appear within the labels in a domain is not unusual. In this case, it is necessary to analyse if they add enough knowledge to keep them in the label or it is hh((( worth suppressing them (e.g. \Mild cognitive impairment(,s(oshtahtehd"). { Word Level Grammar . Implementing derivational morphology rules to transform a term from one part-of-speech into another is interesting because enriching matchers' BK with these words allows those matchers that do not mainly base the matching process on string similarity measures to discover new mappings. Domain words produced by derivational morphology are added to matchers' BK as related forms (e.g. \pathological " is added as a related form of \pathology ). { Character Level Grammar . Depending on the domain, particularly in application domain knowledge resources, orthography follows di erent conventions. This makes necessary to address it optimally in each case. For example, there might be cases in which the content within parentheses or square brackets is meta-information that is not relevant for the meaning of the label (e.g. \Post(-) traumatic stress disorder "), being recommendable its suppression, whereas in other cases this content might be essential (e.g. \Stable lodine (Potassium lodate tables)").
The rules of the di erent domain grammar levels can be extracted both in a supervised or unsupervised way. The latter requires a huge number of documents to apply statistical methods, whereas the former does not need such quantity of documents, but involves more e ort. In general, the rules at the word level can be transferred to any ontology within a domain (e.g. health), while the rules at the phrase and character levels usually are more dependent on the application domain (e.g. Hospitals of North London). 4
The main motivation lies in the need of solving semantic interoperability problems within the health domain. For example, when clinicians have to exchange health records that contain descriptions from multiple o cial classi cations of diseases. To do so, we have developed several extensions to enrich the matcher's BK with health lexical and grammatical knowledge.
Due to descriptions of disorders containing not only technical, but also
general terms, WordNet has been used as a domain-independent BK into which
the extensions are plugged. The extensions have been developed following the
Lexical Markup Framework (LMF) standard [
General Lexicon Princeton WordNet has been used as domain-independent resource. The main reason is that it represents general knowledge and there are multiple approaches that we could apply to enrich WordNet with domain-speci c knowledge resources.
Health Domain Lexicon We have developed an extension for WordNet that
includes health lexical knowledge extracted from the following resources:
{ MeSH is the National Library of Medicine's controlled vocabulary thesaurus
[
General Grammar It has been addressed applying the grammatical techniques included in the matchers by default and including general derivational morphology.
Phrase level Grammar. The tasks applied have been: tokenisation, string similarity and stop word elimination.
Word level Grammar. In this case we applied lemmatisation and the integration of general derivational morphology rules included in SPECIALIST. Table 1 shows examples of these rules. Character level Grammar. The tasks applied have been case normalisation, blank normalisation and diacritics suppression.
Health Domain Grammar It has been addressed using health derivational morphology extracted from SPECIALIST, and considerations identi ed at phrase and character grammar levels. The former was used to enrich matchers' BK, whereas the latter were considered as a preprocessing step prior to the OM process.
Phrase level Grammar. In medical resources there are clues that recurrently appear within descriptions of disorders. Examples are \, unde ned " and \, so stated ". This meta-information does not add special value to labels, particularly a ecting to those matchers that mainly use string similarity measures. The main reason is that they are penalised by irrelevant characters, which results in a lower similarity degree. Considering the previous issue we decided to suppress these interpretational clues from descriptions of diseases in a preprocessing step prior to the matching process.
Word level Grammar. Several domain-speci c derivational morphology rules have been extracted from the SPECIALIST lexicon and integrated into WordNet. Examples of these rules are shown in table 2.
Derivational rule Example ose$(verb)! osis$(noun) sclerose(verb) ! sclerosis(noun) physeal$(adj) ! physis$(noun) adenohypophyseal(adj) ! adenohypophysis(noun) sis$(noun)! ze$(verb) dialysis(noun) ! dialyze(verb) a$(noun)! iasis$(noun) candida(noun) ! candidiasis(noun) Character level Grammar. We have identi ed a particular use of parentheses, square brackets and commas in the health domain. Examples of the use of parentheses and square brackets might be the following:
2. No Diagnosis or Condition on Axis I / No Diagnosis on Axis II [DSM-IV] 3. Premature (early) ejaculation 4. Trichotillomania (hair-pulling disorder) 5. Obstructive sleep apnea (adult) (pediatric) In case 1, the square brackets are used to specify an equivalent expression of \sleep terrors ". Similarly, in case 3 parentheses are used to indicate a synonym of \premature". Case 2 is di erent as brackets are used to point out the DSM version in which the description was included. In case 4 the content within parentheses categorises the kind of disorder that \trichotillomania" is. Finally, case 5 uses parentheses to indicate the domain to which the disorder is applicable, in that case to adults and children.
Similarly as in the previous cases, commas are utilised with di erent purposes in the medical knowledge. Below there are some examples:
2. Adverse e ect of unspeci ed antidepressants, sequela 3. Circadian rhythm sleep disorder, shift work
In example 1, the comma is used to specify the degree of the disorder, whereas in example 2, it is used to de ne the kind of adverse e ect. Finally, in example 3, the comma is used to specify the cause of the disorder.
This diverse use of parentheses, square brackets and commas, complicates labels, penalising matchers' performance. Thus, we decided to suppress commas and all content within parentheses and square brackets to avoid this penalisation. This simpli es labels and reduces irrelevant content. Nonetheless, in the future, we should investigate less aggressive solutions to reduce matchers penalisation while taking advantage of the content within parentheses.
The hypothesis has been evaluated by an experiment in which matchers with di erent con gurations had to match several descriptions of the two most important classi cations of diseases for mental health: the Diagnostic and Statistical Manual of Mental Disorders, fth edition (DSM-5) and the ICD-10. To evaluate the quality of the matchers, we used as gold standard the correspondences between both classi cations published in DSM-5, where it is speci ed to which code in ICD-10 corresponds each description in DSM-5.
The input schemas were a source dataset with 200 entries randomly selected from DSM-5, and a target dataset with 177 descriptions included in ICD-10, which are the correspondences of the entries chosen from DSM-5.
The matchers selected were S-Match [
The experiments were executed 4 times with each matcher, computing the standard metrics within the information retrieval community: precision, recall and f-measure. Firstly, with the vanilla version, which was our baseline in each case; secondly, with the lexicon extension; thirdly, with the grammar extension, and nally, with both extensions.
Figure 1 and gure 2 depict the results of the experiments executed in S-Match and LogMap, respectively. We can see how both matchers, S-Match and LogMap, improve their performance in terms of f-measure around 20% and 7% respectively. It is also noticeable, that overall both matchers achieve low results which are caused by the nature of the input labels, which on average are descriptions with more than 5-6 words, so this results in complex label formulas and low string similarity values.
Regarding S-Match, the vanilla version only has a general BK and the matcher is penalised mainly for the way in which it manages commas (each comma is considered as a disjunctive operator). This caused a huge number of false positives, which negatively a ected precision, but also discovered, as side e ect, a high number of correspondences, resulting in the highest recall. An example is the label \Mild cognitive impairment, so stated " which is transformed into the following label formula:
mild & cognitive state & impairment j state From this label formula S-Match computes the following node formula:
(mild j state) & (cognitive state j state) & (impairment j state) That means that if \state" has a relationship with a lemma within any label of the other ontology, the matcher will output a mapping even if the rest of the label is not related.
The lexicon extension considerably improves the performance (11%) by adding health lexicon knowledge, but this extension also avoids some of the correspondences discovered as side e ect, mainly with the inclusion of lexical entries that were considered as single tokens in the vanilla version and now are compound tokens, so the recall slightly decreases.
The grammar extension is the one that drastically reduces the number of false positives mainly with the techniques applied at phrase and character grammar levels that were employed as a preprocessing step prior to the matching process. In addition, it also discovers new mappings thank to the derivational morphology implemented at word grammar level.
The combination of both lexicon and grammar extensions is the con guration that performs better in terms of f-measure, complementing each other and improving the baseline around 20%. However, the false positives of both extensions are also aggregated, being precision slightly penalised.
As for LogMap (see Figure 2), the vanilla version includes biomedical knowledge by default, resulting in a baseline with a performance over 60%.
The lexicon extension added knowledge coming from SPECIALIST, MeSH and WordNet, but it was the latter which produced the major impact as it added domain-independent knowledge contained in the labels. This new knowledge also produced some false positives, but on average this con guration improved the baseline around 6.3%. An example of false positive is:\Narcolepsy without cataplexy but with hypocretin de ciency " \Narcolepsy with cataplexy ", while an example of new true positive is: \Acute stress disorder " \Acute stress reaction".
The grammar extension had a similar e ect mainly because it also incorporated WordNet. In this case, tasks for word and character grammar levels had a low impact on LogMap. Nonetheless, phrase level grammar preprocessing had a signi cant impact, and the performance improved 6.5% with respect to the baseline. Examples of new true positives are: \Trichotillomania (hair-pulling disorder)" \Trichotillomania", and \Overweight or obesity " \Obesity, unspeci ed ".
The combination of both extensions was the con guration that obtained the best performance, achieving the highest number of true positives discovered. In this case, the baseline is improved more than 7%. 6
In this paper, we have presented an approach in which matchers can take advantage of both, domain lexicon and grammar to improve their performance when matching domain-knowledge resources. After evaluating our approach by matching some descriptions of mental health disorders included in DSM-5 and ICD-10 with S-Match and LogMap, we can conclude that our hypothesis is true, as both matchers improve their f-measure compared with the vanilla version.
It is interesting to highlight how the use of domain lexicon and grammar a ects di erently depending on the matcher. Whereas the domain lexicon extension has the major impact on LogMap, S-Match experiences its major improvement with the grammar extension. The main reason is that LogMap now can discover new mappings thank to domain-independent knowledge, and S-Match has label formulas signi cantly simpli ed. This information is useful in order to optimise e orts in the future, and help to decide whether is more valuable investing time focusing on integrating domain lexicon or grammar knowledge into matcher's KB.
As future work we should explore other factors that may a ect matchers when matching domain-knowledge, such as the impact of each kind of knowledge represented within knowledge resources according to their levels of speci city. Moreover, it is interesting to delve into methods to aggregate lexicon and grammar results in order to optimise matcher's performance.
Acknowledgements This research was partially supported by the European with the grant agreement No. 607062 (ESSENCE Marie http://www.essence-network.com/).