OntoVerbal-M is an ontology verbaliser that transforms OWL into fluent natural language paragraphs in multiple languages. We describe the application of OntoVerbal-M to SNOMED CT, whereby SNOMED CT classes are presented as textual paragraphs in both English and Mandarin through the use of natural language generation. SNOMED CT is a large description logic based terminology for recording in electronic health records. Often, neither the labels nor the description logic definitions in SNOMED CT are easy for users to understand. Furthermore, information is increasingly being recorded, not just using individual SNOMED CT concepts, but using dynamically created description logic expressions (“post-coordinated” concepts). Such postcoordinated expressions can have no pre-assigned labels. In this context automatic verbalisation into multiple languages will be useful both for understanding and quality assurance of SNOMED CT definitions, and for helping different language-speaking-users to understand and share postcoordinated expressions.
We present OntoVerbal-M, a multi-lingual verbaliser for ontologies tailored to be
used with SNOMED CT, a large medical terminology. Such ontologies and
terminologies are increasingly authored in description logics, such as the W3C
recommendation, the Web Ontology Language, OWL [
Verbalising these expressions in natural language is therefore attractive as a means to communicate with users [3; 4; 6]. Verbalisation has the added advantage that it should be possible to re-use some of the same language generation components in the generation of verbalisations in multiple languages.
SNOMED CT [19; 21] (Systematized Nomenclature of Medicine Clinical Terms) is big and potentially widely used OWL based terminology in any field. It attempts to provide a comprehensive terminology for use in medical records across all of medicine, including diseases, diagnoses, procedures, anatomy, microorganisms and pharmaceuticals. It is maintained by the International Health Terminology Standards Development Organisation (IHTSDO)1, and has been mandated or advocated for use in more than 50 countries. Today SNOMED CT is available in US English, UK English and Spanish. Translations to several other languages are currently taking place.
We have taken SNOMED CT as an example to demonstrate our techniques for verbalisation. In its OWL form, SNOMED CT is often awkward and even obscure. For example, the rendering of even just the definition of a simple concept such as heart disease in the raw OWL version of SNOMED CT is several lines long: Class: Heart disease EquivalentTo: Disorder of cardiovascular system and RoleGroup some (Finding site some Heart structure)
By contrast, an English “verbalisation” of this definition in natural language as shown below will be easier for domain experts to understand, although it still seems somewhat stilted:
A heart disease is a disorder of the cardiovascular system that is found in the structure of the heart.
The verbalisation also omits the technically necessary, but to the domain expert mysterious, expression “RoleGroup”.
When we attempt to present, not just the definition, but the information present in
the ontology about a concept – e.g. Heart disease – the OWL expressions become
more complex. Worse, they may not all be located together in the ontology. Hence
the advantage of a verbaliser that presents the entire description of a concept in a
single natural language paragraph, according to the discourse rules expressed in
Rhetorical Structure Theory [
Using Rhetorical Structure Theory, furthermore, gives us a major component that appears to be re-usable across languages. The same mechanisms that produced the English above can generate Mandarin as: heart disease is from heart structure disorder caused cardiovascular system disorder Such verbalisations could be produced manually, but this is time consuming and, as mentioned, not possible for the dynamically created “post-coordinated” expressions for concepts.
OntoVerbal-M provides natural text descriptions with the aim of helping nonontology experts understand the concepts in SNOMED CT. Currently, we have produced an English version using the official SNOMED CT labels and an experimental Mandarin version using ad hoc translations by a native speaker. The Mandarin must be taken with caution, as the translations of the individual labels are ad hoc and the validation has so far been only opportunistic. Nonetheless, the results have been sufficiently well received that we are strongly encouraged to extend the
study to a more formal analysis. In future, we hope to extend this to other languages and to compare verbalisations from OntoVerbal-M with manual translations.
It must be emphasised that OntoVerbal-M is not a machine translation system from one string to another. Rather it generates texts in multiple languages from the same underlying conceptual structure – ultimately a set of expressions in a description logic and the lexicon associated with those concepts in a particular language, as in other multilingual Natural Language Generation Systems [13; 18]. 2.
OntoVerbal-M is an extension of OntoVerbal. OntoVerbal was initially built for
verbalising ontologies into English text [
!"#$%&'()*"+',-.* ?"#%),&'.'()*"+',-.* @&3%&'()*4"#%),&'.%3*"+',-*)&,A0.** /%01,2'()*,&3%&%3*4"#%),&'%.*'(#,*"*#,051%6%1*789*.4$%-"* :()1'.$* <"=%1.>** <%+'4,(* :()1'.$* !%(%&"#,&* :()1'.$* 9%+#* ;"(3"&'(* !%(%&"#,&* ;"(3"&'(* 9%+#*
OntoVerbal-M utilises the intuitive correlation between axioms and sentences to
produce paragraphs that are more than simple collections of individual sentences.
Instead, the sentences are structured and ordered [
OntoVerbal-M currently has two language generators as shown in Fig. 1, but they share a single discourse structure. Each generator has its own input labels as well as its own lexicons. Table 1 shows some examples of class labels in both languages.
Axioms in different notions are also transformed into sentences respectively according to the role of the focused class in the axiom as shown in Table 2. So, for example, a focus class “X” is to be expressed as a sub class of Y in an axiom, then this axiom is to be transformed into English as an X is a kind of Y, and in Mandarin as XY .
There are several natural language (NL) techniques that have been embedded in OntoVerbal-M. The first one is aggregation [8; 16]. For example “an X is a kind of an O”, “an X is a kind of a P” and “an X is a kind of a Q”. The three sentences are aggregated as “an X is a kind of an O, a P and a Q”. The same technique is also applied to Mandarin to have the sentence as “ ”.
The second NL technique used is topic-maintenance-device [
The third NL technique is the use of discourse markers [5; 20]. Discourse markers are applied when a focus class contains several axioms to be verbalised. Using discourse markers ensures the maintenance of fluency and coherence in a paragraph. For example, to connect an additional sentence from the above example, we use “additionally” in English and “” in Mandarin to produce the following paragraphs: “an X is a Y. A more specialised kind of X is Z. Additionally, an X is defined as a P that …”, and “XY Z X P… …”.
The fourth NL technique uses a set of key phrases to signal a change of topic in the generated text. Without such signalling, the text will lack coherence and fluency and be harder to understand. In cases where extra information should be given to the focus class, we introduce “Another relevant aspect of” or “Other relevant aspects of ” as key phrases in English and “...” in mandar in to signal the topic change.
!"#$%&''()*+,-'./0' !"#$$%&'()#$*+"),&-./01*02$)+23& &456!"#$$78%&'9./01*02$):0&;)$+1;013& !"#$$%&'45$*#)20;&;)#$*+"),&-./01*02$)+23& 456!"#$$78%&'()#$*+"),&-./01*02$)+23& !"#$$%&<#6)"0&;)#$*+"),&-./01*02$)+2& 456!"#$$78%&'()#$*+"),&-./01*02$)+23& !"#$$%&'40,+2;#1.&;)#$*+"),&-./01*02$)+23& 456!"#$$78%&'40,+2;#1.&-./01*02$)+23& #2;&'()#$*+"),&-./01*02$)+23& 72*+B016#"C?&
*$%#$%&',"12345' ()#$*+"),&-./01*02$)+2&)$&#&=)2;&+8&-./01*02$):0& ;)$+1;01>&?+10&$/0,)#")$0;&=)2;$&+8&;)#$*+"),& -./01*02$)+2
& • $5$*#)20;&;)#$*+"),&-./01*02$)+2& #2;& • "#6)"0&;)#$*+"),&-./01*02$)+2>& @2+*-01&10"0:#2*&#$/0,*&+8&;)#$*+"),&-./01*02$)+2&)$& *-#*%&$0,+2;#1.&;)#$*+"),&-./01*02$)+2&)$&;0A20;&#$&#& $0,+2;#1.&-./01*02$)+2&#&;)#$*+"),&-./01*02$)+2>& *$%#$%&'+6"7683"'
Our primary goal is to provide text that not only has the structure of SNOMED CT concepts, but also to have them made clear. The textual output of OntoVerbal-M is thus faithful to the ontological input; a more idiomatic verbalisation is not our current goal. We show here some typical output in two languages; the input is the module extracted2 using as a signature Hypertension and all of its inferred subclasses from SNOMED CT full. The underlined words are system-generated words, and the nonunderlined words are formal SNOMED CT terms. The following are five outputs from OntoVerbal-M. Each of them can be a verbalisation from one simple axiom to several complex axioms. (a) Goldblatt hypertension is a kind of renovascular hypertension. (b) Disorder of the pelvis is defined as a disorder of the trunk that has a finding site in the pelvis. (c) The cell is a kind of anatomical structure. More specialised kinds of the cell are • entire cell and
• subcellular structure. (d) Disorder of pregnancy is defined as a finding related to pregnancy and a disease. Another relevant aspect of a disorder of pregnancy is that: complication related to pregnancy is defined as a complication and a disorder of pregnancy. (e) Kidney disease is defined as a disorder of the genitourinary system that has a finding site in a kidney.
Other relevant aspects of a kidney disease include the following: • renal impairment is a kind of kidney disease that has a finding site in a kidney; • uremia is a kind of metabolic disease that is a kidney disease, and has a finding site in a kidney; • renal vascular disorder is defined as a vascular disease of the abdomen that is a kidney disease, and has a finding site in a vessel of kidney; • toxic nephropathy is defined as a kidney disease that has a causative agent in a substance; • renal hypertension is defined as a secondary hypertension that is associated with a kidney disease; • hypertensive renal disease is defined as a hypertensive disorder that is a kidney disease, and has a finding site in a kidney.
Most SNOMED CT class IDs have several associated terms [
The use of SNOMED CT supplied terms directly causes articles to be missed in the generated text, especially when definite articles are needed in the text. For example, if we used only the SNOMED CT supplied terms, we would have just “pelvis” and “trunk” in the output rather than “the pelvis” and “the trunk”. The ontology alone does not provide sufficient information to deal with articles and plurals completely. For example, the use of singular or plurals in anatomy depends on whether the body normally has just one or more than one of a particular kind of part. The naming convention is, however, to take an unadorned singular form, such as “heart”, rather than “the heart”. Therefore, instead of naming a class “the heart structure”, SNOMED CT actually names this class “heart structure”. Our approach on dealing with the definite article and plurals is to look up online resources that can provide examples of usage of articles in human anatomical terms, so that we could replace the SNOMED CT labels with English expressions; Table 3 shows some examples from this approach.
This approach is, however, not perfect, as some terms in SNOMED CT are either missing or in different phrasing order from our looked-up resources. For example, we are able to change “pelvis” to “the pelvis” but unable to change “pelvic structure” to “the pelvic structure”. A further process, such as the use of the Unified Medical Language System (UMLS)3, would be needed to improve this problem to change the adjectival form for this case.
In Mandarin, articles and plurality are not as a concern, as they are in generating English text. The difference between these two languages is that English has count nouns in a singular or plural form, or even to have mass nouns, while Mandarin can only modify a noun by adding an adjective or numeral in front of the noun. For example: “ (your) (heart)”, “ (my) (heart)”, “ (one) (heart)” or “ (two) (heart)”. Wherever the “ ” appears in a text, it never becomes “the ” or “ s”. This means the fixes applied for articles in English NLG of SNOMED CT are not needed in the Mandarin version.
Logical conjunction is one of the important methods used in designing SNOMED CT terms. It reveals the semantic relationship between the clinical medical concepts and their logical triple structure in order to present clinical information. For example: Class: Acute metabolic disorder EquivalentTo: Acute disease and Metabolic disease
The Acute metabolic disorder is an intersection between Acute disease and Metabolic disease. This axiom is transformed into English as “Acute metabolic disorder is defined as both an acute disease and a metabolic disease”, where the intersection in this axiom is not shown directly in the English verbalisation. This verbalisation can be ambiguous for non-native English speakers in understanding that Acute metabolic disorder = Acute disease Metabolic disease rather than Acute metabolic disorder = Acute disease ∩ Metabolic disease.
Why not just simply transformed the above axiom as “Acute metabolic disorder is defined as an intersection between an acute disease and a metabolic disease”? The reason is that the word “intersection” is a mathematical word, which is not commonly used outside Mathematics. Also, translating “intersection” unambiguously into English is awkward at best.
In comparison, the same axiom is transformed into Mandarin as “(Acute)(metabolic) (disorder) (is)(acute) (disease) (a nd)(metabolic) (disease) (apostrophe)(intersection) ”, where (intersection) is the word – intersection, and is just simply to be used to express its role. In fact, the phrase is not awkward in Mandarin’s daily conver sation. For example, if A decides to break a relationship with B. A can say to B “your life has no (intersection) with me”, or A complains about B and says “my conversation with B has no (intersection)”.
Properties are one of the problems in ontology verbalisation due to the lack of
NLG orientated guidelines for labelling properties [
With each different initial morphology, the verbaliser needs a Part Of Speech (POS) checking in order to choose a correct verb for generating a sentence such as using ”is” or “has” or without adding verbs. Our experience has suggested that a standardised term modelling approach will represent ontologies well, and will also save much time when building ontology verbalisers. The quality of the verbalisation will improve dramatically if ontology classes and properties are well phrased or annotated such that the linguistic behaviour of the concept or property is apparent. For example: class labels are noun phrases; property labels start with a third personal singular verb and end with a preposition. In this case, a property would act as a nice predicate between its subject and object classes. This way would free a verbaliser from concerning itself with or without a verb and the text fluency while transforming properties without appropriate prepositions.
In SNOMED CT the property “after” is a good example. If it were lexicalised as “has an after effect in” then the postoperative complication concept can be verbalised as “postoperative complication is defined as a complication of a surgical procedure that has an after affect in a surgical procedure”. In this case the verbaliser only needs to concern itself about the article in this sentence.
The issues on phrasing class and property labels in verbalising English SNOMED CT has led us to be more careful in translating Mandarin labels. We adapt the standardised term modelling approach suggested from English labels to translate Mandarin labels. So every class is translated into a noun phrase, and every property starts with a verb in the Mandarin labels.
Table 5 shows examples of our manually annotated property labels in both English and Mandarin according to the suggested standardised term modelling approach. However, because of the different sentence structure between English and Mandarin, instead of ending each property with a preposition, Mandarin properties are ended with nouns. Therefore each axiom can be transformed to start with a noun subject class, and what ever happens in between, then it ends with a property in a Mandarin sentence.
OntoVerbal-M currently produces well-structured English and Mandarin natural languages for the fragment of SNOMED CT so far studied. Natural language texts are easier to understand than DL based terminologies and ontologies, especially for nonDL users. This motivates the need for automatically generated verbalisations. When users in multiple languages for which there are no full translations want at least limited access to the content, then multilingual generation becomes important.
It is striking that the same rhetorical structure schema appears to be applicable across two such different languages, despite marked differences in grammar and syntax. This significantly reduces the effort required to produce verbalisers in different languages as significant portions of the verbalisation machinery can be reused.
Clearly, there are dangers of erroneous verbalisations, particularly in “new” languages such as Mandarin. Our Mandarin labels were not developed by a team of Mandarin speaking medical experts, but purely from our own team knowledge. At this time they must be regarded as experimental and used for OntoVerbal-M’s purpose only. However, a small survey from Mandarin speaking doctors without SNOMED CT knowledge has indicated that the non-underlined medical terms of OntoVerbal’s output are appropriate to express the meaning of the output texts. The survey also indicates that the underlined words we have chosen are generally suitable for text fluency purposes.
Although OntoVerbal-M’s output needs some linguistic polish, especially in plurality and articles in English, its design in, first: organising information into super, sub and equivalent classes, second: transforming OWL classes and properties into text, and third: the use of discourse structure for generating text, would apply to ontologies from any domain. Specific to SNOMED CT, OntoVerbal-M has annotated classes and properties for text fluency purposes, and particularly to deal with human anatomical phrasing. Its experiences have raised the issues in phrasing ontology terms in English and Mandarin.
In the future, we plan to evaluate the text generated by OntoVerbal-M in two aspects: a) whether the text is faithful to the ontology, so the subjects need to be ontologist to be able to read text and regenerate ontology axioms; b) whether the translation into Mandarin is equally faithful. The participants in evaluation a), and ideally b), need to be SNOMED CT experts so that they understand both the axiomatic descriptions and a natural language text. Failing this, we expect to ask a broad team of domain experts to identify definitions that appear questionable and then consult a more limited team of SNOMED CT experts about those identified as questionable. We will also explore if OntoVerbal-M’s system architecture can adopt more language generators such as French, Spanish and German.
Acknowledgments. This work is part of the Semantic Web Authoring Tool (SWAT) project (see www.swatproject.org), which is supported by the UK Engineering and Physical Sciences Research Council (EPSRC) grant EP/G032459/1, to the University of Manchester, the University of Sussex, and the Open University. We are extremely grateful to Professor Donia Scott for her critical feedback on an earlier draft of this article. The Mandarin survey has been supported by doctors from Chang Gung Medical Foundation LinKo Taiwan. We also thank Dr. Wu from Kaohsiung to give his useful comments on the medical terms translation.