There is this phenomenon called the uncanny valley where, in situations like CGI, robotics etc., if as the human-like thing gets closer and closer to being human, but not quite human, then the human observer is sort of weirded or creeped out. If the human-like thing obviously isn’t human, say a simple cartoon, all is OK, but if it is almost human then people really don’t like it.
In our recent work on natural language generation (NLG) from OWL I’ve noticed a related phenomenon; readers aren’t weirded out by the generated and somewhat clunky English, but are irritated or piqued by it in a way they wouldn’t be by, for instance, some Manchester Syntax for the same axioms or some hand-crafted, but not perfect, text. The Manchester Syntax is further away from “naturalness”, but apparently less irritating – perhaps because the expectations are less. Manchester syntax for some axioms is easy enough to make a “correct” instance of what it is (Manchester Syntax); it’s not so easy to make natural language “natural”, but if we get close-ish, we’ve met a “valley”, perhaps of irritation. It’s not really an uncanny valley we’ve seen in our work with natural language generation from OWL ontologies, but when we generate sentences and paragraphs from OWL readers like the NLG form, but are caused irritation by English that is almost English, but not quite “natural” natural language. As we’ll see, this may be the nature of the errors; they’re basic errors in, for instance, the use of articles and plurals – not grammar fascism.
Doing NLG from an OWL axiom is sort of obvious; an axiom is a correlate of a sentence and we have nouns (and adjectives) in the form of classes and individuals, then properties (relationships) often do verby like things. A class or concept is the correlate of a paragraph; it’s what we want to say on a topic. So, we can take a set of axioms for classes from Snomed CT like
Class: Heart Disease
SubClassOf: (Disorder of Cardiovascular System) and (is-located-in some Heart Structure)
and similarly for hypertensive heart disease:
Class: Hypertensive heart disease
SubClassOf: (Heart Disease) and (is-associated-with some Hypertensive disorder)
And produce paragraphs like
A heart disease is a disorder of the cardiovascular system that is found in a heart structure.
and
A hypertensive heart disease is a heart disease that is associated with a hypertensive disorder.
These paragraphs are oK (produced by OntoVerbal), but are not “beautiful” English prose. In these cases, we’ve got the articles right etc, but it all seems a little plodding. There is some clunkiness that is a little irritating, but overall I think they’re pretty good and give a decent view on a set of axioms that can be fairly hard work to read. It is possible to produce better English, but at the cost of making a bespoke verbaliser for each ontology, especially for the “unpacking” of complex class labels to get articles and plurals correct; OntoVerbal is generic (though we did a little local fixing to help out with articles for Snomed classes). However, what we did do in OntoVerbal is to try and generate coherent, structured paragraphs of text for a class’ axioms. To get this coherence (rather than a set of sentences from unordered axioms for a class) we used rhetorical structure theory (RST) and mapped various types of OWL axioms to roles within RST. Example RST roles are evidence, motivation, contrast, elaboration, result, cause, condition, antithesis, alternative, list, concession and justification. these may be implicit within a text, but are often signalled by “discourse markers”; as “because” for evidence, “in order to” for enablement, “although” for antithesis, “but” for concession, “and” for list, “or” for alternative, etc You can see how we put all of this together in our IJACSA paper.
In the IJACSA paper we did an evaluation to look at the acceptability of these types of rendering and whether they were faithful enough to the OWL to allow “round-tripping” – that people experienced in OWL could take the verbalised OWL and re-produce the OWL axioms in Manchester Syntax. We also looked at the quality by comparing the machine verbalisations to human generated verbalisations. The argument is that a human produces good quality text (under the constraints given), so if the OntoVerbal text is similar to the human written text, then it should be of reasonable quality. Below are the OWL axioms from SNOMED for 10 classes. There are natural language verbalisations generated by a human (not in the project) and by OntoVerbal.
Number
|
OWL input
|
OntoVerbal
|
Ontologist
|
1
|
pelvic structure SubClassOf: lower trunk structure
lower trunk structure SubClassOf: structure of subregion of the trunk
pelvic structure SubClassOf: the pelvis and lower extremities and the abdomen and the pelvis and lower trunk structure
|
A lower trunk structure is a kind of structure of subregion of the trunk. A more specialised kind of lower trunk structure is pelvic structure. Another relevant aspect of lower trunk structure is that a pelvic structure is defined as the pelvis and lower extremities, the abdomen and the pelvis and a lower trunk structure.
|
A lower trunk structure is a structure of the subregion of the trunk. A pelvic structure is a subtype of a lower trunk structure.
|
2
|
procedure on the brain SubClassOf: intracranial procedure
intracranial procedure SubClassOf: procedure on the central nervous system
intracranial procedure SubClassOf: procedure on the head
intracranial procedure EquivalentClass procedure by site and has a procedure site some intracranial structure
|
An intracranial procedure is a kind of procedure on the central nervous system and procedure on the head. A more specialised kind of intracranial procedure is procedure on the brain. Additionally, an intracranial procedure is defined as a procedure by site that has a procedure site in an intracranial structure.
|
An intracranial procedure is a procedure on the central nervous system and a procedure on the head. Any procedure by site in which the procedure site is an intracranial structure is also an intracranial procedure. A procedure on the brain is a subtype of intracranial procedure.
|
3
|
abdominal vascular structure SubClassOf: abdominal and pelvic vascular structure
abdominal and pelvic vascular structure SubClassOf: vascular structure of the trunk
abdominal vascular structure SubClassOf: abdominal structure and abdominal and pelvic vascular structure
|
An abdominal and pelvic vascular structure is a kind of vascular structure of the trunk. A more specialised kind of abdominal and pelvic vascular structure is abdominal vascular structure. Another relevant aspect of abdominal and pelvic vascular structure is that an abdominal vascular structure is defined as an abdominal structure and an abdominal and pelvic vascular structure.
|
An abdominal and pelvic vascular structure is a vascular structure of the trunk. An abdominal vascular structure is a subtype of an abdominal and pelvic vascular structure.
|
4
|
chronic disease of the genitourinary system SubClassOf: chronic disease
chronic disease of the genitourinary system SubClassOf: disorder of the genitourinary system
chronic hypertensive uraemia SubClassOf: chronic disease of the genitourinary system
chronic disease of the genitourinary system EquivalentClass: chronic disease and disorder of the genitourinary system and has a finding site some structure of the genitourinary system
|
Chronic disease of the genitourinary system is a kind of chronic disease and disorder of the genitourinary system. A more specialised kind of chronic disease of the genitourinary system is chronic hypertensive uraemia. Additionally, chronic disease of the genitourinary system is defined as chronic disease that is a disorder of the genitourinary system, and has a finding site in a structure of the genitourinary system.
|
A chronic disease of the genitourinary system is a chronic disease and a disorder of the genitourinary system. Any chronic disease which is also a disorder of the genitourinary system and is found in the structure of the genitourinary system is also a chronic disease of the genitourinary system. A chronic hypertensive uraemia is a subtype of a chronic disease of the genitourinary system.
|
5
|
finding of the head and the neck region SubClassOf: finding of the body region
head finding SubClassOf: finding of the head and the neck region
finding of the head and the neck region EquivalentClass: finding of the body region and has a finding site some head and neck structure
head finding EquivalentClass: finding of the head and the neck region and has a finding site some head structure
|
A finding of the head and the neck region is a kind of finding of the body region. A more specialised kind of finding of the head and the neck region is head finding. Additionally, A finding of the head and the neck region is defined as a finding of the body region that has a finding site in a head and neck structure. Another relevant aspect of finding of the head and the neck region is that a head finding is defined as a finding of the head and the neck region that has a finding site in a head structure.
|
A finding of the head and the neck region is a finding of the body region. Any finding of the body which is found in a head and neck structure is also a finding of the head and neck region. A head finding is a subtype of the finding of the head and the neck region.
|
6
|
nephrosclerosis SubClassOf: degenerative disorder
degenerative disorder SubClassOf: disease
arteriosclerotic vascular disease SubClassOf: degenerative disorder
degenerative disorder EquivalentClass: disease and has an associated morphology some degenerative abnormality
|
Degenerative disorder is a kind of disease. More specialised kinds of degenerative disorder are nephrosclerosis and arteriosclerotic vascular disease. Additionally, degenerative disorder is defined as disease that has an associated morphology in a degenerative abnormality.
|
A degenerative disorder is a disease. Any disease which has an associated morphology of degenerative abnormality is also a degenerative disease. Nephrosclerosis and arteriosclerotic vascular disease are subtypes of degenerative disease.
|
7
|
kidney graft material SubClassOf: urinary tract material
kidney graft material SubClassOf: solid organ graft material
kidney graft material SubClassOf: urinary tract material and solid organ graft material
transplant of the kidney EquivalentClass: kidney operation and solid organ transplant and renal replacement and has a method some surgical transplantation action and has a direct substance some kidney graft material and has an indirect procedure site some kidney structure
|
A kidney graft material is a kind of urinary tract material and solid organ graft material. Another relevant aspect of kidney graft material is that a transplant of the kidney is defined as a kidney operation that is a solid organ transplant, and is a renal replacement, and has a method in a surgical transplantation action, and has a direct substance in a kidney graft material, and has an indirect procedure site in a kidney structure.
|
Kidney graft material is a urinary tract material and a solid organ graft material. A kidney operation, solid organ transplant and renal replacement which has a method of surgical transplantation action, a direct substance of kidney graft material and an indirect procedure site of kidney structure is a type of transplant of the kidney.
|
8
|
graft SubClassOf: biological surgical material
tissue graft material SubClassOf: graft
tissue graft material SubClassOf: graft and body tissue surgical material
|
A graft is a kind of biological surgical material. A more specialised kind of graft is tissue graft material. Another relevant aspect of graft is that a tissue graft material is defined as a graft and a body tissue surgical material.
|
A graft is a biological surgical material. Tissue graft material is a subtype of graft as well as a body tissue surgical material.
|
9
|
benign essential hypertension complicating and/or reason for care during pregnancy SubClassOf: essential hypertension complicating and/or reason for care during pregnancy
essential hypertension complicating and/or reason for care during pregnancy SubClassOf: essential hypertension in the obstetric context
essential hypertension complicating and/or reason for care during pregnancy SubClassOf: pre-existing hypertension in the obstetric context
essential hypertension complicating and/or reason for care during pregnancy SubClassOf: essential hypertension in the obstetric context and pre-existing hypertension in the obstetric context
benign essential hypertension complicating and/or reason for care during pregnancy SubClassOf: benign essential hypertension in the obstetric context and essential hypertension complicating and/or reason for care during pregnancy
|
Essential hypertension complicating and/or reason for care during pregnancy is a kind of essential hypertension in the obstetric context and pre-existing hypertension in the obstetric context. A more specialised kind of essential hypertension complicating and/or reason for care during pregnancy is benign essential hypertension complicating and/or reason for care during pregnancy. Another relevant aspect of essential hypertension complicating and/or reason for care during pregnancy is that benign essential hypertension complicating and/or reason for care during pregnancy is defined as benign essential hypertension in the obstetric context and essential hypertension complicating and/or reason for care during pregnancy.
|
An essential hypertension complicating and/or reason for care during pregnancy is an essential hypertension in the obstetric context and a pre-existing hypertension in the obstetric context. A benign essential hypertension complicating and/or reason for care during pregnancy is a subtype of essential hypertension complicating and/or reason for during pregnancy.
|
10
|
procedure on artery of the abdomen SubClassOf: procedure on the abdomen
procedure on artery of the abdomen SubClassOf: procedure on artery of the thorax and the abdomen
abdominal artery implantation SubClassOf: procedure on artery of the abdomen
procedure on artery of the abdomen EquivalentClass: procedure on artery and has a procedure site some structure of artery of the abdomen
|
A procedure on artery of the abdomen is a kind of procedure on the abdomen and procedure on artery of the thorax and the abdomen. A more specialised kind of procedure on artery of the abdomen is abdominal artery implantation. Additionally, a procedure on artery of the abdomen is defined as a procedure on artery that has a procedure site in a structure of artery of the abdomen.
|
A procedure on artery of the abdomen is a procedure of the abdomen and a procedure on artery of the thorax and the abdomen. Any procedure on artery which has a procedure site of structure of artery of the abdomen is also a procedure on artery of the abdomen. An abdominal artery implantation is a subtype of procedure on artery of the abdomen.
|
You can see that the verbalisations are fairly similar. Given the task of being faithful to the OWL and enabling “round-tripping”, very similar texts are produced by a human and OntoVerbal; the machine and human verbalisations are of a very similar quality. Evaluators could use both to round-trip to the OWL axioms, but did better with the OntoVerbal generated axioms. This is, we think, at least in part due to OntoVerbal being more “complete” in its verbalisation. The human verbalisation is smoother, but presumably not as smooth as a description written by a human domain expert could be (though do look at James Malone’s blog on this topic). However, I suspect that such smooth, natural language texts would be much harder to match to a set of OWL axioms.
Where does this leave my uncanny valley or valley of irritation for generated natural language? Domain expert humans writing definitions without the constraint of being faithful to the ontology’s axioms will probably avoid the valley of irritation; if there’s too much “ontologising” in the verbalisation there will be irritation (this came up in another paper on an earlier verbalisation); if there’s clunky English there is irritation. In a generic tool like OntoVerbal this is probably inevitable and I suspect it’s irritating as these are minor English errors that are always irritating as they disrupt reading. However, the use of rST does seem to give OntoVerbal’s NLG verbalisations a good level of coherence and fluency, even if they’re not perfectly fluent. They are also cheap to produce…. As they are close to the OWL they give an alternative view on those axioms – one thing I’d like to find out is if a verbalised view is any better (or worse) at allowing error spotting – and whther it is the verbalisation or just the alternative that does the job). One could also provide a variety of verbalisations – hand-crafted, luxury ones; ones close to the OWL and ones with and without the often inpenetrable words used in ontologies (especially for relationships).