A strategy for the cohabitation of conflicting theories of emotion within the Emotion Ontology and its annotations
Scientific ontologies aim to represent what is agreed upon by a consensus of scientific researchers. They aim to represent the best of scientific knowledge in a particular domain.
This objective is open to several objections. One is that the goal is a moving target -- after all, scientific knowledge is a constant progression. But this objection is a merit in disguise. Isn't it great that science progresses? And shouldn't we represent the best of scientific knowledge within a framework that is flexible enough to plug into multiple applications and to be extended without breaking any legacy systems? What framework could possibly be that cool? -- oh, yes, OWL. The second commonly raised objection carries more weight, however, and will be the main subject of this post. It is that, at least in some areas or domains, there is no consensus yet -- science isn't `settled', and scientists are still arguing about conflicting theories. One domain in which this seems to be the case is that of research into the emotions.
The Emotion Ontology is an ontology for the description and annotation of emotions and affective phenomena across research in the domains of pharmacology, psychiatry, neuroscience, psychology, genetics and any other domain that is interested in human affective functioning. It is being developed within the context of the OBO Foundry and as such is based on the upper-level Basic Formal Ontology and is designed to interoperate with other ontologies in the bio-sciences domain such as ChEBI for chemicals and the Gene Ontology for gene functions. Yes, knowledge about chemicals and gene functions IS important also in understanding the emotions -- it is well known, for example, that certain neurotransmitters are essential to the experience of emotion, and that drugs which alter the effect of those neurotransmitters by blocking the activity of their receptors can change how humans experience emotions. So, interoperability is important. The image to the right shows the current upper levels of the ontology aligned with BFO and the Mental Functioning Ontology as a mid-level bridge between the root BFO and the affect-specific domain.
This version of the Emotion Ontology has been largely based on the framework described in (Ceusters and Smith, 2010) and on the appraisal theory of the emotions as described in (Scherer and Mulligan, Emotion Review, forthcoming; Scherer et al., 2001; Deonna and Scherer, 2009). The key feature of this version of appraisal theory is that it bases emotions on an appraisal of relevance of some eliciting object or event (which can be external or internal). An example of an appraisal is when I see something black and elongated beneath the waves when I'm swimming in the sea, and something in me assesses -- very rapidly -- this could be a shark! and sharks are dangerous! and thereby elicits the emotion fear, which is accompanied by all sorts of processual changes including physiological responses (raised heart rate), behavioural action tendencies (urge to run away), subjective feeling (Aaaaaaah!) and an actual behavioural process including changes in facial expression. In other words, according to the appraisal theory, it all starts with the appraisal, and the appraisal is what represents the "content" of the emotion -- what, in other words, the emotion is about. But there is more: the theory also entails that different emotions are differentiated by their appraisals (which vary along differnent dimensions, one of which is valence -- positive vs. negative). Fear is the sort of emotion that results from an appraisal of dangerousness, anger is that sort of emotion that results from an appraisal of offense, and so on. The theory doesn't entail that emotions are only these appraisals, as it recognises the other aspects of emotion (the boxes in red in the image). Neither does it entail that there aren't also differences in these other aspects of emotion, such as differences in characteristic facial expressions for different emotions, differences in physiological response, and differences in how they feel to the persons undergoing them. But it makes the appraisal somehow central, essential -- at least one appraisal is necessary for an emotion occurrent to take place, and this appraisal triggers the rest of the (synchronized) changes that unfold during the occurrent experience of the emotion. In OWLSpeak, this can be minimally expressed as: 'emotion occurrent' subClassOf has part some 'appraisal process' (1). Every emotion occurrrent has an appraisal process as a part.
This is contested. Within the emotion research community, several different theories of the emotion abound, and not all of them assign such a central role to appraisal, or even include a role for appraisal at all. One such alternative theory was presented at the 2012 InterOntology conference in Tokyo earlier this year by Berit Brogaard. Brogaard raised several objections to the use of Scherer's appraisal theory as the basis for the Emotion Ontology, and suggested a "perceived response" theory to take its place. I'm not going to try to defend appraisal theory against the objections -- at least not yet -- although I believe that the objections rest on a mistaken view of what appraisal really is; for the interested, a very good exposition on this topic is given in the first part of Jenefer Robinson's brilliant 'Deeper than Reason'. For the moment, I will limit myself to discussing these different theories as a backdrop to tackling the pressing question of whether one ontology can sensibly be used to annotate scientific data despite the fact that different researchers subscribe to different theories. So, what are the sorts of differences involved?
Perceptual theories of the emotion view as essential component the internal perception of bodily changes that take place as a result of the emotion. Brogaard writes: "Suppose I see an aggressive-looking person approaching. The visual image activates sympathetic nervous system activation, which in turn gives rise to changes in my body state by acting on the muscles and hormonal levels. This change in body state then activates nerve cells in the brain. This causes a fear response." She adds that within her theory, an emotion is "a perception of an object or event causing a physiological response. For example, a state of fear may be a perception of an aggressive-looking person causing the muscles to tense up, the heart to pound, and the breath to shorten." So, we can summarise this as the assertion that, on the one hand, every emotion occurrent has this complex type of perceptual experience as a part, and on the other hand, that the object causes some or other the change in physiological state and then this causation is perceived or represented to the self. So, a change in physiological state is also a necessary, although not sufficient, component of emotion under this theory. Again, in OWLSpeak, this would look something like: 'emotion occurrent' subClassOf ( has part some 'perception of something causing physiological change' and has part some 'physiological change' ) (2). But, we can hypothesise at least, the physiological change here is no different to the physiological change process that is called 'physiological response to emotion process' in the existing Emotion Ontology -- but within the perceived-response theory, the physiological changes are not in response to the emotion, they precede the emotion, so they would never be called 'physiological response to emotion process'. Let's call this a Naming discrepancy between the two theories. Also, we can observe that 'appraisal' is present in one theory and absent in another. Let's call this an Additional entity discrepancy between the two theories. Finally, we note that the 'emotion occurrent' entity -- shared between the two ontologies -- is axiomatized differently (with different necessary parts) in the two theories. Let's call this an Axiomatization discrepancy. I'll now discuss our planned "cohabitation" approach in the context of each of these types of discrepancy.
1. Naming discrepancies
Naming discrepancies are in some senses the easiest to deal with in a single ontology. Although ontologies are annotated with labels that help people to understand what they represent, and to use the ontology correctly in annotations, the "core" content of the ontology is not the label but the logical structure and the definition. So, with a careful definition for physiological changes, there is no obstacle to reuse one entity for physiological changes involved in an emotion and to allow two different labels -- i.e. a primary label and an alternative label -- to be annotated. It is common in bio-ontologies such as ChEBI and GO to include multiple labels for each ontology entity -- and to create semantics-free stable identifiers so that even the preferred or primary label can change without affecting the underlying structure of the ontology. Better yet, when needed, synonyms can be scoped -- that is, given types that assign them to a particular community -- such that that community can use their preferred labels as the primary label for the ontology entity within their applications, but still re-use the same ontology, reducing effort and duplication.
But what about entities that appear or disappear across different ontologies? Trickier, but still possible:
2. Additional entity discrepancies
When one theory contains an entity - such as 'appraisal' -- that another theory simply does not, we propose that the underlying ontology should include the entity, regardless of which theory it is that postulates the entity, subject of course to the usual quality control -- i.e. there should be good evidence that the entity is a sensible one, fits the domain, will be used in annotations and so on. But the proponents of one theory, as consumers of the shared ontology, may not want to see the offending entities from the other ontology in their GUIs / tools / annotations and so on. The solution here is to provide a capacity for subsets of the overall ontology to be reliably extracted and redistributed, as is done for the Gene Ontology. Subsets are a way for different communities to flag up those entities that they are interested in and thereafter use only those entities in their ontology-based tools and interfaces. It is a kind of ontology modularity -- where the modules are chosen by the community who want to use them -- but, crucially, the modules also contain overlapping entities, meaning that each community still benefits, in terms of lower costs of maintenance thanks to pooled resources, from the shared underlying ontology.
But what about axioms? Can the same approach work there, i.e. can we include both (1) and (2) in the same ontology?
3. Axiomatization discrepancies
The short answer is no, it can't. OWL is monotonic, which means inter alia that while you can add knowledge to the ontology, the added knowledge can't contradict what is already known. If the ontology contains both axioms (1) and (2), and assuming that additional axioms such as disjointness and closure are added to the ontology to correspond to the different intentions of the two theories, it will be inconsistent and therefore unusable by logic-based reasoning tools. Oh dear. What are we to do?
One approach would be to take a pragmatic stance and simply privilege one theory above the other -- accepting that by so doing you will be annoying proponents of the second theory. In a sense, this is what the current version of the Emotion Ontology is doing. However, this cannot be the long-term perspective if the project aims to become a shared community-wide resource, which we do. So, a different approach is needed. Another strategy would be to engage with the community -- holding meetings, for example, during which the proponents of the two theories are brought together with the ontology developers as mediator and brought to hold constructive discussions towards resolution and agreement on a compromise that is acceptable to both parties. This approach has worked in previous contexts, such as in ontology alignment efforts (e.g. between ChEBI and GO). And -- it should be noted -- that even with the offending axiom in place, the entities in the ontology can be used in annotation of data across research arising from both theoretical perspectives. Who knows, once a sufficient quantity of data is annotated, maybe an ontology-based analysis of the data will start to speak for itself in favour of one theory or another theory?
Suffice to say -- we're still optimistic.
... or, what makes an ontology an ontology?
Every now and then, thankfully not that frequently, somebody asks me "What's the point of ChEBI, when MeSH already does chemical classification?"
MeSH does indeed provide an organised controlled vocabulary that serves a similar function to ChEBI with respect to the organization of text-based literature. In fact, you can now compare ChEBI and MeSH classification side-by-side in the newly released PubChem classification component of the PubChem chemical knowledge base, along with the chemical classification from KEGG and others. MeSH (Medical Subject Headers) is used to index and organise literature in MEDLINE. They have a dedicated chemical section in the Supplementary Concept Records, which included about 300,000 chemical terms in 2010. More information about chemicals in MeSH can be found in Stefan Schulz's presentation to the 2010 ChEBI User Group workshop on the topic of the representation of chemicals in large medical vocabularies.
That all sounds quite good, actually. So what's the down side?
Simply put, MeSH is not an ontology. It's not even a chemical database. It's a thesaurus. And while it is well optimised for the application scenario for which it was designed -- indexing of primary literature -- it suffers from several limitations with respect to more general use cases for chemical knowledge representation.
1) MeSH is name-based rather than structure-based, making it largely inaccessible to cheminformatics applications and hindering its application to integration of chemical entities appearing in different chemical or biological databases where different naming conventions are used. ChEBI maintains chemical structures in MOL format internally and exports popular cheminformatics formats SMILES and InChI as well as maintaining cross-references of identifiers such as CAS and Reaxys where we can find those as open data. ChEBI is thus structure-searchable, and indeed enables chemical structure searching across all databases that use ChEBI IDs for chemical annotation, such as the IntAct protein interaction database.
2) MeSH uses only a small number of relationships, such as is-a or part-of, which are not formally (logically) defined and are in some cases ambiguously used, .such as using the same classification relationship to link chemicals to their pharmacological action, classification parents and parts. That means that the relationships cannot be used as a basis for automated reasoning with logic-based tools, such as is needed in the context of the Semantic Web. ChEBI uses the is-a and has-part relationships strictly, linking chemicals to their pharmacological actions using a has-role relationship, and using several structural relationships to link biologically relevant structurally related chemicals together such as conjugate bases and acids and tautomers.
3) MeSH does not provide stable unique identifiers for chemicals and chemical classes that can be used to annotate chemicals in the context of biological databases. ChEBI does, following the OBO Foundry ID policy.
Given all of the above, ChEBI enables a much broader range of applications in bioinformatics, metabolomics and chemical biology than MeSH does. Yes, ChEBI is smaller. But we're slowly but surely catching up.