Textual emotion analysis is a new eld whose aim is to detect emotions in user generated content. It complements Sentiment Analysis in the characterization of users subjective opinions and feelings. Nevertheless, there is a lack of available lexical and semantic emotion resources that could foster the development of emotion analysis services. Some of the barriers for developing such resources are the diversity of emotion theories and the absence of a vocabulary to express emotion characteristics. This article presents a semantic vocabulary, called Onyx, intended to provide support to represent emotion characteristics in lexical resources and emotion analysis services. Onyx follows the Linked Data principles as it is aligned with the Provenance Ontology. It also takes a linguistic Linked Data approach: it is aligned with the Provenance Ontology, it represents lexical resources as linked data, and has been integrated with Lemon, an increasingly popular RDF model for representing lexical entries. Furthermore, it does not prescribe any emotion model and can be linked to heterogeneous emotion models expressed as Linked Data. Onyx representations can also be published using W3C EmotionML markup, based on the proposed mapping.
From the tech-savvy to elders, our society is exponentially moving its social and professional activity to the Internet, with its myriad of services and social networks. Facebook1 or Twitter2 are only two of the most successful examples, producing ooding streams of user-generated data. Unluckily, quite often that information is just meant for human consumption and is only formatted to be displayed. This prevents us from automatically processing these massive streams of information to aggregate, summarize or transform them and present human users with a bigger picture. In other words, data mining techniques require machine-formatted data input.
In an attempt to shorten that gap, the multidisciplinary eld called Sentiment
Analysis or Opinion Mining was born, which aims at determining the subjectivity
of human opinions. Many tools have been created to enrich or make sense out of
human generated content by applying natural language processing and adding
the results as annotations or tags. Whilst this solves the issue at a small scale,
for each ad-hoc solution, it raises another problem: data collected by di erent
programs presents di erent and sometimes incompatible formats. Linked Data
introduced a lingua franca for data representation as well as a set of tools to
process and share such information. Many services embraced the Linked Data
concepts and are providing tools to interconnect the previously closed silos of
information [
The Sentiment Analysis eld is now evolving to determine also human
emotions. An important fact about emotions is that they change the way we
communicate [
Social networks aside, another eld of application of Emotion Analysis is
A ective computing. There are a variety of systems whose only human-machine
communication is purely text-based. These systems are often referred to as dialog
systems (e.g. Q&A systems). Such systems can use the emotive information to
change their behaviour and responses [
On the other hand, the rise of services like microblogging will inevitably lead to services that exchange and use a ective information. Some social sites are already using emotions natively, giving their users the chance to share emotions or use them in queries. Facebook, for instance, recently updated the way its users can share personal statuses.
These sites have started making heavy use [
Combining the objective facts already published as Linked Data with subjective opinions extracted using Sentiment and Emotion analysis techniques can enable a wide array of new services. Unfortunately, there is not yet any widely accepted Linked Data representation for emotions. This paper aims at bridging this gap with the de nition of a new vocabulary, Onyx.
This paper is structured as follows: Section 2 introduces the technologies that Onyx is based upon, as well as the challenges related to Emotion Analysis and creating a standard model for emotions, including a succinct overview of the formats currently in use; Section 3 covers the Onyx ontology in detail and several use cases for this ontology; Section 4 presents the results of our evaluation of the Ontology, focusing on the coverage of current formats like EmotionML; Section 5 completes this paper with our conclusions and future work.
2.1
To work with Emotions and reason about them, we rst need to have a solid understanding and model of emotions. This, however, turns out to be a rather complex task.It is comprised of two main components: modelling (including categorisation) and representation.
There are several models for emotions, ranging from the most simplistic and
ancient that come from Chinese philosophers to the most modern theories that
re ne and expand older models [
Other models cover a ects in general, which include Emotions as part of
them. One of them is the work done by Strapparava and Valitutti in
WordNetA ect [
Despite all, there does not seem to be a universally accepted model for
emotions [
EmotionML [
In the eld of Semantic Technologies, Grassi presented Human Emotion
Ontology (HEO). This ontology presents an ontology for human emotions for its use
for annotating emotions in multimedia data. Another work worth mentioning is
that of Hastings et al. [
For Opinion Mining we nd the Marl vocabulary [
Lastly, it is worth mentioning lemon, the Lexicon Model for Ontologies. As
its name indicates, it is a model that supports the sharing of terminological and
lexicon resources on the Semantic Web as well as their linking to the existing
semantic representation provided by ontologies [
Provenance is information about entities, activities, and people involved in
producing a piece of data or thing, which can be used to form assessments about its
quality, reliability or trustworthiness. The PROV Family of Documents de nes
a model, corresponding serializations and other supporting de nitions to enable
the inter-operable interchange of provenance information in heterogeneous
environments such as the Web [
There are many advantages to adding provenance information in Sentiment Analysis in particular as di erent algorithms may produce di erent results. By including the Provenance classes in our Emotion Mining Ontology we can not only link results with the source from which it was extracted, but also with the algorithm that produced them. 3
Onyx is a vocabulary to represent the Emotion Analysis process and its results, as well as annotating lexical resources for Emotion Analysis. It includes all the necessary classes and properties to provide structured and meaningful Emotion Analysis results, and to connect results from di erent providers and applications.
At its core, the Onyx ontology has three main classes: EmotionAnalysis, EmotionSet and Emotion. In a standard Emotion Analysis, these three classes are related as follows: an EmotionAnalysis is run on a source (generally in the form of text, e.g. a status update), the result is represented as one or more EmotionSet instances that contain one or more Emotion instances.
The EmotionAnalysis instance contains information about: the source (e.g. dataset, website) from which the information was taken, the algorithm used, and the emotion model that was used to represent emotions. Additionally, it can make use of Provenance to specify the Agent in charge of the analysis, the resources used (e.g. dictionaries), and other useful information.
An EmotionSet contains a group of emotions found in the text or one of its parts. As such, it contains information about: the original text (extractedFrom); the exact excerpt that contains the emotion or emotions (emotionText); the person that showed the emotions (sioc:has creator); the entity that the emotion is related to (describesObject); the concrete part of that object it refers to (describesObjectPart); the feature about that part or object that triggers the emotion (describesFeature); and, lastly, the domain detected. All this properties are straightforward, but a note should be given about the domain property. Different emotions could have di erent interpretations in di erent contexts (e.g., fear is positive when referred to a thriller, but negative when it comes to cars and safety).
When several EmotionSet instances are related, an AggregatedEmotionSet can be created that links to all of them. For instance, we could aggregate all the emotions about a Movie, or all the emotions shown by a particular user. An AggregatedEmotionSet is a subclass of EmotionSet which contains additional information about the original EmotionSet instances it aggregates.
Considering the lack of consensus on modeling and categorizing emotions, our model of emotions is very generic. In this Emotion model we include: an EmotionCategory or type of emotion (although more could be speci ed), through the hasEmotionCategory property (e.g. \sadness"); the emotion intensity; action tendencies (ActionTendency) related to this emotion, or actions that are triggered by the emotion; appraisals and dimensions. Appraisals and dimensions are de ned as properties, whose value is a oat number. On top of that generic model, we have adapted two di erent systems: the WordNet-A ect taxonomy, and the EmotionML vocabularies for categories, dimensions and appraisals.
WordNet-A ect [
Regarding EmotionML, we have converted its vocabularies [
This is further developed in Section 4.
It is also possible that two separate emotions, when found simultaneously, imply a third emotion. A more complex one. For instance, \thinking of the awful things I've done makes me want to cry" might reveal sadness and disgust, which together might be interpreted as remorse. In such situation, we could add an AggregatedEmotion that represents remorse to the EmotionSet, linking it to the primary emotions with the aggregatesEmotion property.
To group all the attributes that correspond to a speci c emotion model, we created the EmotionModel class. Each EmotionModel will be linked to the di erent categories it contains (hasEmotionCategory), the AppraisalProperty or DimensionProperty instances it introduces (through hasAppraisalProperty and hasDimensionProperty), etc.
Figure 2 shows a complete overview of all these classes, as well as all their properties.
After this introduction of the ontology, we will present several use cases for it. This should give a better understanding of the whole ontology by example. Rather than exhaustive and complex real life applications, these examples are meant as simple self-contained showcases of the capabilities of semantic Emotion Analysis using Onyx. For the sake of brevity, we will omit the pre x declaration in the examples.
Case An example Analysis.
EmotionProcessing \I lost one
hour today because of
the strikes!!", by the user
JohnDoe
Example of annotation of
a lexical entry using Onyx
and lemon [
N3 Representation : customAnalysis a onyx : EmotionAnalysis ; onyx : algorithm " SimpleAlgorithm "; onyx : usesEmotionModel wna : WNAModel . : result1 a onyx : EmotionSet ; prov : wasGeneratedBy : customAnalysis ; sioc : has_creator [ sioc : UserAccount <http :// blog . example .
com / JohnDoe >. ]; onyx : hasEmotion [ onyx : hasEmotionCategory wna : anger ; onyx : hasEmotionIntensity :0.9 ]; onyx : emotionText "I lost one hour today
because of the strikes !!" ; dcterms : created "2013 -05 -16 T19 :20:30+01:00"
^^ dcterms : W3CDTF . : fifa a lemon : Lexicalentry ; lemon : sense [ lemon : reference wn : synset - fear - noun -1; onyx : hasEmotion
[ oyx : hasEmotionCategory wna : fear . ].
];
lexinfo : partOfSpeech lexinfo : noun .
Evaluating ontologies is always a di cult task. Evaluation methodologies are
highly debatable and there are no standards [
Comparing two Emotion Mining algorithms by comparing the discrepancies in the results obtained using both.
Query SELECT DISTINCT ? creator ? cat WHERE { ? set onyx : hasEmotion
[ onyx : hasEmotionCategory ? cat ]; dcterms : created ? date ; sioc : has_creator ? creator . ? cat skos : broaderTransitive * wna : negative
- emotion .
FILTER ( ? date >= xsd : date (" 2012 -12 -31 ")
? date <= xsd : date (" 2013 -01 -01 ") ) } SELECT ? source1 ? algo1 ( GROUP_CONCAT (? cat1 )
as ? cats1 ) WHERE { ? set1 onyx : extractedFrom ? source1 . ? analysis1 prov : generated ? set1 ;
onyx : algorithm ? algo1 . ? set1 onyx : hasEmotion
[ onyx : hasEmotionCategory ? cat1 ].
FILTER EXISTS { ? set2 onyx : extractedFrom ? source1 . ? analysis2 prov : generated ? set2 . ? set2 onyx : hasEmotion
[ onyx : hasEmotionCategory ? cat2 ].
FILTER ( ? set1 != ? set2 ).
FILTER ( ? cat2 != ? cat1 ).
} } GROUP BY ? source1 ? algo1
ORDER BY ? source1
adaptation of a well-known Emotion Analysis tool to output Onyx, Synesketch
[
For the EmotionML part, the evaluation process is split into two parts:
transforming the EmotionML categories into a semantic format, and representing
EmotionML cases with Onyx. The result of the former can be seen in [
Synesketch is a library and application that detects emotions in English texts and can generate images that re ect those emotions. Originally written in Java, it has been uno cially ported to several programming languages (including PHP), which shows the interest of the community in this tool. The aim of the PHP port was, among others, to o er a public endpoint for emotion analysis, which later had to be taken down due to misuse. The relevance of this tool and its Open Source license were the leading factors in choosing this tool. Our approach has been to develop a proof-of-concept web service that performs Emotion Analysis using Synesketch's emotion analysis. The service can be accessed via a REST API and its results are presented in Onyx, using the RDF format.
The Synesketch library uses the big-6 emotional model, which comprises: happiness, sadness, fear, anger, disgust and surprise. Each of those emotions are present in the input text with a certain weight that ranges from 0 to 1. Additionally, it has two attributes more that correspond to the general emotional valence (positive, negative or neutral) and the general emotional weight. In other words, these attributes together show how "positive", "negative" or "neutral" the overall emotion is.
To represent the big-6 emotion category in Onyx we used EmotionML's big-6 category, which we previously mapped to Onyx. The Synesketch weight directly mapped to hasEmotionIntensity in Onyx.
However, the General Emotional valence and weight do not directly match any Onyx property or class. To solve it, we simply added an AggregatedEmotion with the PositiveEmotion, NeutralEmotion or NegativeEmotion category (as de ned by WordNet-A ect) depending on the value of the valence. The general emotional weight is then the intensity of this AggregatedEmotion, just like in the other cases.
The nal result is a REST service that is publicly available at our website3. EmotionML < emotionml xmlns =" http: //.../ emotionml " xmlns:meta =" http: //.../
metadata " category - set =" http: //.../#
everyday categories " > <info > < classifiers:classifier
classifiers:name =" GMM "/ > </ info > < emotion > < category name =" Disgust " value
=" 0.82 "/ > 'Come , there is no use in
crying like that ! ' </ emotion > said Alice to herself rather
sharply ; < emotion > < category name =" Anger " value ="
0.57 "/ > 'I advise you to leave off
this minute ! ' </ emotion > </ emotionml >
Onyx : Set1 a onyx : EmotionSet ; onyx : extractedFrom " Come , there is no use in crying like that ! said Alice to herself rather sharply ; I advice you to live off this minute !"; onyx : hasEmotion : Emo1 onyx : hasEmotion : Emo2 : Emo1 a onyx : Emotion ; onyx : hasEmotionCategory
emlonyx : disgust ; onyx : hasEmotionIntensity 0.82; onyx : hasEmotionText " Come , there 's no use in
crying like that !" : Emo2 a onyx : Emotion ; onyx : hasEmotionCategory emlonyx : anger ; onyx : hasEmotionIntensity 0.57; onyx : hasEmotionText "I advice you to leave
off this minute !" : Analysis1 a onyx : EmotionAnalysis ; onyx : algorithm " GMM "; onyx : usesEmotionModel emlonyx : everyday
categories ; prov : generated Set1 .
With this work we have introduced an option to represent Emotions that takes advantage of the work conducted in the eld of Semantic Web. This ontology presents characteristics that are particularly bene cial for any process of Emotion Analysis. Onyx provides a structured format for Emotion Analysis. It addresses the problem of supporting heterogeneous categories of emotions, and new categories and features can be added, using the recommended taxonomy to link them and retain compatibility.
We also presented how Onyx would be used in several scenarios. Furthermore, we adapted some of the existent resources and services to Onyx, making them publicly available.
Although this paper is focused on Emotion Analysis, emotive information can also be directly provided by users. Either given explicitly or extracted via an automated process (Emotion Analysis), the information they represent is the same. A single ontology should thus cover both scenarios. This is possible with Onyx, as we demonstrate in this paper.
We would like to note that our proposal is compatible with EMO, since
EMO can be easily mapped to Onyx using the property usesEmotionModel.
The situation is similar with the proposal of Lopez et al. [
With all this in mind, we consider that using Onyx to represent Emotion Mining results is highly bene cial.
As part of the future plans for Onyx, it will be actively used in the
EUROSENTIMENT4 project, whose aims is to create a language resource pool for
Sentiment Analysis. Together with Marl [
There is also room for experimentation emotion composition and inference using tools such as SPIN6. It is possible to infer complex emotions whenever other simple emotions are present, and vice versa. The same techniques could be used to work with di erent emotion models
Finally, our research group will use the integration with EmotionML to develop intelligent personal agents that bene t from the potential of the Semantic Web.
nif-core.html
This work was partially funded by the EUROSENTIMENT FP7 Project (Grant Agreement no: 296277)