Understanding societal debates on the Web and how they are impacted by the spread of biased narratives and falsehoods are becoming increasingly important issues. The notion of a claim is central in a number of related studies into fake-news propagation or computational fact-checking. While the understanding of this notion varies from one eld to another, there are few studies that have focused on the conceptual modeling of claims and their context. We attempt to contribute to this area by proposing a novel conceptual model for claims and related notions, such as attitudes, reviews and annotations, that aims to take into consideration the claims inherent complexity, distinguishing between their meaning, linguistic representation and context. We provide an example of an implementation of this model by using established vocabularies, such as schema.org, Open Annotation and PROV-O, and discuss the challenges related to this work. 1
The spread of biased narratives and falsehoods on the Web and the analyses
of online discourse have become increasingly important issues [
Whereas techniques for knowledge graph construction and augmentation
often deploy methods strongly related to the aforementioned computational
methods related to claims, e.g., when aiming to verify facts from the Web for
augmenting knowledge bases [
In order to facilitate the advancement of tasks such as claim veri cation or fact checking, it is crucial to capture the complexity of a claim in a way which enables unambiguous interpretation by both humans and machines. However, both the used terminology and the underlying conceptual models are still strongly diverging in academic literature (Sect. 2) as well as in the conceptual models deployed by fact-checking sites.
Therefore, capturing the meaning of a claim requires both the preservation of the actual claim utterance as natural language text, often carrying a range of statements and sentiments embedded in complex sentences which are easy to process by humans but hard to interpret by machines, as well as structured knowledge about a claim, its context and constituents, which enables machineinterpretation, discoverability and reuse of claims, for instance, to facilitate research in the aforementioned areas.
This paper makes the following main contributions: i) a conceptual model
and corresponding terminology of claims and their constituents and context,
grounded in both the scienti c state-of-the-art in related elds such as
argumentation mining as well as the actual practices of representing and sharing
claims on the Web, for instance, as part of fact-checking sites; ii) an RDF
implementation of the proposed conceptual model that uses W3C standards for data
sharing, namely RDFS, and is informed by established vocabularies, such as
schema.org, Open Annotation, and the PROV data model, in order to facilitate
Web-scale sharing, discovery and reuse of claims and their context, for instance
through semi-structured Web page markup or as part of dedicated knowledge
graphs such as ClaimsKG [
While the analysis of claims plays a crucial role for a number of elds, the de nition of the very concept of a claim is often left to the intuition of the reader. Existing de nitions vary considerably across and also within elds.
According to the Oxford English Dictionary, a claim is a statement or
assertion that something is the case, typically without providing evidence or proof.2
Platforms dedicated to journalistic fact-checking refer to claims as statements
supported by (a group of) people or organizations that appear newsworthy,
signi cant and veri able.3 An RDFS-based model for such fact-checked claims is
introduced in [
In argumentation mining, claims denominate the conclusion of an argument
or the assertion the argument aims to prove [
Thus, what is identi ed as \claim" in a particular work may or may not be called \claim" in another. While it is the belief of a person about a fact that is called \claim" in argumentation mining, it is the fact itself that is coined \claim" in the fact-checking community. Similarly, the belief and opinion about certain consequences are the argumentative \claim", while fact-checking may verify whether the anticipated consequences would indeed follow an action. Statements expressing the position of a person towards a proposition or target are not susceptible to fact-checking (unless the correctness of the quotation is to be veri ed) but are a prevalent claim type in argumentation mining. Moreover, what is used as premise or evidence in an argument is often selected as check-worthy \claim" by fact-checking sites. Generally, the distinction of argumentative units such as claims and evidence is based on the statements' usage in an argument while fact-checking classi es statements as claims depending primarily on features inherent to the statement itself.
In an e ort to reconcile these di erent understandings of the concept of a
\claim", we propose a model considering requirements from various research
elds. While in argumentation mining, the meaning of a claim in the context of
the current discourse is the signi cant part, many tasks from the fact-checking
community, e.g. those aiming at matching unchecked statements to fact-checked
claims [
Overview. We distinguish three main components of a claim, represented by three central classes: (1) claim proposition, (2) claim utterance, and (3) claim context. A claim proposition is the meaning of a statement or assertion that something is the case. It is usually related to a controversial topic and can be 2 https://www.lexico.com/en/definition/claim 3 https://www.truthorfiction.com/about/ https://checkyourfact.com/about-us https://www.politifact.com/truth-o-meter/article/2018/feb/12/ principles-truth-o-meter-politifacts-methodology-i/ factual or subjective (expressing an opinion). A claim proposition can be expressed in many di erent ways and in di erent contexts, thus it has one or more claim utterances. For example, it may be expressed in di erent languages, using di erent words in the same language, or uttered by di erent persons and/or in di erent points in time. On the contrary, a speci c claim utterance can be associated to only one proposition, i.e., it has a single meaning. The claim proposition can be represented in di erent ways, for example, by selecting a representative utterance or through a more formal model. Each claim utterance is related to a speci c claim context, like the author of the claim or its date. It provides the means to interpret the claim utterance and thus understand its proposition. Below, we provide details and the main properties of each of these three main classes (without repeating the associations among them).
Claim Proposition. A claim proposition re ects the meaning of one or more semantically equivalent claim utterances expressed in di erent linguistic forms or contexts. A claim proposition is associated with i) one or more preferred representations, ii) one or more reviews, and iii) one or more attitudes (Fig. 1). A representation can have the form of free text, e.g., a sentence that best describes the proposition (like the text of one of the corresponding utterances), or be more complex, e.g,. a rst-order logic model. A review is a resource (e.g., a document) that analyzes one or more check-worthy claim propositions and provides a verdict about their veracity or trustworthiness. An example of such a review is an article published by a fact-checking organization. Note that not all claims have a review or verdict. For instance, the claim \the presence of a gun makes a con ict more likely to become violent" represents a hypothesis and is di cult to be associated with a correctness score (there may be mixed evidence supporting and contradicting it). An attitude is an opinion on a given topic (e.g., a viewpoint), which often underlies a set of speci c values, beliefs or principles. For instance, pro-Brexit and pro-Remain are two di erent attitudes for the Brexit topic. A claim proposition can be associated with several attitudes for di erent topics. For example, the claim \immigrants are taking our jobs" supports both the against immigration attitude (for the Immigration topic) and the pro-Brexit attitude (for the Brexit topic).
Claim Utterance. A claim utterance is the act of expressing a claim proposition in a speci c natural language and form (like text or speech). Among other things, it may be something said by a politician during an interview, a text within a news article written by a journalist, or a tweet posted by a celebrity about a controversial topic. It is associated with i) one or more linguistic representations (subclass of representation in Fig. 1), and ii) one or more sources (Fig. 2). A linguistic representation can be, for example, a text in a speci c language that best imprints the claim as it was said/appeared, or a sound excerpt from someone's speech. A source provides evidence of the claim existence. For instance, it can be the URL of an interview video, a news article, or a tweet. A linguistic representation can have one or more annotations which provide formal linguistic characteristics, like an entity or date mentioned in the text of the claim utterance, the polarity of this text (e.g., positive, negative, neutral), or the linguistic tone of a speech (like irony). The annotation can enable advanced exploration of the claims (e.g., based on mentioned entities) and can be manually provided by a domain expert or automatically produced using a NLP or speech processing tool (like an entity linking tool for the case of entity annotation in text). Claim Context. The claim context provides background information about the claim utterance (Fig. 3). Together with the linguistic representation of the claim utterance, it can provide an answer to the Five W's : i) what was said (linguistic representation of claim utterance), ii) who said it (author of the claim), iii) when it was said (date the claim was said), iv) where it was said (location the claim was said), and v) why it was said (event or activity in the context of which the claim was said). The claim context provides the necessary information for interpreting the claim utterance (and thus understanding its proposition), and can be extended with more concepts that allow describing additional context information about the claim utterance (like the topic of the underlying discourse or the medium used for uttering the claim). Instantiation Example. Fig. 4 depicts an instantiation example of the proposed conceptual model. The example shows information for two claim utterances (in pink background): i) one said by David Dimbleby during a topical debate in Dover (\We are going to be paying until 2064, apparently"), and ii) one extracted from a news article of The Independent (\UK will be paying Brexit `divoce bill' until 2064"). Both utterances correspond to the same claim proposition (in green background) and each one has its own context information (in yellow background). The linguistic representation of the rst claim utterance has been annotated with one date annotation (2064) and that of the second claim utterance with one entity annotation (UK). The claim proposition has two representations, a textual one (\Britain will be paying its Brexit bill for 45 years after leaving the EU") and a formal one (\cost = fof=Brexit, for=UK amount=?, until=2064g"), and supports the against-Brexit attitude for the Brexit topic. In addition, there is a review of this claim proposition with verdict \true", published by Full Fact (UK's independent fact-checking organisation). We can also see the URL of the review article as well as a reference to a PDF le which provides evidence for its correctness. The context of each claim utterance provides additional metadata about the claim. For example, we see that the rst utterance was said by David Dimbleby on 15.03.2018, in the context of a debate about Brexit which took place in Dover. 4
We introduce an RDF/S implementation of the proposed conceptual model using established vocabularies, in particular schema.org,4 the Open Annotation (OA) Data Model,5 the Marl Ontology,6 the NLP Interchange Format (NIF),7 and the PROV Data Model.8 The selection of these vocabularies was based on the following three main objectives: i) relying on stable term identi ers and persistent hosting, ii) being supported by a community, iii) being extensible.
Fig. 5 depicts the proposed schema. For representing the main concepts of our conceptual model, we exploit classes and properties of schema.org, a collaborative, community activity with a mission to maintain and promote a common schema for structured data on the Web and beyond. We make use of the class schema:Claim (currently under integration in schema.org) to describe a claim utterance. According to schema.org, this class represents a speci c, factuallyoriented claim. For the claim proposition, we use the class schema:Intangible, a utility class that serves as the umbrella for a number of `intangible' things. Although this class does not su ciently re ect the semantics of a claim proposition, it appears to be the most reasonable term for representing a proposition. For the same reason, we use schema:Intangible to describe a claim context. 4 https://schema.org/ 5 http://www.openannotation.org/ 6 http://www.gsi.dit.upm.es/ontologies/marl/ 7 https://persistence.uni-leipzig.org/nlp2rdf/ 8 https://www.w3.org/TR/prov-dm/
> ) 6 & 6 & 6 & ; ,& 5 8 $ , & - ) , % # ) " 3 24 -& -& -& &
< ? & # # ! , = -& & ( " & & '%
" & " 1 ) 9 2 < 6 # $ -& # $%!
" # $ 156 32 # " 18 8 2 " & // . )/ & /, 9 # 9, 9 / " & // . , . 0/3879: 6 "+ '!.& = -& > ) 6 &
-& # 1B ) ) , & - ) && -24 ; * & ((
" " & //$$$.,, . . 0/ & ) /, $ ) + & ' " ) & -& " ' (
An alternative solution is to bypass the claim context class and directly link an instance of schema:Claim to instances of the four classes connected to the claim context (author, date, location, event ). These four classes are described through corresponding schema.org classes: schema:Thing (e.g., a person, an organization, a blog, etc.), schema:Date, schema:Place, schema:Event. For connecting a schema:Claim to a schema:Intangible, we can use the property schema:about or its inverse schema:subjectOf.
For representing a source, we use the class schema:CreativeWork (or one of its sub-classes). Thereby, we take advantage of its properties and can describe additional information about the source, such as headline, language, keywords, publisher, etc. The linguistic representation of a claim utterance, as well as the (preferred) representation of a claim proposition, can be described through the class schema:Text (for textual representations) or schema:MediaObject (for image, audio or video representations). For describing annotations, we make use of the widely-used OA and NIF data models, while provenance information is represented though the PROV data model. NIF allows us to include detailed information about the outcome of an NLP process on textual representations (like begin/end indexes and con dence scores). The review of a claim proposition is described through the class schema:ClaimReview, which in turn is connected to a schema:Rating for assigning a rating score about the veracity of the claim proposition. Finally, we exploit the Marl ontology to represent attitudes. Marl is a data schema designed to annotate and describe subjective opinions, and provides the attributes that enable to connect opinions with contextual information. 5
We propose a conceptual model and an implementation based on existing RDF vocabularies to represent and contextualize claims and related entities. Our work is meant to advance a shared understanding of claims and related terminology across communities, and the semi-structured representation of claims and their contexts to foster transparency, reproducibility and a joint advancement of research in related elds.
An open challenge is the detection and representation of the inherent relations between claims as well as their relations to other entities or resources. In particular, the semantic relatedness of claims is re ected by the relations between their proposition components. Establishing relations across the model classes, e.g., relating an utterance to a proposition, allows the uncovering of paraphrased claims with identical meaning, while topical relations between claims are of crucial importance to enable retrieval and search of claims.
Other challenges concern information extraction techniques geared towards the extraction of utterances from text or audio together with the attitude towards a particular topic, as well as additional contextual information, such as authors or sources. From a knowledge representation perspective, we emphasize the need for a formal representation of propositions (e.g. by applying dynamic predicate logics), or for extending the Marl model in order to represent speci cally viewpoints instead of general opinions on objects. This reveals the need for the development of a dedicated ontology for claims representation.