3.1. The structure of claims

Claims may have a large diversity of forms: a fact, a thesis, an opinion, a judgment, a position on a topic, a question, a rhetorical question, etc. Claims must have a debatable character: they express something that is plausible but may be questioned in various ways. This plausible character is often realized by means of evaluative terms or expressions. Statements become claims when it is possible to find related justifications or attacks. This section presents the main linguistic cues that are typical of claims. However, a number of claims are not marked by any linguistic cues: their argumentative character is pragmatic. Furthermore, some of these linguistic cues can also be found in justifications and other types of pragmatic constructions, which makes the identification of claims difficult.

The first difficult point in claim identification is to delimit the text span that could constitute a claim. In general, it is a sentence, ending with a full stop, an exclamation or a question mark. A claim may be longer if it includes restrictions, comments, definitions or circumstances of various kinds. A sentence may also be composed of several claims, possibly hierarchically organized, for example:

Due to the necessity of a systematic vaccination against Ebola, the authorities of country C have decided to reduce the funding of education in English.

This statement contains two potential claims (systematic vaccination and funding reduction), which can be equally debated. Furthermore, the causal link between the two can be also debated. If the vaccination necessity is attacked, then the latter claim is also attacked. The need to delimit claims has led to the emergence of a research topic, as well as the notion of argumentative discourse unit (ADU), derived from the notion of elementary discourse unit (EDU) in discourse analysis. ADUs are segments that, in general, contain an argument, a claim, a support or an attack.

Claim delimitation is addressed in Chapter 6, which is dedicated to argument annotation. So far, it is essentially an experimental process based on annotations, even if cues such as punctuation, connectors and relative clauses can be taken into account.

The second challenge is to identify statements that have a debatable character. This identification can be realized partly by analyzing the linguistic elements in the statement, which introduce this debatable character. A first set of linguistic cues are those that are proper to the content of the claim. These cues include, but are not limited to, the following:

– Scalar adjectives such as easy, difficult, efficient, unsafe, which have an evaluative dimension in the context of a claim. These adjectives evaluate a property or a characteristic that is central to concepts of the claim or of closely related concepts. For example, in:

Vaccine development is very expensive, the adjuvant is toxic,

expensive evaluates the property “cost” of the vaccine, which has the general properties of artifacts such as costs and functions. In the second claim, where adjuvant is a derived concept, which refers to a part of the vaccine that is injected, “toxic” refers to a more indirect property of medicines, for example, their “side effects”. Evaluative adjectives can also appear in comparative expressions:

electricity produced from nuclear energy is cheaper than electricity produced from coal or wind.

Nuclear wastes are more dangerous than coal wastes.

Boolean adjectives are less frequent since they tend to express a position that is more definitive: Ebola vaccination must be compulsory...

• – Verbs that have a clear positive or negative polarity, e.g. to protect, to prevent, to damage. This polarity introduces a flexible character that easily suggests discussions and debates.
• – Adverbs that focus on specific features such as necessity, obligation and possibility such as necessarily and inevitably where these features can be debated. Adverbs associated with the notions of frequency or completion are also frequent in claims since the level of frequency they induce can be debated, e.g. generally, almost never, seldom, rarely.
• – Modals and modal expressions that convey an idea of possibility or plausibility such as

might, should, would, could be and should be possible; the vaccine against Ebola should be compulsory in infected areas.

In Romance languages, conditional and subjunctive modes are also used to convey plausibility.

• – Forms such as why should that typically introduce rhetorical questions:

Why should citizens save money? To support SMEs.

In this example, the debatable segment is the response provided by the utterer.

• – Typical types of punctuations such as exclamation marks:

Let our committee make a decision!

Such a statement, if not ironic, could be followed by evaluations of various kinds on the committee’s ability to make decisions.

Based on the notion of evaluative expression given above, the following constructions are typical of simple claims:

1. 1) If the main verb of the claim has a clear positive or negative polarity (e.g. to protect, to prevent, to pollute), then language realizations of a claim include:
   1. (1a) a proposition with this verb,
   2. (1b) the negation of the verb or verb phrase (VP),
   3. (1c) the use of adverbs of frequency, completion, etc., possibly combined with a negation: in general, almost never, seldom, rarely, not frequently, very frequently, fully, systematically, or
   4. (1d) the use of modals expressing doubt or uncertainty: seem, could, should.

For example:

Vaccination prevents Ebola dissemination.

Vaccination does not prevent any disease dissemination.

In general, vaccination should prevent disease dissemination.

2) If the evaluative expression applies to the head noun of the claim, in general, the subject, then language realizations involve attribute structures with one or more adjectives with a clear polarity in the context at stake that evaluate the concept: toxic, useless, expensive or their negative forms. These can be modified by intensifiers such as: 100%, totally. These intensifiers can also generate by themselves a debatable character proper to claims. The main verb is often neutral such as to be or a light verb. For example, in:

Women’s rights are not totally respected in this part of the world.

“not totally respected” applied to the noun “rights” introduces potential controversies.

Control constructions, where head verbs subcategorize for a proposition or a sentential complement, may transform a statement into a claim by introducing some form of uncertainty. These control constructions may also appear in supports or attacks. In both cases, their scope is the claim or the entire argument. As far as claims are concerned, control constructions can be organized according to the following linguistic categories:

– Propositional attitude verbs and expressions that express a position. In this class, verbs are included and expressions such as:

think, believe, I am convinced that.

The semantics of these verbs is investigated in depth in [WIE 87]. Propositional attitude constructions can be modified by a negation or by a modal such as would, could, have to as in: I would think that, I have to believe that.

– Psychological expressions or expressions denoting a position or an experience. They include verbs and expressions such as:

I feel, I am worried about, I am intrigued by, dream of, be encouraged by, tend to.

These terms are often in an initial position with respect to the statement they introduce, or in a final position for constructions such as worries me as in:

the obligation of vaccination worries me,

where the nominalized sentence is raised to be the main focus of the sentence, it then becomes the grammatical subject.

– Performative verbs, which partly overlap with the other classes presented here. They clearly introduce the idea of a thesis or of a position that could be attacked. Verbs in this category are, for example, pretend, support, recommend, assume and claim. Their corresponding nominal expressions are, for example, my position, thesis and recommendation is. These cues explicitly indicate that the statement they subcategorize is a claim.

– Report verbs and associated constructions. They introduce claims in a direct manner or as a reported speech from, for example, other participants in a debate or from external persons, frequently considered as experts. Similarly to the above categories, these constructions can be modified by a negation or a modal. In this category, we include verbs and expressions such as report, say, mention, state, announce, discuss and claim and their morphological variants. For some of these verbs, it is not straightforward to transform a factual statement into a claim since they do not lend themselves easily to controversies, this is the case for report, say, announce and declare, which are rather neutral and need additional linguistic cues. Conversely, terms such as claim, affirm, maintain and allege are much more typical of claims and easily induce debates. The first utterance below is not as typical of a claim as the second one, in spite of the evaluative “dangerous”:

A doctor said that Ebola is a dangerous disease.

A doctor claimed that Ebola is a dangerous disease.

– Epistemic verbs and constructions. These also occur quite frequently. They include expressions such as:

know, my understanding is that, I am convinced that, I suppose, I realize, it is reasonable to assume, infer, imply, I can see, it is reasonable to, this can mean, this may mean.

These mainly introduce doubts while others are clear positive statements.

– Modal expressions. These are left-adjuncts to a proposition that modify some of the expressions described above. Most of them either weaken the statement or introduce a hypothesis:

it should be possible, it is probable that.

For example, the following claim and its justification:

it should be possible to have a systematic vaccination that creates sanitary belts to avoid the proliferation of the disease.

lend themselves to debates and controversies.

– Adverbials related to the expression of opinion. They indicate the strength of a claim but also its debatable character. They include:

probably, necessarily, most definitely, definitely, surely, usually, frequently, often, certainly, obviously, generally speaking, of course, indeed.

Finally, given a statement that is factual, a number of discourse structures adjoined to that fact may introduce a debatable character such as illustrations. It is not possible to develop in this section a systematic analysis of these discourse structures, however, a few relevant examples can be outlined to show the importance of this phenomenon.

Illustrations often take the form of an enumeration of elements to serve as examples. As a result, illustrations introduce some plausibility for supports or attacks because the elements in the enumeration can be used as starting points for these supports or attacks. For example:

the Ebola vaccine is easy to use for emerging countries (cheap, can be transported without any need for refrigeration, active for a long time).

Even if the main claim is consensual, some of the elements of the enumeration can be subject to debates, in particular when they contain fuzzy lexical items or expressions, such as for a long time. The term “cheap” could also be discussed.

Claims that include a purpose or goal may be subject to debates not on the main content of the claim but on its purpose. In:

Vaccination against Ebola is necessary to avoid any disease proliferation in territories close to those where Ebola is present.

the purpose can be debated in particular on the forms and speed of proliferation. Here, the purpose plays the role of a justification of the main claim. It can be attacked without really questioning the necessity of the vaccination, because it is not a very relevant justification or because it is rather weak. Similarly, the following purpose may weaken a factual statement:

[_purpose in order to avoid any form of panic or, worse, of bio-terrorism], the authorities of Guinea closed their borders.

For example, the permeability of the border can be outlined and discussed.

3.2. The linguistics of justifications

Identifying justifications to a claim is also very challenging both from a linguistic and conceptual point of view, in particular when justifications are not adjacent to the claim. In this section, we consider justifications to a claim C (supports) as well as justifications to its negation ¬C, which are attacks of C.

The first difficulty is to relate a potential justification to a claim. A justification may support a facet of a claim instead of the whole claim. The terms used in a justification may require inferences and knowledge to relate it to a claim because the linguistic, conceptual and referential links to that issue are rarely direct and explicit. This is one of the main challenges of argument mining. Justifications are often evaluative natural language statements that become justifications because of the specific relations they have with another evaluative statement considered as a claim. Besides their evaluative dimensions, justifications may be causally related to a claim.

Let us illustrate the need for knowledge and inference when relating a claim and one of its justifications with the following example:

Claim: The situation of women has improved in India,

Statement found in another text: (a) early in the morning, we now see long lines of happy young girls with school bags walking along the roads.

(a) is a justification (support) of the claim, but it requires knowledge and inferences to make explicit, and possibly explain, the relationships between women’s conditions and young girls carrying school bags. Here, the relationship is education. Let us now consider:

(b) School buses must be provided so that schoolchildren do not reach the school totally exhausted after a long early morning walk.

(b) is an attack of (a), indeed: these young girls may not be so happy to be walking in the early morning, but it is not an attack of the claim: the facet that is concerned in the relation between (b) and (a) does not concern women’s conditions in particular.

So far, the relatedness problem has not been addressed in depth in argument mining, although it is crucial to identify arguments. In Chapter 4, we develop a few examples that show the complexity of the problem, which is a major limitation to the development of argument mining technology.

The next difficulty is identifying that the statement that is related to a given claim is a justification. The term justification must be considered in a broad sense: a justification adds content that contributes to the acceptability of a claim. It makes the claim more plausible, but it does not prove that it is true. For example, a statement that is a reformulation or an elaboration of a claim is not a justification, since the content that is added simply makes the claim clearer or more precise. An illustration can be considered as a justification in some contexts:

Claim: Vaccination against Ebola is dangerous,

illustration: One person died after an injection in Monrovia and two patients got sick in Guinea last week.

The illustration does not really justify why the vaccination is dangerous, it however adds strength to the claim.

When a justification is adjacent to a claim or adjacent to a referential expression associated with that claim, it may be linked to that claim by means of different types of connectors, among which:

• – causal connectors such as because, since, as, so that, due to, resulting in
• – purpose connectors and expressions such as to, in order to, so as to, to avoid, to prevent can also be used to introduce supports as well as attacks,
• – connectors and expressions describing a result may also introduce a justification, for example: thus, therefore, as a consequence, hence, as a result,
• – illustrations may play the role of justifications: they are introduced by typical terms such as: for example, e.g., for instance, such as, like, including,
• – weak or strong forms of attacks can be characterized by concessive connectors such as: although, though, even if, despite, in spite of, however, nonetheless, nevertheless, or by
• – contrast connectors such as: while, whereas, but, in contrast of, yet, instead of, rather, unlike.

More details can be found in [ANS 83] where a large number of connectors with an argumentative orientation are presented and illustrated for French. An equivalent set can be developed for English. This investigation considers that argumentation is essentially a linguistic and language act rather than a discursive action. Claim–justification pairs are investigated from the perspective of their possible sequences in language.

Let us come back to justifications. Since they support claims, justifications are also evaluative expressions. Similarly, attacks support the negation of the claim at stake. They use some of the linguistic categories presented above for claims, in particular scalar adjectives, adverbs, verbs with a positive or negative polarity, report verbs, epistemic constructions and adverbials. The class of propositional attitude verbs is extended to take into account verbs that denote forms of acceptance of rejection such as:

agree, deny, argue, refute, acknowledge, reckon, disagree, accept, reject,

possibly combined with modals such as would, negation or temporal or frequency adverbs: I always disagree with...

3.3. Evaluating the strength of claims, justifications and arguments

In an argument it is of much interest to evaluate the strength of a claim and the strength of a support or an attack separately. Strong claims call in general for stronger forms of supports or attacks than claims of a moderate strength. Therefore, the strength of a claim or of a support or an attack can be evaluated separately. Then the evaluation of a support or of an attack must be tuned with respect to the a priori strength of the claim to determine the overall strength of an argument.

We view the strength of a claim, a justification or an argument as a kind of metrics, which is essentially induced by linguistic factors, even if an accurate measure of the strength of a number of linguistic cues is domain dependent and difficult to evaluate. It follows that persuasion is a more pragmatic measure of strength where personal and contextual factors play an important role. These factors include readers’ or listeners’ profiles and expectations as well as their global cultural environment. In this chapter, strength is viewed as a linguistic phenomenon. However, strength may also be determined by other factors such as images, gestures, mimics and intonation in interactive situations (see Chapter 9).

There are several ways to measure the strength of arguments. Strength can be measured from a logical and pragmatic perspective or it can be measured from a linguistic point of view. Both approaches are not necessarily coherent but they must be combined to produce a relatively accurate measure of strength. Argument strength may be measured for each argument in isolation or for groups of related arguments, taking into account their relations and structure.

Investigations on argument strength share some aspects with the evaluation of strength in opinion analysis. This latter area is well developed and a number of resources are freely available. We do not go into it in this book. However, the notion of strength in argument mining is more complex since it involves a number of specific features, including the impact of argument schemes.

So far, investigations on argument strength have focused on a few aspects such as (1) teaching how to organize written essays and how to organize arguments and give them an appropriate strength, (2) research on persuasion which is, in our view, an analysis of strength in contexts (the domain and the audience of the argumentation are taken into account) and (3) in theoretical analysis of argumentation where graphs of attacks and supports are developed. Let us note for example [GRA 11] that deals with an in-depth analysis of persuasion and [ZHA 11] which investigates the content of persuasive messages. Sensitivity to argument strength of various populations is developed in [COR 04]. The impact of strength in conjunction with rhetorical questions has been addressed in [MUN 88]. A number of linguistic factors are analyzed in [ANS 83], and later in [EEM 01] and [EEM 92]. The notion of strength is an important element in opinion analysis, for example, to measure whether a specific feature in a product is appreciated or not, and to what extent. A number of resources has been developed to evaluate strength, in general out of context. Let us note the Stanford Sentiment Treebank (https://nlp.stanford.edu/sentiment/treebank.html). This resource is a well-developed and stable resource frequently used in opinion analysis. However, to the best of our knowledge, little has been done to characterize argument strength from a linguistic point of view and from the perspective of argument mining.

This section is organized as follows. In the first section, the contribution of individual lexical items found in propositions P_i is investigated. The hypothesis is that such propositions have an intrinsic strength independently of any claim. Next, we introduce structures from lexical semantics that enable the organization of the linguistic data. Then, the strength variations induced by the combination of several lexical items in a proposition and the support construction in which it may be embedded are explored and tested experimentally. Since it turns out that contextual effects in their broad sense are crucial to have an accurate estimate of the strength of an argument, several contextual parameters are discussed, in particular the impact of the discourse structures that are adjoined to a statement and the kind of argument scheme on which the argument relies.

[[ruin] - [break, demolish] - [hurt, alter, lessen, undermine, damage] - [affect] - Neutral - [preserve, guarantee] - [benefit] - [improve, consolidate, strengthen] - [optimize]].

[[dangerous, harmful] - [risky, polluted] - [unsafe] - Neutral - [useful, beneficial] - [safe] - [certified]].

[[deny - refute - reject] - [disagree] - Neutral - [believe, think, accept] - [agree, acknowledge, reckon] - [argue]].

[[probably] - [indeed, usually, of course] - [often, frequently, generally speaking] - [definitely, surely, obviously, necessarily] - [most definitely]].

3.4. Rhetoric and argumentation

Rhetoric is an important component of argumentation and vice versa. Rhetoric, from a linguistic point of view, has developed models to organize claim justifications, argument supports and attacks in a structured way with the goal of being as efficient and convincing as possible. Planning issues are essential in rhetoric. For example, starting with the less prominent justifications and ending with the most crucial ones with which almost everybody agrees is a frequent strategy to convince an audience. This planning activity may also anticipate attacks from opponents by presenting them and refuting them. Rhetoric also deals with linguistic aspects such as lexical choice and syntactic structure. Finally, rhetoric offers a unified view of linguistic expression paired with non-verbal means such as gestures, mimics, various forms of sounds and images. Rhetoric is therefore a major component of argumentation in real contexts. In this section, the main features of rhetoric are introduced. There is an important literature on rhetoric, starting from [PER 58].