In this paper, we present a model to study rephrase in argumentative settings which encompasses an annotation scheme along with a rephrase analytics toolset to capture the dynamics of rephrase relation in conjunction with other communication layers such as appeals to ethos (speaker's character) and expressed sentiment (speakers' emotions). To this end, we study posts on social media, TV presidential debates and parliamentary debates. The proposed account of corpus and analytical devices to study language of rephrase allow for a new approach for mining rephrase in natural language texts.
In this study, only the rephrase (MA) of the whole IAT layer is used. More specifically, we investigate the speaker’s initial statement (input), e.g., CLINTON is going to increase regulations all over the place, that has been rephrased by another or the same speaker (output), e.g. CLINTON is going to regulate these businesses out of existence. The input and the output may carry slightly diferent messages, in particular they may express: (i) either ethotic support, see Example (1), or attack, Example (2), i.e. a positive/negative description of the skills, charisma, virtues, character of the speaker or the institution: E+ or E-, respectively; and (ii) either positive emotions, see Example (3), or negative emotions, Example (4), i.e. positive/negative sentiment of the speaker(s): S+ or S-, respectively. In other words, input and output can have diferent ethos and sentiment values. (1) (2) (3) (4)
olivecorgi7: New York is falsifying data [Label E-]
[Label S-]
Building upon an AI-based tool: Dynamics of Rephrase Analytics (DynRephAn) [
Preliminary experimental findings in studying the persuasive efects of rephrase on the
audience with crowd-sourced experiments have shown that rephrase – though not being as much
persuasive as arguments are – is both be quite frequent in the pilot corpora and have a
significant impact on changing audience’s beliefs [
Existing models and approaches to rephrase encompass both accounts of rephrase relation to
inference and argumentation, and the study of specific rephrase uses. In [
The phenomenon of rephrasing, thanks to its relational nature, gives us the opportunity to
compare original utterances with modified ones, and thus provides us with an excellent research
laboratory for tracing the dynamics of ethos and sentiment as key communicative phenomena
carried by rephrase which might be decisive in successful persuasion. Rephrase is typically
conceived in terms of relation between two text spans we call rephrase input and rephrase
output, where the latter, while keeping the core content of a former, also introduces some
novel information [
The study of the dynamics of ethos in rephrase is aimed at tracking the dynamics of ethos across corpora in order to gain insight into the scale of how rephrase ‘behaves’ in terms of adding or subtracting ethotic load to messages. Consider Example (5), in which Anderson Cooper’s statement in (5-a) about Lincoln Chafee’s being Republican in the past is rephrased in (5-b): (5)
Given that in this particular context of the democratic Preliminary Debate, the fact of being Republican can be typically perceived as an accusation and constitute an attack on the candidate’s ethos, Chafee’s response can be interpreted as an ethos support (being a ‘liberal Republican’ helped Chafee to defend his position as being closer to Democrats even back then). This example can be thus discussed in terms of using rephrase to turn an alleged ethotic attack done in (5-a) towards an ethotic support in (5-b), and so be an instance of dynamical role of rephrase that makes a shift from positive to negative ethos (we called this dialogical move a pejorativisation).
A clear instance of ethos dynamics as carried by rephrase is illustrated with Example (6):
pletentious_asshore: He’s (Donald Trump’s) not answering the question about how to bring jobs back [Label E-]
Pokemongotrainer: he (Donald Trump) never did [Label E-] In (6-b), the Reddit user, Pokemongotrainer, makes an ethotic attack on Donald Trump, introduced in the input (6-a), stronger, as claiming that Trump has never answered the question about bringing jobs back. Hence, the intensification of the attack on Trump’s ethos can be identified with the use of the general quantifier ‘never’.
Another rhetorical device used in rephrase is to bring into an output either not present or more intense emotional load than in the case of an input (see Example (7)). We say that rephrase in such cases serve speakers as a positive or negative sentiment carrier, depending on specific rhetorical gains that are to be achieved.
Sentiment, added in rephrase output, quite typically overlaps with either adding or intensifying the previous attack on or support of ethos. The issue of the overlap between ethos and sentiment on the one hand constitutes an additional challenge for designing the mining tasks (see Section 4.1), but on the other hand, may serve as a promising research field to explore e.g. how rephrased arguments convey ethotic attacks or supports done by means of emotionally loaded words and phrases. In Example (8), Sanders, by elaborating on Kang’s view on the relations between Obama and the Republicans, conveys in (8-b) a clear negative sentiment carried with the word ‘terrible’ (used twice), and the phrase ‘total obstructionists’.
S-] This move, being at the same time an ethotic attack on the Republicans, relies on conveying sentiment by adding a lot more linguistic material than the one used in (8-b). Even a greater elaboration on the content of an input can be observed in example (9):
have so much and so many have so little by which Wall Street’s greed and recklessness wrecked this economy? [Label S-] In (9-b), Sanders, by elaborating on the content of Cooper’s assertive question (9-a) (which can be reconstructed as: “Sanders does not consider himself capitalist”), adds his, rich in emotionally loaded words, assessment of capitalist’s practices. Such a dialogical move of adding negative sentiment is typical for the US 2016 Presidential Debates record.
In this section, we explore the linguistic features of rephrases, which contain ethos and sentiment elements, that can be accounted for on the basis of relatively simple language factors with available linguistic resources. Still, their number is significant and involve pragmatic factors or complex linguistic constructs, thus their modelling is beyond the scope of this paper. 4.1. Language challenges Let us briefly evoke the underlying challenges to rephrase modelling and automatic analysis. These are important to understand how rephrase in conjunction with emotion and ethos can be characterised. First, rephrase may be quite dificult to identify. There is indeed a kind of continuum between cases where, e.g. the rephrase content is rather weak or shallow w.r.t. the original statement. Annotator decisions are therefore prone to subjectivity.
Next, the speaker profile or the context must be considered to tune rephrase characteristics. In a political debate for example, even if positions are similar, the language used by politicians may significantly vary which pose a challenge to the identification of their position and attacks on other participants.
Third, even simple rephrases may lead to complex forms of analysis, consider: The campaign was aggressive rephrased as: the campaign was very aggressive but with no systematic obstructions. Does the contrast (after ‘but’) introduce a negative ethotic dimension to the rephrase and if so, how is the rephrase oriented?
Fourth, a standard statement often has several facets, due to the lexical semantics of the diferent terms it contains. Some facets may be more crucial than others in argumentation in general. A rephrase often concerns one of these facets. Facets can be implemented, by example, by the Qualia structures of the head terms of the original utterance. Facets may describe, for each head term, their uses and functions (telicity), how they came into being (agentivity), their components, etc. Consider for example: (10)
some communities Here, (10-b) elaborates (10-a) by supporting the statement ‘but’ with the introduction a consideration on an agentive feature ‘the cost’ with a negative orientation. The other facets of vaccination are not addressed. Another rephrase could concentrate on e.g. the toxicity of the adjuvant (vaccine component), or on some political aspects (making vaccination compulsory for adults). These examples illustrate the dificulty to identify rephrases and to categorise them (orientation and strength) and their associated ethos with some accuracy. The examples given in this paper outline the impact of language factors as well as of domain knowledge. This would be extremely dificult to handle for an automatic system.
From a more technical perspective, the extreme diversity of language uses found in our corpora makes it dificult to foresee the development of large repositories of annotated rephrases to be used as entries of automatic learning algorithms (confirmed in Table 1 in Section 4.3 below, where we show a rephrase rate of 15.2%). This is true even in hybrid contexts where corpus data is paired with language analysis. Our approach in this research is first to observe how the diferent types of rephrases are realised in diferent types of corpora, to annotate them, and then to develop tools which would allow humans to develop the linguistic characteristic of rephrase in various utterance contexts. Rephrase mining remains our ultimate goal, in conjunction with ethos and sentiment, but in the long term, when abstract models are established and tested. 4.2. An approach to modelling rephrase in analytics Given the above considerations, our strategy is twofold: (i) to develop relatively abstract principles which account for the relation between an initial statement and subsequent statements identified as potential rephrases with their ethotic or sentiment features; and (ii) to use already existing or develop additional linguistic elements (lexical, syntactic) which could serve as indices of rephrase situations and as the basis of a future automatic recognition of rephrase.
The first point is realised by human analysts on the basis of quite large corpus annotation. A theoretical framework could be developed based on principles describing how rephrase is linguistically (and probably cognitively) elaborated in conjunction with ethos and sentiment. However, these principles will probably never be comprehensive (necessary and suficient) and more concrete considerations such as those advocated in point (ii) must be considered, leading to a kind of hybrid system pairing abstract principles with linguistic considerations. Statistical data will need to be introduced to indicate preferences observed in corpora.
An exploratory analysis of these linguistic considerations shows that most of the linguistic
levels are concerned in rephrase, the main ones being lexical semantics and pragmatics. Syntax
probably plays a more marginal role, due, for example, to poor language uses, ellipsis, gaps of
various sorts, etc. To identify and organise the linguistic aspects which are crucial to identify
rephrases and ethos and sentiment, a DynRephAn tool has been developed [
The first three datasets comprise online discussions on X – Twitter on COVID-19 vaccines
(PolarIs1); on X – Twitter on Climate Change (PolarIs4); and reactions on Reddit to televised
presidential debates in the US in 2016 (US2016reddit). The US2016tv dataset comprises transcripts
of 3 televised presidential debates in the US from 2016 [
Polaris1 Polaris4 US2016tv US2016reddit Hansard Total
The observed rephrase rate is 15.2%, that is, among the ADUs (Argumentative Discourse
Units) 15.2% are identified as rephrases of others. Now, if we consider the total number or
proand con- arguments, rephrases represent 29% of this total, which underlines their importance.
Similarly, if we consider the total number of ethotic supports and attacks, these are present
in 55% of the total number of rephrases (therefore 45% are present in other statements). The
confidence level for annotator judgements (how certain they feel about their decision) is about
0.57, this is in particular due to numerous borderline cases. However, the average inter-annotator
agreement over these corpora is 0.77 which is quite high for such a task.
4.4. Evaluating the lexical level to characterise ethos and emotions in rephrase
Let us concentrate in this contribution on the lexical level used to produce rephrases and ethotic
and emotive structures in our corpora. As already indicated in [
Rephrase is basically characterised by two dimensions: (i) the context: an original statement shares quite a lot with its rephrase, in particular verbs and nouns are frequently similar, or in a synonymy relation; and (ii) the lexical essence of the rephrase, which characterises what has evolved and how, in a positive or a negative direction and along what semantic dimension (or facet). This is, in simple cases, relatively frequently accounted for by the use of scalar adjectives of various types or intensity adverbs modifying verbs or possibly adjectives. It is also possible to observe, in more limited cases, nouns and verbs, which, while preserving the context advocated in (i), worsen or improve the original statement.
From an implementation perspective, in lexical semantics, scalar adjectives and intensity
adverbs are structured on the basis of non-branching proportional series [
frozen - cold - mild - hot, warm - torrid.
Other adjective series are more subtle to structure, they may depend on context or on the speakers, e.g. for a political discourse (or a research project):
awful - terrible - bad - drab - average , acceptable - good - excellent.
The position of terms on a given series has an impact on the rephrase orientation and on ethos.
For adverbs, very, dramatically, etc. increase the initial intensity of the verb whatever its
orientation. This data is well-known and accessible on several platform, most notably WordNet
and Sentinet. Results produced by our platform include these data repositories and more specific
data collected from our corpora [
For those rephrases which do not enter into the above scenario, the following experiments are being carried out: • consider pairs where the rephrase is a form of negation of the initial statement, with the use of an explicit negation, • similarly, consider pairs where the rephrase contains a verb which is an antonym of the head verb in the initial statement, • consider pairs where the original statement remains unchanged but a restriction (e.g. a contrast) has been adjoined via a dedicated connector (e.g. but, nevertheless or a comma), • consider those pairs where the rephrase contains a noun which is in the same hyperonymy class of a noun in the initial statement instead of the initial noun.
In order to facilitate the linguistic analysis of the dynamics of ethos and emotions of rephrase, a
DynRephAn_v02 tool (Dynamics of Rephrase Analytics version 02, available at https://newethos.
org/technologies/) has been developed [
The current version extends the implementation to allow for the analysis of linguistic properties of pairs of argumentative discourse units that are linked through the relation of rephrase. DynRephAn_v02 enables to search for (i) selected n-grams (unigrams, bi-grams, etc); (ii) PartOf-Speech; or (iii) their combination (see Figure 2). This allows for understanding the rhetorical mechanism of rephrasing arguments in discourse, and at the same time gives the perspective of enhancing the automated argument detection through the exploration of the characteristics of the linguistic surface and linguistic patterns that accompany such an argumentation.
In this paper, we explored rephrase dynamics using five corpora with discussions on Covid-19 and climate change on X – Twitter, televised US 2016 Presidential Elections Debates, Reddit reactions to those debates, and UK Parliamentary debates from the 1979-1990 period. The study of the relational character of rephrase allowed us for developing rephrase analytics tool, DynRephAn, to capture the linguistic changes of ethos and sentiment when speakers rephrase their own or someone else’s words. Whereas the ethos layer has been annotated manually, an automatic annotation was done for the sentiment dimension. The emphasis in the design of the DynRephAn tool, which has been put on the linguistic aspect of the analysis, makes it possible to automatically obtain visualisations showing which specific expressions (parts of speech or sequences of typically used words, called ‘n-grams’) are associated with the dynamics of the rephrase in terms of amelioration, pejorativisation and neutralisation. In this way, we have provided a tool to further study the language of rephrased arguments for the sake of the future inquiry into tendencies of modifying the meaning to achieve persuasive aims.
A possible line of future inquiry should, for instance, focus on the lexical characterisation of ethos which, despite being largely dependent on pragmatic factors, may be studied in terms of some elements at the lexical level that may play a prominent role independently of any pragmatic consideration. For example, using in a rephrase an adjective or an adverb which is very positive or negative (to the extreme right or left in a non-branching proportional series) may indicate e.g. a form of excessive support (positive), or aggressive behaviour (negative orientation probably). Also, using terms which are quite far from the original ones may be analysed as such. This may also be the case for verb antonyms. Conversely, a rephrase where the lexical variations are very limited in intensity could indicate a rather consensual behaviour or consideration for the other speakers. In that case, the terms used in the rephrase are close on the series considered. The facets considered in the rephrase may also be of importance to characterise ethos. There is a diference between facets which are central (such as those associated with the telicity) and those which are more peripheral (such as those associated with the agentivity). Then, for example, a rephrase which refers to a peripheral facet may be interpreted as more negative or aggressive, but this remains an open research problem. These and other aspects of the language of rephrasing, once implemented in the rephrase analytics tool, may turn out to be instructive in identifying and making sense of sophisticated rhetorical strategies of the use of rephrase with (un-)loading messages with ethos and sentiment, and thus complement to the pragmatic characteristics of manoeuvering with rephrased arguments.
The work reported in this paper was supported in part by the Polish National Science Centre under grant 2020/39/I/HS1/02861 and in part by VW foundation (VolkswagenStiftung) under grant 98 542.