Proceedings of CMNA 2016 - Floris Bex, Floriana Grasso, Nancy Green (eds) How Natural is Argument in Natural Dialogue? Shauna Concannon, Patrick Healey, Matthew Purver Queen Mary University of London s.concannon@qmul.ac.uk as academia, politics or law and online data from product review and debate sites. In contexts such as legal or parlia- Abstract mentary debate stylised language, rhetoric and persuasion are employed, and arguments are typically prepared in ad- Exposed disagreement is extremely rare in natu- vance. In natural dialogue it is often through the process ral dialogue. Although informal argumentation of dialogue that individuals come to know and refine their features frequently in natural dialogue, the ways own opinions, as well as those of others, making natural di- in which individuals make and evidence claims alogue a particularly rich source for understanding opinion and position their opinions in relation to those formation. In these aforementioned contexts the expecta- of others is often achieved through more subtle tion is established that opinions will be freely expressed and oblique methods. This makes natural dia- and there is no social obligation to mitigate the impact of logue distinct from more formal or institution- exposing contrary opinions. In natural dialogue, this pre- alised contexts. With increasing availability of defined expectation for argumentation is often not present, natural dialogue datasets and with increasingly di- and the implications of challenging another person’s opin- verse contexts within which the application of ar- ion can be potentially problematic. Social interactions in- gumentation modelling could be beneficial, be- volve the management of a person’s public self image, or ing able to identify and interpret argumentation face, in Erving Goffman’s terms. in natural dialogue becomes more important; so In order to access the abundance of informal argumen- too does an understanding of why argumentation tation that is increasingly taking place on the social web, is enacted differently in natural dialogue and how closer attention should be paid to how opinion, agreement factors such as politeness impact upon this. In and disagreement are enacted in natural dialogue. In par- this paper we highlight some of the ways in which ticular, we suggest that a starting point is empirical studies argumentative content is produced differently in of face-to-face dialogue. Furthermore, as emerging appli- natural dialogue compared to formalised debate cations of online technologies are used in ever diverse con- contexts and highly structured documents. We texts in which inter-personal relationship management is present some initial findings that demonstrate how important, such as health care dialogues, understanding the existing models such as the Penn Discourse Tree- social dynamics of dialogue and disagreement is ever more bank need further development if they are to adapt crucial. to the more dialogic data created on the social web. This work contributes to the existing literature on Com- putational Models of Natural Argument by addressing how the processes of disagreeing with a conversational partner 1 Introduction is executed in natural dialogue. We demonstrate that ex- In natural dialogue individuals take to care to make state- plicit disagreement is quite rare in natural dialogue and ments in such a way as to not cause offence, especially highlight some of the more implicit mechanisms that are when presenting a stance that may be contrasting or chal- used to position a stance as oppositional, and achieve lenging to another speaker’s prior contribution. Exposed disagreement without enacting disagreement in the more disagreement is rare in natural dialogue and the ways in recognisable forms. We discuss how politeness theory which individuals present their own and others’ positions can guide our interpretations of interactions and demon- on a given topic are influenced by efforts to maintain po- strate the interactional significance of paralinguistic fea- liteness. tures, such as hesitations and disfluencies. Finally, we Computational modelling of argumentation has typi- present some preliminary findings on how discourse rela- cally drawn on textual data from institutional contexts such tions manifest differently in natural dialogue compared to 43 Proceedings of CMNA 2016 - Floris Bex, Floriana Grasso, Nancy Green (eds) news articles. the large problems that are increasingly effecting our so- ciety’ [15]. The need for new and adapted approaches for argumentation on the social web has been acknowledged 2 Related work [14, 15]. Data on the social web is often more closely Classifying and extracting argumentative content automat- aligned with conversational dialogue in structure than writ- ically has been demonstrated in such contexts as parlia- ten text; consequently, the importance of developing sys- mentary debate [11], legal documents [10], news articles tems that can interpret incremental, fragmentary and col- [9] and online debate forums [1, 3]. While there is some loquial content is essential. In order to create intelligent work addressing dialogic argumentative interactions, the systems that can interpret a wider range of strategies used focus so far has been on highly structured argumentative by people in the process of argumentation, that will apply texts. Previous work has shown that discourse relations are to multiple contexts beyond formal argumentation contexts closely related to argumentative relations, most notably led such as law, we need to better understand the way in which by the creation of two annotated corpora, the Rhetorical argumentation is performed in everyday contexts. This is Structure Theory Discourse Treebank (RST-DT) [5] and particularly essential as argumentation begins to spread to the Penn Discourse Treebank (PDTB) [13]. Such anno- diverse contexts such as pedagogy, health consultation and tated corpora have been valuable resources for training au- e-democracy. tomatic classifiers, but as the source material for both is news articles, how useful they will be for natural dialogue 2.1 Politeness and social conventions is unclear. Recent consideration of how to develop effective approaches to argumentation on the social web, has empha- Qualitative studies show that exposed disagreement is gen- sised that dialogue is structured differently, and warns that erally avoided in conversation [12]. This is normally at- meaning may be lost if messages are extracted individually tributed to politeness strategies that mitigate potentially and out of context [14]. face threatening behaviour [4]. Disagreeing or expressing Furthermore, as we will demonstrate in this paper, ar- a view in opposition to that of your interlocutor can be so- gumentation in natural dialogue relies much more heav- cially problematic. Brown and Levinson [4] explain the ily on vague and implicit arguments, which are challeng- predisposition for the avoidance of disagreement in terms ing to identify through existing argument mining methods. of face, i.e. the public self-image or identity of an indi- Machine learning approaches, such as [2], which include vidual in interaction with others [8]. Direct challenges to textual entailment, stance alignment and semantic textual a speaker can constitute a Face Threatening Act, i.e. it can similarity analysis have gone some way to improve per- threaten the hearer’s public identity. Conversation Analy- formance, but are typically applied to highly structured sis (CA) has also shown that when people produce initial datasets, i.e. forum posts labeled in support or attack of assessments of situations or events, positive responses are a given argument. Argumentation in ‘online dialogue’ [1], made more quickly and clearly than negative or unaligned although arguably more closely aligned to natural dialogue responses. Negative responses are normally produced more as the content is generated on forums by those not specif- slowly and are often prefaced with some form of agree- ically trained in rhetoric and debate, is still distinct in a ment (e.g. ‘Oh yes... but’); the negative assessment is often number of ways: the structured post-and-response format, delayed by several turns and produced with some sort of the time available for formulation and consideration before mitigating account [12]. Although research has shown that publishing, and (in many cases) explicit meta-tagging of incivility occurs more freely online, the negative social im- content as ‘support’ of ‘attack’ of an argument. A cor- pact of exposed and unmitigated disagreement persists in pus study highlighted that the markers of agreement and computer mediated dialogues between acquaintances [6]. disagreement employed in the Internet Argument Corpus, were very uncommon in naturally occurring conversation 3 Argumentation in natural dialogue [7]. In natural conversational dialogue data the strategies How people enact disagreement is socially important, and used in argumentation are likely to be more diverse and less more often than not it is achieved through subtle means. formalised. If argumentation frameworks can also account Politeness theory suggests that interlocutors employ strate- for and adapt to such data, they will be applicable to more gic conflict avoidance techniques to mitigate the effect of contexts; in order to tap into the wealth of data available any disagreement that may surface. Care is taken to make via social media and other online sources, it is necessary to disagreement indirect, thus making a rubric for identifying adapt argumentation models for the conversation of the lay disagreement challenging. commentator, not just the trained professional. The web is an increasingly social space, in which huge 3.1 The span of disagreement in natural dialogue quantities of informal interactions are captured, many of which feature argument structures and ‘could provide real In natural dialogue, because of the preference to minimise insight into the stated beliefs and reasoning of people into disagreement and emphasise agreement, speakers often de- 44 Proceedings of CMNA 2016 - Floris Bex, Floriana Grasso, Nancy Green (eds) A: D’yuh li:ke it? produce a response to a previous assessment, if the con- (+) D: .hhh Yes I do like it= (-) tent is positive it is made more quickly and directly than if D: =although I rreally::= it is an unaligned response that might challenge the prior C: =Dju make it? speaker’s face. Negative or dispreferred responses are typ- A: No We bought it, It?s a .hh a Mary Kerrida print. ically prefaced with a delay or an agreement token [12]. D: 0:h (I k-)= Consequently, argumentative content can span quite a num- A: =Dz that make any sense to you? ber of turns in a dialogue, and failing to consider this fully C: Mn mh. I don’ even know who could lead to misinterpretation and false classification of she is. stance. Disagreements can be socially problematic and A: She’s that’s, the Sister so speakers often delay issuing contrasting or challenging Kerrida, who, propositions. This can be signalled through turn initial hes- D: Oh that’s the one you to:ld me itations, disfluencies and discourse markers, or by prefac- you bou:ght.= ing any disagreement content with an agreement. This can C: Oh- make automatic extraction of disagreement from natural di- A: Ye:h alogue extremely challenging. D: Ya:h. A: Right. Consider example 1; in this transcription, Evaluation (1.0) of an artwork, taken from (JS:I. -1) [12], participant A is A: It’s worth something, inviting the others to provide their opinions on the artwork (1.0) at which they are currently looking. Critical assessments A: There’s only a hundred of’m are indicated in the transcript by Pomerantz with a ‘-’ sign, (0.5) while a ‘+’ sign indicates a positive assessment. The way in D: Hmm which A structures their questions, ‘D’yuh li:ke it?’, con- E: which picture is that. strains the range of appropriate responses to a polar yes/no A: The one that saysLife. response. D, although issuing a slight hesitation (as indi- (1.5) cated in the transcript as ‘hhh’), provides a positive appre- A: ( ). (-) D: ‘hhh Well I don’t- I’m not a ciation in the turn directly following the initial question. great fan of this type of a:rt. Notably, this is followed by the contrastive conjunctive ‘al- There are though’, which initiates D’s next turn, and provides some certain ones I see thet I like, indication that they have more to add on this subject. How- But I like the w- = ever, it is not until some 18 turns later that D manages to E: =Is there ano thu way of contribute that they are ‘not a great fan of this type of art’. spelling Life?. In the final turn of the example D explains that that they (-) D: -more realistic-. find it reminiscent of a magazine advertisement, and state A: hhmh! that their taste in art is more realistic. Without ever directly E: That’s all I wd loo(hh)k fo(h), saying that they do not like it, it becomes clear that they D: hh! don’t despite having explicitly said that they do. (-) D: Yih d-know why don’t got fer this type of uh: art, Becuz A great deal of conversational context must be taken into it- it account in order to identify the position each speaker is tak- strikes me ez being the ing. The polar interrogative that A initially offers, leaves D magazine adverti:sement yt:pe. with the choice of being polite, and providing the preferred Which some response, or offering a more accurate but dispreferred re- uh-uh some a’ them are really sponse (i.e. that she doesn’t like the artwork), which di- great. But tuhm I-my, taste in rectly positions her in opposition to her interlocutor. As art is this example highlights, offering an opinion can be signifi- for the more uhit-t-treh- it cantly affected by the social factors of the interaction. If we tends tuh be realistic. had considered only the first two lines a different summary of the discussion would have been concluded (example 2): A: D’yuh li:ke it? Example 1: Evaluation of a new artwork from (JS:I. -1) D: .hhh Yes I do like it= (-) [12] lay the delivery of dispreferred responses. CA is an ap- proach that without introducing additional theory, looks at Example 2: Detail of Evaluation of a new artwork from language used by the speakers to interpret the sequential meaning of the language. CA has shown that when people By examining only this segment we could conclude that 45 Proceedings of CMNA 2016 - Floris Bex, Floriana Grasso, Nancy Green (eds) A and D both like the painting. D’s response taken in iso- A: cause those things take working lation could lead to erroneous analysis; if the full context at, of the dialogue is included, then a different interpretation is (2.0) B: (hhhhh) well, they do, but possible. However, even if we manage to extract all of the A: They aren’t accidents, propositional content from the dialogue, it is still difficult B: No, they take working at, But to make a accurate interpretation (example 3). on the other hand, some people A: Do you like it? are born with uhm (1.0) well D: Yes I do like it. I’m not a a sense of humor, I think it’s great fan of this type of art. something you are born with It strikes me as being the Bea. magazine advertisement type. A: Yes. Or it’s c- I have the- Some magazine advertisement eh yes, I think a lotta people type art is great. But my, are, but then I think it can be taste in art is for the more developed too. realistic. Example 4: Example of a disagreement from [12] Example 3: Summary of Evaluation of a new artwork from From example 3, it would be valid to conclude that D likes this artwork, although in general they are not a fan of this style of work as they prefer more realistic art. ‘Yes I ing room to avert it, that is, the conversant can resume with do like it’, is direct and seemingly unequivocal; thus, when a modified assessment that may lead onto agreement. As interpreting it alongside the summarised content, it carries such, there are times when honest appraisals are simply not more weight and seems more directly connected to the orig- a part of interaction: ‘It is not only that what would be a inal question than what follows. However, when consider- disagreement might not get said, but that what comes to ing the full transcript, the dialogue reads quite differently, be said may be said as an agreement’ [12]. In addition to and the likelihood that D simply says they like it out of po- hesitation, speaker B also uses the discourse marker well liteness, before providing an account for why they don’t, in line 3. A turn-initial well typically (but not exclusively) seems much more plausible. This example highlights the indicates that a disagreement is forthcoming or what fol- importance of paralinguistic features, such as hesitation. lows will be in some way contrary to a prior statement Before D asserts that they do like the artwork they issue [12]. Speaker B performs an initial agreement, signalled a breathy hesitant delay. While this may seem like noise in through a turn-initial No (typically regarded as a marker the data, it is actually an important indicator that D is strug- of disagreement) and a repeat back ‘they take working at’, gling to formulate and appropriate response. Such paralin- before delivering a contrasting point of view, namely that guistic content can prove vital to an accurate interpretation certain traits are innate. In response, speaker A begins with of the interaction. a token agreement, chiming in with accord, before revert- Making and responding to assessments and assertions ing back to their previous, contrary stance: ‘I think it can occurs frequently in natural dialogue. When responding to be developed too’. By adding ‘too’ at the end of the utter- an initial assessment, an agreement may be signalled by ance, it enables A to maintain their line of argument while repeating back the original assessment, but subtle details conceding to the possibility that they both could be right, such as whether it is an exact repeat or a modified repeat thus mitigating any face threat and enabling the difference can signal whether it is a strong agreement or weaker vari- of opinions to be left unresolved. ation, modifying or downgrade the original assessment or even acting as a disagreement. Example 4, taken from [12], These two extracts highlight many of the devices, such illustrates a disagreement. A pause and delay, ‘(hhhhh) as hesitation, negation, and discourse markers, that are em- well’, is inserted, followed by a partial agreement, before ployed when managing disagreement in natural dialogue. the contrastive conjunctive ‘but’ is uttered, signalling that They also demonstrate how a disagreement can be withheld this is not in fact an agreement. Such mechanisms enable initially and argumentative content can span across multi- the speaker to take some time to formulate their disagree- ple turns, making the process of delimiting relevant context ment, to search for a tactful way to deliver it, and prevent problematic. The importance of context is evident through- the response coming across as blunt or aggressive. out; the turn-initial ‘no’, without the consideration of the Pomerantz highlights that people have a tendency to previous turn, which features a negative verb (aren’t), could minimize disagreements; respondents to initial assessments easily be misleading, but example 1 demonstrated that con- employ backdowns to hint at disagreement while still leav- text often spans more than adjacent turns. 46 Proceedings of CMNA 2016 - Floris Bex, Floriana Grasso, Nancy Green (eds) 4 Studying disagreement in natural dialogue and imagine, are a key resource through which individu- als present their own stance in conversation. If we take the The CA observations, as demonstrated by the examples in propositional attitude verb think, there are 14264 instances section 3, highlight the ways that people normally make in 150 of a total 153 files; think occurs in nearly every con- effort to avoid exposing disagreements directly (unless of versation file in the demographic portion of the BNC. Be- course they intend to be abrupt or confrontational). In nat- lieve, while slightly less common, still features in 116 of ural dialogue the presentation of opinions, evidence and 153 transcripts. While the espousal of internal states and counterclaims, are not always marked as agreement or dis- subjective positioning may be inappropriate in a news ar- agreement, rather they often remain implicit and can span ticle, in dialogue it is a key resource for positioning your over many turns of talk. In addition to this, dialogue is argument, and can also act as a form of hedging, helping fragmentary and metalinguisitic features (e.g. discourse face management. These very preliminary insights suggest markers) can be highly context dependent, making mod- that more detailed investigation of how argumentation in elling argumentation in natural dialogue particularly chal- natural dialogue is marked could be very useful, particu- lenging. One alternative approach may be to include more larly if applied to CMC data from informal contexts. lexical features that relate to stance and politeness in com- putational models of argumentation. These linguistic fea- tures are particularly important in dialogue as they enable 5 Conclusion a speaker to position an utterance in opposition to a prior In this paper we have highlighted some of the ways in proposition without necessarily enacting a direct challenge which argumentative content is produced differently in or disagreement. Better understanding the ways in which natural dialogue compared to formalised debate contexts. individuals construct argumentative content, whilst still ad- Some initial findings were presented that demonstrate how hering to norms of politeness, could be extremely bene- existing models such as PDTB need further development ficial for computational argumentation models of natural if they are to adapt to conversational data created on the dialogue. If we look to face-to-face dialogue as a start- social web. We emphasise the importance of considering ing point, and think more about politeness and the socially social factors, such as politeness, when modelling disagree- problematic aspects of dialogue we may be able to under- ment in natural dialogue and offer some potential ways to stand the challenge at hand better and approach it in a more interpret and account for this in interactional data. sophisticated manner. Two main objectives that our future work will set out to Acknowledgments achieve therefore, will be: to develop a more robust frame- work of what argumentation does look like in natural di- This work was funded by EPSRC through the Media and alogue, and to explore the limitations of existing models. Arts Technology Programme, an RCUK Doctoral Training In order to establish whether existing argumentation mod- Centre EP/G03723X/1. els are less suited to natural dialogue a preliminary corpus approach was developed using the British National Corpus References (BNC). Although these are very preliminary results, they provide helpful indicators of some of the most crude dif- [1] Rob Abbott, Marilyn Walker, Pranav Anand, Jean E ferences between natural dialogue and more formal debate Fox Tree, Robeson Bowmani, and Joseph King. How data. can you say such things?!?: Recognizing disagreement in informal political argument. In Proceedings of the An initial cursory comparison of the ‘spoken demo- Workshop on Languages in Social Media, pages 2–11. graphic’ sample of the BNC with the Penn Discourse Tree Association for Computational Linguistics, 2011. Bank (PDTB) corpus highlights some of the differences observed in natural dialogue when compared to news ar- [2] Filip Boltuzic and Jan Šnajder. Back up your stance: ticles. Discourse connectives are key to interpreting argu- Recognizing arguments in online discussions. Citeseer, ment structure and deducing the relationship between ar- 2014. gument content items. The explicit connective for example does not occur at all in the spoken demographic portion [3] Filip Boltuzic and Jan Šnajder. Identifying prominent of the BNC, but occurs 196 times as an explicit connec- arguments in online debates using semantic textual sim- tives in the PDTB. In the PDTB corpus, assertion propo- ilarity. In Proceedings of the 2nd Workshop on Argu- sitions, indicated by communicative verbs (say, mention, mentation Mining, pages 110–115, 2015. claim, argue), account for the overwhelming majority of relationships between agents in the corpus (98%), with the [4] Penelope Brown and Stephen C. Levinson. Politeness: other types (beliefs, factives and eventualities) occurring Some Universals in Language Usage. Studies in Inter- very infrequently. However, in the BNC, propositional at- actional Sociolinguistics. Cambridge University Press, titude markers, such as think, believe, feel, expect, suppose, 1987. 47 Proceedings of CMNA 2016 - Floris Bex, Floriana Grasso, Nancy Green (eds) [5] Lynn Carlson, Daniel Marcu, and Mary Ellen Okurowski. Building a discourse-tagged corpus in the framework of rhetorical structure theory. In Proceedings of the Second SIGdial Workshop on Discourse and Dia- logue - Volume 16, SIGDIAL ’01, pages 1–10, Strouds- burg, PA, USA, 2001. Association for Computational Linguistics. [6] Shauna Concannon, Patrick GT Healey, and Matthew Purver. Shifting opinions: An experiment on agreement and disagreement in dialogue. In Proceedings of the 19th SemDial Workshop on the Semantics and Pragmat- ics of Dialogue (GoDial), Semdial Workshop, 2015. [7] Shauna Concannon, Patrick GT Healey, and Matthew Purver. Taking a stance: a corpus study of reported speech. In Proceedings of the 19th SemDial Workshop on the Semantics and Pragmatics of Dialogue (GoDial), Semdial Workshop, 2015. [8] Erving Goffman. Interaction ritual: essays on face-to- face interaction. 1967. [9] Eleni Miltsakaki, Rashmi Prasad, Aravind Joshi, and Bonnie Webber. Annotating discourse connectives and their arguments. In Proceedings of the HLT/NAACL Workshop on Frontiers in Corpus Annotation, pages 9– 16. Boston, MA., 2004. [10] Marie-Francine Moens, Erik Boiy, Raquel Mochales Palau, and Chris Reed. Automatic detection of argu- ments in legal texts. In Proceedings of the 11th inter- national conference on Artificial intelligence and law, pages 225–230. ACM, 2007. [11] Nona Naderi and Graeme Hirst. Argumentation min- ing in parliamentary discourse. CMNA15, 2015. [12] Anita Pomerantz. Agreeing and disagreeing with as- sessments: Some features of preferred/dispreferred turn shaped. 1984. [13] Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Milt- sakaki, Livio Robaldo, Aravind K Joshi, and Bonnie L Webber. The penn discourse treebank 2.0. In LREC. Citeseer, 2008. [14] Jodi Schneider, Tudor Groza, and Alexandre Passant. A review of argumentation for the social semantic web. Semantic Web, 4(2):159–218, 2013. [15] Simon Wells. Argument mining: Was ist das. CMNA14, 2014. 48