We would like to thank Injenia S.r.l. for supporting this research.

1. Introduction

1One of the most exciting innovations that we are experiencing in the last decade is the massive widespread of conversational interfaces, such as chatbots or virtual assistants (Tsvetkova et al. 2017; Chaves et al. 2019; Dale 2016). The various attempts that have been made to classify these technologies (Radziwill and Benton 2017; Følstad, Skjuve, and Brandtzaeg 2019; Hussain, Sianaki, and Ababneh 2019; Mathur and Singh 2018) and the absence of an unequivocal taxonomy (Braun and Matthes 2019) surely contribute to the lack of a methodological approach for designing conversational agents. They are perceived as something in between humans and web search engines characterised by a conversational way of expression and the capability of managing input and output in natural language (Dale 2016; Braun and Matthes 2019).

2We are witnessing a flourishing literature about technologies, techniques and applications for building conversational interfaces (Ahmad et al. 2018; Adamopoulou and Moussiades 2020). Unfortunately, we cannot say the same for the elaboration of methodological guidelines that can be pursued for the designing of conversational interfaces, especially from a linguistic point of view.

3This is the context in which our research comes to light. We think that a solid anchor in linguistics and therefore a scientific knowledge of what human conversation is may be the key for identifying a generalizable methodological approach for designing conversational agents.

1.1 A methodological lack in conversation design

4A relevant set of studies have been produced on this topic. The first agents were implemented with simple pattern-matching techniques and template-based responses (Weizenbaum 1966; Colby, Weber, and Hilf 1971). They could support a continuative concatenation of utterances but they were still far away from today’s state of the art. The same pattern recognition model shaped the architecture of A.L.I.C.E., a chatbot annotated with AIML, a mark-up language derived from the metalanguage XML (Wallace 2003; Shawar and Atwell 2007). After 2016 there has been a growing interest for chatbots in various areas and applications, mostly because they were considered new productive and entertaining objects not reducible as mere assistants, but capable of a way of interacting that brings them closer to users (Dale 2016).

5Today, most of them rely on machine learning algorithms and Natural Language Understanding modules, but still even the more conversational of the agents can just vaguely simulate conversational exchanges enriched by mutual understanding as we know it as humans.

6Adamopolou and Moussiades (2020) distinguish between two ways of developing chatbots: using any programming language like Java, Clojure, Python, C++, PH, Ruby and Lisp or using state-of-the-art platforms. At this time, the leading NLU cloud platforms supported by machine learning are: Google’s Dialogflow in both versions ES and CX, Facebook’s wit.ai, Microsoft LUIS, IBM Watson Conversation and Amazon Lex. These platforms share a common information-retrieval approach based on what Moore and Arar (2019) identify as Intent-Entity-Context-Response (IECR) paradigm. An intent “[...] represents a mapping between what the user says and what action should be taken by the chatbot” (Adamopoulou and Moussiades 2020, 377). Intents recognize the conversational action a user is performing, while entities are tools used to extract particular details and parameters values from natural language inputs. They can be either system-defined or customized by the developer. Contexts are “[...] strings that store the context of the object the user is referring to” (Adamopoulou and Moussiades 2020, 378) used to capture the context of the current topic. Responses consist of what the chatbot actually answers in chat. This approach aims to extract contextual and detailed information from users’ inputs and respond accordingly to the users’ intention, extracting domain-specific entities and associating the corresponding intent, which means that machine learning algorithms are used for intent identification and entities extraction tasks, but responses are typically pre-authored by a designer (Moore and Arar 2019).

7Generative models capable of automatically generating answers considering current and previous user messages are also in production, but there are still difficulties in building and training them and they are not available in the major commercial platforms (Adamopoulou and Moussiades 2020). Studies on the evaluation of these platforms show similar performances in terms of combined f-score (Liu et al. 2019; Braun et al. 2017), with slight differences in intent identification task (Canonico and De Russis 2018), especially with longer utterances (Zubania et al. 2020).

8Although the technological progresses and the wide technical landscape here outlined, we are facing today a lack in designing domain specific conversational interfaces. The current state of conversational interfaces is limited in terms of established user interface design patterns: it is still unclear when chatbots should be text-based or button-based, or which are the best practices in designing a chatbot conversations. The question about how to structure the interaction with this new medium for creating efficient conversational experiences is still opened. Schiavo and Fadhil (2020) investigate the available scientific literature about interaction patterns and design principles in healthcare and identify four common theoretical themes in which the specific features are categorized: bot-user interaction, bot-response, bot-development, user experience. Since our work has a predominant linguistic focus, we mainly concentrate our attention on linguistic features, such as tone of voice, flexibility of responses, conversation length and user engagement in general. Schiavo and Fadhil (2020) treat each feature separately, offering relevant suggestions, but no univocal applicable design methodology in that sense.

9Some studies embrace sociolinguistic theories: Chaves et al. (2019) apply to the design of a specific use case the register theory, Bennett (2018) and Dippold et al. (2020) identify Interactional Sociolinguistics as key to express chatbot’s personality through language in responses and prompts design.

10There are multiple works concerning the users’ perceptions while texting with chatbots and what they would expect from a satisfying conversation with them: Hill et al. (2015) demonstrate that users hold long conversations also with conversational agents, adapting to chatbots language without overlooking that they were actually chatting with robots. Svenningsson and Faraon (2019) identify the factors of perceived humanness in chatbots’ responses and underline their possible applications in terms of design guidelines. Jain et al. (2018) focus on new chatbots’ users identifying guidelines more related to flow buildings. Kvale et al. (2019) draw practical and theoretical implications from a manual analysis of chatbots conversations, such as the value of cross-disciplinary teams and the need of diligence in chatbots training. Although these studies end with practical advice, they are far too generic for laying the foundations for a methodology.

11There is consistent number of systematic guidelines on how to design conversational interfaces with a practical-computational procedure on how to approach chatbots’ projects, reported also in McTear (2020). Some of them do have a commercial vocation (Hall 2018), other focus on technical issues (Shevat 2017; Dasgupta 2018) but even the more linguistically or cognitive oriented ones do not display a clear affiliation to a complete framework of analysis rooted in linguistics (Pearl 2016; Cohen, Giangola, and Balogh 2004).

12We think that it is fundamental for creating effective conversational agents that should actually converse with humans and whose aim is to simulate the mechanism of human interaction to refer to a solid linguistic framework. Since pragmatics is the area of linguistics that primarily focuses on language in use also in interactional contexts, we agree with Bianchini et al. (2017) on the importance of pragmatics in developing new chatbots examples. Furthermore, we also agree with Bennett (2018) with the identification of Conversation Analysis as a methodological key to design better conversational flows.

2. Theoretical analysis of dialogue and conversation

2.1 A pragmatic perspective

13The term pragmatics is conventionally credited to Charles W. Morris who first introduced a “pragmatic dimension” in the context of relations between signs, interpreters and objects (Bazzanella 2008). Influenced by Charles P. Peirce (1932) and in agreement with Carnap (), he distinguished three “dimensions of semiosis” (Morris 1938, 21), in which pragmatic1938s addressed the relations between signs and who use and interpret them (Horn and Ward 2006). Since pragmatics is an interdisciplinary “hardly a well-integrated field of research” (Dijk 2009, 13), it is preferable to speak about a pragmatic perspective towards language instead of a pragmatic theory (Bazzanella 2008; Bublitz and Norrick 2011). The adoption of a pragmatic perspective also allows to embrace the contributions from philosophy, psychology, sociology, linguistics and the multiple definitions it has (Levinson 1983; Leech 1983; Katz and Fodor 1963; Ariel 2010; Turner 1999).

14The philosophers of language Austin (1962), Grice (1962) and Searle (1969) influenced a notion of pragmatics in contrast with the chomskyan analysis of language as an abstract instrument independent by the context of use. Reflections in this direction are a consistent part of the most common handbooks of pragmatics, such as Levinson (1983), Leech (1983), Mey (1993), Yule (1996), and Verschueren (1999). It thus seems reasonable "[...] to claim that the ensuing pragmatic turn was most notably induced by J.L. Austin, J.R. Searle and H.P. Grice, who were interested in utterance meaning rather than sentence or word meaning, i.e. in studying unique historical events created by actual speakers to perform linguistic acts in actual situational contexts in order to accomplish specific goals” (Bublitz and Norrick 2011, 2).

15The approach adopted in this work is functionalist, that is, “[...] that it attempts to explain facets of linguistic structure by reference to non-linguistic pressures and causes” (Levinson 1983, 7). This perspective aims to explain linguistic phenomena relying on pragmatics principles (Givon 1979; Hymes 1962) and opens to different developable possibilities such as an ethnomethodological method rooted in sociology (Garfinkel 1996; Goffman 1983; Sacks, Schegloff, and Jefferson 1992) and a psycholinguistic approach such as the alignment model (Pickering and Garrod 2004; Branigan, Pickering, and Cleland 2000; Szmrecsanyi 2005).

16In both cases, the study of pragmatics is connected to the use of language in communication. Since communication inevitably involves at least two parties, the primary focus of pragmatics are “language use and language users in interaction” (Bublitz and Norrick 2011, 4). Towards this intersubjective dimension, Kerbrat-Orecchioni speaks of pragmatique interactionniste, whose main objects of research are manifestations of verbal interactions, such as dialogue and conversation.

17Dialogue could be taken as “[...] the elementary and universal form of human communication” (Luckmann 1990, 58), whose basic principles are most salient in conversations and authentic discourses (Linell 2001). Levinson (1983, 284) defines conversation as “[...] the predominant kind of talk in which two or more participants freely alternate in speaking, which generally occurs outside specific institutional settings”. In the broadest sense, it includes both face-to-face social communications and technology-mediated forms of interactions: all these different manifestations can be classified according to different criteria. A common distinction bases on the final scope of the exchange: social interactions’ aim is building and maintaing rapports, while transactional interactions mainly fulfill practical goals (Brown and Yule 1983; Clark et al. 2019). Hakulinen’s classification (2009) takes into account the degree of institutionality, the activity type or genre, the channel and participation framework; Linell and Luckmann (1991) consider the degree of asymmetry between the interlocutors. According to Schegloff, the ordinary conversation is the most general and flexible type of conversation from which the other types are adapted for particular purposes. It is defined as “[...] the basic medium of ‘interactional exchange’ [...] in whatever practices it is embodied in those settings” (Schegloff 1999, 413). Moore and Arar (2019) identify service, teaching and counseling conversations as derived typologies from the ordinary conversations. They are all charachterized by roles’s fixedness and influenced by their settings. Since Moore and Arar (2019) embrace a strictly conversationalist point of view, their focus is on identifying the underlying structure of conversation, which is suitable for slight adaptations according to the settings and contexts real conversations may occur in.

2.2 Pragmatic frameworks of analysis

18There are several useful frameworks to analyze dialogical and conversational interactions. In the field of pragmatics, Haugh (2012) discerns two key trends to place conversational interactions: at the level of meaning and abstract principles referring to the works of language philosophers such as Grice (1989) and Searle (1969); at the level of the performance the analysis of authentic data referring to Conversation Analysis (Sacks, Schegloff, and Jefferson 1974) and Interactional Sociolinguistics (Gumperz 1982). In this section, we summarise the key points of each and the possible adaptation to a human-computer interaction.

19The first trend is situated on a cognitive level and understands conversation as a “joint activity” whose progression is determined by the concatenation of “joint actions” (Clark 1996, 30). They are the result of the coordination of individual actions on two levels: "There is coordination of both content, what the participants intend to do, and processes, the physical and mental systems they recruit in carrying out those intentions.” (Clark 1996, 59). Regarding language use, “[...] a central problem is coordinating what speakers mean” (1996, 73). The idea of conversation as action determined by an undercurrent of communicative intention between the participants was formulated by Austin (1962) and Searle (1996, 73). These studies have been enormously influential in the pragmatic approach of conversation: they allow to “[...] formalise rules and principles by which speakers mean (and to a lesser extent do) things” in conversation, abstracting from the conversation itself (Haugh 2012, 251). For example, the austinian notion of performativity frames new perspectives in human-computer-conversations, such as the collaborative action of “We” Human-and-Technology (Cho and Yoon 2013; Cho 2015) and the methodology of the Performative Experience Design (Spence 2016).

20Intentionality and delivery of implicit meanings are Grice’s main objects of inquiry (1975). He can be considered one of the pioneers of inferential pragmatics (Ariel 2012). Most of all the cooperation principle and the conversational maxims proposed by Grice (1975) and later updated by Sperber and Wilson (1995) are a consistent part of the study of pragmatics today and involved in the implementation of dialogue systems from a methodological point of view: Jacquet et al. (2018; 2019; 2019b) evaluate the violation of the gricean maxims in textual online conversations; Saygin and Cicekli propose an empirical study study of human-computer interactions within the context of the Loebner Prize Contest.

21Lakoff’s theory of politeness (1973) is an attempt of expansion of Grice’s conversational maxims. This theory has been extensively criticized because it is hardly generalizable (Al-Duleimi, Rashid, and Abdullah 2016) and the key terms used in it are culturally determined and they therefore need to be clearly defined (Brown 1976; Tannen 1984).

22Brown and Levinson (1987) propose instead an expansion of the studies on politeness made by Goffman (1967) introducing the concepts of positive face and negative face, which are respectively the need to be approved by the others and the need of autonomy. The importance of politeness in the realization of conversational interfaces is attested among the others by Følstad et al. (2018) and Nordheim, Følstad, and Bjørkli (2018), who list politeness as a factor perceived to affect trust in chatbots for customer service. As far as practical chatbots realizations are concerned, Hall (2018) includes it as an element to take into account during the conversation design, while De Jong, Theune, and Hofs (2008) adapt the model of politeness strategy elaborated by Brown and Levinson (1987) and of the linguistic alignment (Pickering and Garrod 2004) to a virtual museum guide.

23In relation to the second trend, the dominant perspectives are Interactional Sociolinguistics (IS) and Conversation Analysis (CA).

24The core idea of Interactional Sociolinguistics is that what happens in a sequence of talk can be analyzed in its social contexts and that humans in talk accomplish social goals. A central concept of this approach are contextualization cues, “[...] by which speakers signal and listeners interpret what the activity is, how the semantic content is to be understood and how each sentence relates to what precedes or follows” (Gumperz 1982). Feine et al. (2020) offer an overview of the implementation of social cues in different kinds of conversational agents, while Bennet (2018) translates conversational cues in the realm of text-based chats, arguing that a strategic manipulation of orthography to convey conversation cues could help the design of chatbots personality and could situate them on different levels of enthusiasm or considerateness. Dippold et al. attest how a microlevel of design linguistic analysis based on Interactional Sociolinguistics can be useful for chatbot designers for creating engaging interactions and provide specific guidelines. Relevant studies in the field of Computer-Mediated Communication (CMC) describe the modification on the different levels of language that apply on digital mediated communication: Crystal and Herring are pioneers for the English language, while for Italian the work from Pistolesi is certainly a relevant reference point.

25Another approach to dialogical interaction is Conversation Analysis. Levinson includes in his textbooks on pragmatics a chapter entitled Conversational structure, in which he compares Conversation Analysis and Discourse Analysis as two opposite methodological frameworks with a preference for the first one, characterized by an empirical and inductive vocation.

26CA is a subfield of sociology whose origin is influenced by ethnomethodological studies (Garfinkel 1996), but the publications of the firsts and more influential CA contributions in 1974 A simplest systematics in the flagship Journal of the Linguistic Society and of The preference of self-correction in 1977 on Language both by Schegloff, Sacks and Jefferson established a relationship between CA and Linguistics from its origins (Fox et al. 2018).

27The aim of this discipline is providing a systematic description of oral language practices (Schegloff 1992, 120) between humans, in order to formalise it into key structural elements that occur in the variety of contexts in which conversations may take place. Unlike ethnomethodology, whose observations are based on memory and intuition (Pallotti 2007), the methodological approach of CA is fundamentally empirical and its focus was a description of language as a tool used by social actors in interactions. In this respect, the object of study is an interactive activity contextually situated, where sentences are “[...] produced by someone, for someone else, at a certain time, in a certain way” (Hoey and Kendrick 2017). In the words of Schegloff and Sacks (1992, 70), CA is a “[...] naturalistic observational discipline that could deal with the detail of social action(s) rigorously, empirically and formally”.

28The systematic descriptions of such practices leads to discovering the machinery (Sacks 1984, 84) underlying conversations, a mechanics of how people naturally talk in a variety of settings (Sacks, Schegloff, and Jefferson 1974; Schegloff 2007) made of some key concepts that regulate interactions. This approach supports the existence of some structural patterns that occur in conversational exchanges without consideration of the delivered content and with slight adaptations according to the settings and contexts of realization: it is therefore clear the enormous potential it may have in Human-Computer interaction studies.

29We will describe four elements of the descriptive apparatus for analyzing interactional structures, adhering to the analysis made by Pallotti (2007), Moore and Arar (2019) and Hoey and Kendrick (2017), which are today “[...] common stock for everyone doing CA” (Pallotti 2007, 7).

2.3 Key concepts of the conversation machinery

30The strategy people use to manage the conversational traffic in interaction and the distribution of talk among the parties is turn-taking, known to be the feature that makes conversations orderly without significant clashes, overlappings or long pauses. Sacks et al. (1974, 702) describe it as a simplest systematics composed of two components and a coordination of the ending of the turn with the start of the next. The turn-constructional units (TCU) consist of linguistic unit-types such as sentential, clausal, phrasal or lexical constructions that form a “[...] recognizably complete utterance in a given context” (Hoey and Kendrick 2017). These bound units are defined in functional terms, being understood that usually “[...] a complete (linguistic) action corresponds to a complete syntactic unit, so that the TCU boundary turns out to coincide with the clause boundary” (Pallotti 2007, 8). Once a turn is perceived as completed, occur a turn-transfer using turn-allocation techniques such as self-selection and other selection in specific transition-relevance place (TRP). A hierarchically organized set of rules governs the turn construction and coordinates the transition so as to minimize the gaps and overlappings (Sacks, Schegloff, and Jefferson 1974).

31Turns do not occur haphazardly, but are sequentially organized into coherent courses of actions (Schegloff 2007). The minimal unit of sequential organization is a two-move sequence, the adjacency pairs, in which the connection between the parts depends on the conditional relevance: the occurrence of a first pair sets up the relevance of the second part to follow (Schegloff 1968). Schegloff and Sacks (1973) give some examples of adjacency pairs, such as question-answer, greeting-greeting, offer-acceptance/refusal. Moore and Arar (2019, 65) also include farewell-farewell, assessment-assessment, inquiry-answer, request-grant/deny, invitation-acceptance/decline, accusing-admitting/denying. Since the first part sets up an expectation, the absence of the accepted second part is noticeable and the participants may require explanations or justifications for not having answered or for not having chosen the preferred option (Schegloff, Jefferson, and Sacks 1977). Even if the first part is usually directly followed by the second part, it could happen that for satisfying various requirements the completion of the first part has to be suspended for one or more turns. Sequences are therefore inherently expandable through additional turns over and above the two basic units of sequences. Expansions are allowed before the first pair part (pre-expansion), between the first and the second pair part (inserted expansion) and after the second pair part (post-expansion) (Schegloff 2007, I:26). Another kind of sequence are the storytelling sequences (Jefferson 1978), used to express stories, anecdotes, or instructions whose content have to be distributed on multiple turns. They are often introduced by pre-announcement, namely the story preface (Schegloff 2007, I:41). Sequences are an instrument for organizing utterances produced by the participants, while sequences themselves are organized into activities that define the “overall structural conversation” (Schegloff and Sacks 1973, 71), such as conversation openings and closings, instruction giving or troubleshooting.

32Another relevant element in CA is turn design, which deals with how speakers build their turns to achieve some goals and to deliver contents for a specific audience (Drew 2020). Turn design principles concern the speakers’ orientation to contiguity and their tendency to display connections between what they are saying and what the other said in prior turns; the specific lexico-syntactic adopted by the speakers to pursue the desired action and the recipient design principle, which is “[...] a multitude of respects in which the talk by a party in a conversation is constructed or designed in ways which display an orientation and sensitivity to the [...] co-participants [...] with regard to word selection, topic selection, admissibility and ordering of sequences, options and obligations for starting and terminating conversations [...]” (Sacks, Schegloff, and Jefferson 1974, 727). It implies the consideration of the relationship between the interlocutors, their mutual knowledge and common ground (Clark 1996). The other principle to take into account in turn design is minimization, known as the speakers’ tendency to deliver a message or to complete an action without using more words than necessary, but still being recognisable from the recipient the conversation is tailored to (Sacks and Schegloff 2007). In other words, the recipient’s design is prioritarian over minimization: the speaker has to be efficient, using as few words as possible without preventing the interlocutor to understand.

33The last element of conversation machinery are the repair practices spontaneously accomplished by the speakers in case of troubles in speaking, hearing or understanding (Schegloff, Jefferson, and Sacks 1977). Hoey and Kendrick (2017) describe the three basic components of a repair procedure: a trouble source, a repair initiation (i. e. a signal that begins the repair procedure) and the repair solution (i. e. the actual repair, for example paraphrasing or repeating a word part of the prior turns). Both the speaker (self) and the recipient (other) can initiate a repair procedure and/or accomplish a repair solution. Repair mechanisms are thus distinguished relying on who initiates the repair (self-initiated or other-initiated) and on who effectively accomplishes the repair (self-repaired or other-repaired).

34Some attempts have been made adopting CA as a theoretical framework for implementing conversational agents: Luff et al. (1990) early imagined the potentialities of CA both as analytical tool in HCI applications and as inspiration for design methodology. Wooffit et al. (1997) adopt CA as a sociological perspective for studying human-computer dialogues. More recently, Lotze (2016) includes CA in the theoretical approaches used to analyze human-chatbot corpora. Hirst (2001) reviews Luff, Gilbert and Frohlich (1990) focusing on the different conceptual perspectives from which CA and Discourse Analysis are shaped, defending the necessity of considering CA studies in the field of Natural Language Understanding (NLU) technologies and applications. At that time technologies were not mature enough to adhere to such a complex theoretical framework, but at present time more and more studies go in this direction, such as the works from Gervits et al. (2010) and Michael and Möller (2020).

2.4 The Natural Conversation Framework (NCF) as a promising starting point for a design methodology

35The work by Moore and Arar (2019) on NCF represents an innovative proposal in the scientific-industrial landscape. Moore and Arar worked together at IBM-Research in designing prototypes for novel forms of interactions for conversational interfaces and Moore is currently developing a conversational methodology founded on the qualitative models from the field of CA. This is not the first work in this direction: Moore (2013) and Moore et al. (2018) collect interesting contributions on specific design issues generalizable to various use-cases, such as Bickmore et al. (2018) and Candello and Pinhanez (2018) and lay the groundwork for the practical guide published in 2019. Moore and Arar (2019) especially introduce the lack of a methodology for designing conversations and invite us to embrace the complexities of human dialogues in order to create machines we can interact with in a natural way.

37The NCF traces the basics mechanics of conversational patterns documented in CA and consists of four parts: an underlying interaction model based on expandable sequences, a distinctive content format based on the interaction model, a reusable pattern language for common conversational activities, a general method for navigating conversational applications.

38The interaction model is based on the sequential structure of conversations, in which sequences are “[...] general patterns that [...] can be used and reused in all kinds of different situations and settings, for all kinds of different purposes” (Moore and Arar 2019, 65). Like adjacency pair sequences and the storytelling sequences in CA, this model should support sequence expansions, as “[...] natural indicators of the participants’ state of understanding of a turn-by-turn basis” (Moore and Arar 2019). This is a more natural and interactive pattern than the simple two-turn sequence model of the majority of chatbots and virtual assistants currently available.

39In order to apply the interaction model accurately, a particular format is required for the content of the conversational application. (Moore and Arar 2019, 70) express it translating the principle of minimization (Sacks and Schegloff 2007) into three guiding principles: “limit agent utterances to a single sentence or less”, “break paragraphs down into their parts” and “let users control the level of detail”. These criteria enable designers to break-up document-formatted content into bite-sized intents, which can be requested by the users through simple queries.

41The patterns can be divided into three categories: conversational activities, sequence level management and conversation-level management (see Figure 1). Conversational activities include patterns for managing content inside the boundaries of the conversation, such as inquiries, requests or extended tellings. The remaining two categories help users and agents to manage the interaction itself and occur on two levels: management patterns on a sequence level help managing successful sequential interactions through repair procedures, management patterns on a conversation level help managing conversations’ openings, closings and disengaging. Finally, the navigation method consists of six simple conversational actions the users can take to get around the conversational space.

42The so defined set of patterns is much more similar to a “language”than a “library”: these patterns are designed to be used together and are particularly useful in service encounters conversational interactions, which are fundamentally practical dialogues characterized by a goal-seeking nature (Allen et al. 2001).

43To show how the patterns are presented, we provide an example of sequence closer patterns, from the category of the sequence management patterns (Moore and Arar 2019, 142). Sequence closers enable users to communicate that the previous action is completed and that they are ready to move on:

44U: INQUIRY/REQUEST A: RESPONSE U: ACKNOWLEDGMENT/ASSESSMENT/LAUGHTER <OPTIONAL PAUSE> A: LAST TOPIC CHECK

45U: what’s your favourite movie? A: I guess I like movies with a strong AI lead. U: ok (2.0) A: Anything else?

46Each pattern is expandable with other subpatterns or may present different options according to its conversational structure. Patterns display the same format: an abstract model of the pattern with generic social actions or intents in capital letters and an adaptation of the pattern to a particular use-case. Here the user does the first part of an adjacency pair (an inquiry or a request) successfully granted or answered by the agent. If the agent did not succeed, the third position would not have been filled with an acknowledgement but it would have been expanded with a repair initiator, like “what did you say?” or “define AI”. In this case, the agent does not need to say anything more after the acknowledgement and, after 2-3 seconds pause in a vocal interaction or 20-30 seconds pause in a textual interface, the agent checks if the user has other requests or topics.

47The adherence to an Intent-Entity-Context-Response paradigm allow to apply the language pattern of the NCF to various platforms and use-cases. Specifically, the examples shown in the textbook are all realized on Watson, but we think that the same patterns could be realized also on other platforms with the same paradigm, like Google’s Dialogflow. Furthermore Moore and Arar methodology is primarily focused on the realization of vocal conversational interaction in English, but we argue that the same patterns with slight adaptation could be implemented also in textual conversational interfaces and in other languages, like in Italian.

3. A roadmap towards the implementation

3.1 A possible proposal: a summary roadmap

48In this section, we provide a practical-procedural workflow for approaching chatbot projects based on our work experience in the field and on the most influential textbook guidelines on this topic (Cohen, Giangola, and Balogh 2004; Pearl 2016; Shevat 2017; Hall 2018; Dasgupta 2018; McTear 2020). This workflow is made of macro and micro-levels integrated together and should involve stakeholders from different departments, considering at least the joint effort of marketing, linguistic design and the technical development orchestrated by a flexible methodology that opens with the project’s requirements definition up to the implementation.

49The stage of project’s requirements definition produces as output the acknowledgement of the feasibility of the conversational agent’s project, the identification of the potential target users and of the technical requirements the application should possess. We substantially agree with the procedure proposed in Cohen et al. (2004, 46) that covers three micro-levels: understanding the business, understanding the user, technical requirements. There are multiple business issues that need to be understood and questions to be answered. Answering these questions can provide significant guidance to also write down metrics for the overall evaluation of the project.

50The second element of the requirement phase is the understanding of the population of expected target users, which needs to be understood both in terms of characteristics/needs and of usage modality. Cohen et al. (2004, 49) list various elements for understanding and taking into account the final user’s needs and characteristics.

51The second perspective regards how, when, where and why the final users will use the application (Cohen, Giangola, and Balogh 2004, 50). They suggest also two practical ways to get these information: a preliminary overview of the various touchpoints of the company focusing on offered functionalities and transmitted feels, and organizing meeting with the company. The key attitude to develop to gain this information coincides with the first stage of the Design Thinking process assumed in Moore and Arar (2019), empathize. Empathy towards the interlocutors to get an understanding of the business and the users can be obtained through observations of how the final users currently interact with the industry or resolve the task that will be supplied by the chatbots, or through the engagement of the people directly involved. From a linguistic point of view, examples of useful materials to collect and analyze may be emails, call transcripts, or messagges from the final users to the company.

52The third requirement is getting an understanding of the application, focusing on the technical requirements of it. This stage is a prerogative of the technical department and its main point is getting an understanding of the application from a technical point of view, evaluating the feasibility itself and the strategies that need to be elaborated for solving tasks and subtasks.

53The second macro-level point is definition, whose goal is to draw conclusions from the preliminary analysis of the project’s requirements in order to define user needs more formally. We identify three elements that have to be defined: goal statements, user personas and agent personas. With goal statements we mean the definition of key design criteria learned from the analysis of requirements. They involve the definition of user goals grouped into broader statements, defined also with respect to the technical, financial and organizational constraints of the specific project.

54The user personas is a fictional representation of the target user: a systematization of the collected user’s qualities to define a prototype of who will typically interact with the final product. The user personas should also reflect the users’ pain points, in order to improve the user journey. From our personal experience in the field, an efficient way to systematize the users’ pain points is to write them down synthetically and associate them with benefits that the chatbot could provide in relation to them. An example is shown in Table 1.

Table 1: Pain points in the user journey associated with possible chatbot’s benefits.

Pain Points	Chatbot benefits
Long waiting time on the phone	Instant answer or escalation to an operator
Difficult information retrieval	Personalised user journey to the retrieved information starting with an initial disam-biguation
Hard understanding of complex and long documents	Systematisation of information in small slots and simplified language

55The last element of the second section is the definition of some characteristics of the agent personas (Hall 2018; Cohen, Giangola, and Balogh 2004; Pearl 2016). In our experience on the field, we do agree with the synthetical sketch proposed in Moore and Arar (2019) that splits the agent personas design into three components: agent job description, agent personality and agent self-knowledge. A starting point for describing the qualities and the language of an agent, is imagining a job it is supposed to do. What role is the agent supposed to substitute or replace? Trying to list down duties and activities the chatbot is intended to assolve and the expected experiences and qualifications, as if it was a real job candidature, can help to be consistent also in the design itself and throughout the project. This procedure helps to identify jargon and recurrent technical terms the chatbot is supposed to manage. The agent personality needs to be characterized in terms of communicative style, its level of formality and generally the tone of voice the target users expect to find in the conversational agent they are talking to. Two other core issues of the agent personality are the assignment of a gender to a chatbot, with the cultural and social implications that this may bring with, and the opportunity of humor in it. From our experience on the field, strongly anthropomorphized chatbots usually do belong to a gender, which most of the time is female (West, Kraut, and Chew 2019). There are also cases of neutral chatbots that reproduce an animal, a vegetable or a fantasy character. It depends on what kind of character or conception of gender we aim to reproduce in a virtual reality, being aware of the risk of reproducing virtually biases or prejudices belonging to the real word (Strengers and Kennedy 2020).

56Humor can be an efficient strategy to build trust, especially if used in secondary responses that do not cover the main topics of the chatbot. Since the users expect a chatbot to be productive and efficient (Brandtzæg and Følstad 2017, 2018; Piccolo, Mensio, and Alani 2018; Zamora 2017), humoristic responses rather than informative ones can be counterproductive and may indeed frustrate the user. On the other hand, receiving a humoristic answer in an unexpected context such as online conversation with a machine, can increase the surprise effect and it may induce the users to continue the conversation (Jain et al. 2018).

58The third step in the proposed methodology is called mock-up & ground truth. Once the target users and the goal statements have been sketched out, we can move on to the drafting of the conversational perimeter. This term identifies a sort of table that holds together the groups of topics managed by the agent, structured in a way that reflects the Intent-Entity-Context-Response paradigm of the most common NLU platforms. In this phase, it is important to define the intents coherently and functionally to the goals of the agent. If conversational data collected during the preliminary phase such as call phone transcripts or emails are available, we suggest to group them following a bottom-up labelling approach. This approach consists of first grouping collected utterances into wide general categories, like “questions”, “problems”, “getting information” and then proceeds refactoring and splitting the so funded categories into more specific ones, considering for example the topic of the question, the action required to satisfy it and so on. Through this procedure we can both identify intents and train them with authentic linguistic material that constitute the so-called training phrases. A schema of how it works is shown in Figure 3 representing a schematisation of the reasonings behind the identification of the topics that will be handled by the agent, but it is not proper conversational perimeter as we mean it.

59The conversational perimeter is the final elaboration of this reasoning. It also includes a section dedicated to the responses, that should coincide with the actual responses that the agent gives in chat and has to be updated with new information or modifications. Furthermore, it also contains some examples of the training phrases, utterances that final users can potentially say to formulate a request. Training phrases are categorized in intents, in turn grouped according to the topic they refer to. A conversational perimeter offers a systematization of the topics handled by the conversational agent, organized in intents like the platform it will be developed on; and it shows how simple question-answer exchanges will be managed directly in chat, with some examples of utterances the final users may produce and the answers they will be given. Table 2 shows the different section of a conversational perimeter, re-adapting the utterances from Figure 3 to the final elaboration of the perimeter.

Table 2: A section of a conversational perimeter

Topic	Training phrases	Intent	Response
Consegna	Quando consegnate ?	Consegna_orari	Consegniamo tutti i giorni della settimana, compresi i festivi.
	Posso sapere gli orari di consegna?
	Quando fate la consegna?
Consegna	Consegnate all’estero ?	Consegna_estero	Consegniamo solo in Italia, non all’estero.
	Consegnate solo in Italia?
Consegna	Quanto costa la consegna ?	Consegna_costi	Il costo della consegna è gra-tuito per gli utenti registrati e di 4 euro per gli ospiti.
Consegna	La consegna non mi è ar-rivata.	Consegna_ritardo	Se ritieni che ci sia un ritardo di consegna, posso inoltrare una segnalazione al servizio clienti, mi basta solo qualche tuo dato.
	Il pacco è in ritardo.
	Devo segnalare una consegna in ritardo.
	La mia consegna non si è verificata.
Pagamenti	Posso pagare con carta di credito ?	Pagamenti_metodi	Puoi pagare con paypal, satispay e carta di credito registrata.
	Posso pagare con satispay?
	Posso pagare con contanti?
Pagamenti	Mi devo registrare al sito per pagare?	Pagamenti_account	Puoi comprare dal nostro sito sia come ospite sia come utente registrato, a te la scelta!

60We need to consider that a complete conversational perimeter is usually much bigger than this: it depends on the specific project, but in our experience it can contain between fifty and one hundred intents. Furthermore, it contains only the simplest form of interaction: the responses are static, which means that they do not change dynamically taking information from external sources, but they are always the same, even though on the most common NLU platforms they can be randomized, that is to a specific intent can be associated with one or more responses randomly picked-up by the agent to create conversational variety. The form of the responses should respect the basic language-specific pragmatic norms, as shown in section 4.

61The second micro-level step of the mock-up and ground truth mapping is dedicated to the dialog strategy, concerned with the effective building of the dialogue. It answers the question: how will the back and forth between the agent and the final users be? Will generally the agent start the conversation or the user? If the conversation flow is supposed to be more complicated than a simple question-answer and it requires the following of a specific path, we recommend the use of graphical tools for visualising the steps of the path we have imagined and possible variations, such as Google Draw or XMind.

62Mock-up transcripts is thought especially for more complex flows, that need more than one conversational turn to be developed and that may also be represented on one of the graphical tools mentioned above. In this phase there are two alternatives that can be pursued. From one side we can simply write down the dialogues we would like to reproduce with the agent (Cohen et al. (2004) and Pearl (2016) call them sample dialogues) and read it aloud to see if they sound human before implementing it, using either programming languages or commercial platforms that allow to concatenate conversational flows reproducing the characteristics identified by CA mentioned before. Alternatively, instead of creating the flows from scratch and implementing them, our proposal considers the Natural Conversation Framework and especially the sequential patterns proposed by Moore and Arar (2019). The latter approach appears more innovative and more structured, and it is the one we would like to adopt.

63The final macro-level step of the roadmap is the concrete realization of the agent and consequently the effective implementation of the defined flows and selected sequential patterns on a NLU platform, on specific chatbot tools or using the most common programming languages.

64After the implementation, a fundamental step before the roll out is testing. Like McTear (2020) claims, there is still no unified and univocal testing approach, but it depends on the implementation method behind the agent (Deriu et al. 2020) and on the project’s requirements that need to be evaluated. Testing has a double function: to evaluate the efficiency of the developed application and, if possible, to improve the actual functionalities with a tuning activity. From our experience on the field, we suggest various test steps before the final roll-out. This allows developers and designers to evaluate the results and, if necessary, to tune some aspects of the application before the final version. A dialog transversal test (Cohen, Giangola, and Balogh 2004; Pearl 2016) for evaluating the behaviour of the system in every dialogue state and in every condition seems very important, trying out some out-of-perimeter utterances to verify the proper response of the system also in such cases. A fundamental aspect that needs to be considered in this context, is that a conversational agent is a constantly-in-progress creature: even though it is finished and especially if it supports a NLU tool, it needs to be updated with new linguistic materials in order to make the performances better and better. It has to be seen as an alive creature and project, that is nurtured by language and, in some way, produces language: therefore, training and maintenance are continuative activities. There are multiple elements that can be tuned or updated to make the performance of the chatbot better: the training phrases in natural language, the responses in case of changings in the information to deliver, new conversational paths. Table 3 shows a possible evaluation framework taken from our experience on the field and resulted from the combination of two variables: in/out of perimeter, which is referred to the coherence of the utterance produced by the user in relation to the conversational perimeter of the conversational agent; and correct/wrong, which is the effective evaluation of the response given by the conversational agent in that specific context. From every possible combination of these variables we provide possible improvements that can be undertaken.

Table 3: Evaluation grid

	In perimeter	Out of perimeter
Correct	The user’s utterance is in perimeter and the chatbot answers with the expected associated response: no improvements needed.	The user’s utterance is out of perimeter and the chatbot answers properly acti-vating the expected fallback intent: no improvements needed.
Wrong	The user’s utterance is in perimeter and the chatbot answers wrongly with a response associated with another intent. Improvements may be: updating the training phrases of the missed intent, verifying the training phrases of other intents that may cause miss-match, or modify the response with more information.	The user’s utterance is out of perime-ter and the chatbot doesn’t activate the fallback intent,but a response as-sociated with another intent. Improve-ments may be: adding more training phrases to a fallback intent, adding training phrases to the missed intent, verifying training phrases that may have caused miss-match.

4. Adaptation of the methodology and implementation

65After the definition of a theoretical approach rooted in pragmatics and a practical-computational operating procedure towards the implementation, in this section we expand and enrich the procedural workflow with the implementation of a selection of patterns on one of the most authoritative NLU commercial platforms also for the Italian language, namely Google’s Dialogflow (Zubania et al. 2020).

66The selection of the patterns has been made according to two criteria: (a) the adaptation to a text-based modality of interaction; (b) the suitability to a customer-service context (Szymanski and Moore 2018).

67We elaborate a general customer-service case study on which we transpose a selection of patterns in Italian. This practical section on the implementation aims to demonstrate the high generalizability of the approach, still considering the language-specific pragmatic implications in prompt design, and its suitability also to business-oriented contexts of use.

4.1 General case-study description

68The structure of conversation belonging to a customer-service domain is similar to the more general category of service conversations, i.e. dialogic exchanges in which a person (in the role of a customer or a citizen) requests services or information and another person on behalf of an organization or an institution, provides services or information.

69Since the focus of our work is not to build a comprehensive conversational agent but to demonstrate the effectiveness of our methodology for the design of sequential flows, we do not dwell on the details of the conversational perimeter of the case-study.

70It suffices to say that it is a customer-service chatbot, whose goal is to provide repetitive information to support customers on the e-commerce of a chain of shops. It can provide information about typical online-shopping requests, such as deliveries, expeditions, payments and returns and it manages issues and problems related to the state of the orders. Customers may own a fidelity card, a card they collect points on to gain special discounts. Furthermore, customers can register on the website and activate an online profile with all the details on their customer’s situation. Even though the main goal of the conversational agent is not to perform complex task, it should be able to support the final users through the registration process on the website. The chatbot handles it through some questions focused on the extraction of data. We can distinguish between two kinds of data to extract: (a) user’s name and user’s email address are necessary data to complete the registration process. Without the collection of these data, the registration can not be performed; (b) the fidelity card number is an optional data. If the user decides to not provide it or the user does not possess a fidelity card, the registration process it is not compromised.

71The registration process is the interaction that we are going to transpose in Italian using the NCF pattern and implement on Google’s Dialogflow, in order to demonstrate the validity and the high generalizability of the presented theoretical and methodological approach.

4.2 Pattern selection and transposition

72The registration process on the website can be realised basically associating three patterns of the NCF described in Table 4, namely the pattern A2.6 Open Request Summary the pattern A2.7 Warrant Request & Refusal and the pattern A2.11 Open Request Repairs.They all belong the the first group of pattern, conversational activities, and therefore help to manage what happens inside the boundaries of the conversation itself.

73As shown in the section 2.4, each pattern is made of an abstract model of social actions in capital letters and an adaptation of the pattern to an hypothetical dialogue.

74The pattern A2.6 Open Request Summmary manages requests that need a progressive level of detail to be satisfied, like a registration process, in which a set of data need to be collected in order to perform an action or to fill-in a module. The peculiarity of this pattern is the final summary at line (8), that should allow a user’s confirmation or modification of the inserted information. The modification can be performed using the pattern A2.11 Open Request Repairs, which handles repair procedures. The pattern A2.7 Warrant Request & Refusal manages user’s questions concerning a detail request and the possible refusal of giving a specific data. As described in 4.1, we have both compulsory and optional data that need to be handled in different way.

75In the next section, we explain the pragmatic aspect we take into account for the transposition into Italian and we then implement them on Google’s Dialogflow.

Table 4: Selected patterns for the transposition into Italian

Pattern A2.6 Open Request Summary
1 U: PARTIAL REQUEST 2 A: DETAIL REQUEST 3 U: DETAIL 4 A: DETAIL REQUEST 5 U: DETAIL 6 A: DETAIL REQUEST 7 U: DETAIL 8 A: REQUEST SUMMARY 9 U: SUMMARY CONFIRM 10 A: GRANT 11 U: SEQUENCE CLOSER 12 A: RECEIPT	U: I need to register A: Okay. What’s your name? U: Dave Bowman. A: What’s your job title? U: astronaut A: What is your organizational affiliation? U: NASA A: So Dave Bowman, Astronaut at Nasa. Is that correct? U: yes A: Ok. U: thank you A: You’re welcome!
Pattern A2.7 Warrant Request & Refusal
1 A: DETAIL REQUEST 2 U: WARRANT REQUEST 3 A: WARRANT 4 U: REFUSAL 5 A: ACKNOWLEDGMENT	A: What’s your job title? U: why do you need to know? A: Job title enables us to know the range of professions of our partici-pants. U: none of your business. A: No problem. What’s your organiza-tional affiliation?
Pattern A2.11 Open Request Repairs
1 U: FULL REQUEST 2 A: GRANT 3 U: REPAIR INITIATOR 4 A: REPAIR 5 U: SEQUENCE CLOSER	U: Can you recommend a nearby Mex-ican restaurant? A: Mario’s located at Beach and Main. U: What do you mean? A: The Mexican Restaurant, Mario’s, is located at the intersection of Beach Boulevard and Main Street. U: thanks

4.3 Pragmatic aspects for the transposition into Italian

76The pragmatic-functionalist perspective adopted in this work emerges both in the turn-taking sequential structure of the patterns of NCF and in the transposition of the chosen sequential patterns into the Italian language. The observations made by Cohen et al. about prompts design are still actual and rich in ideas, but they are concerned exclusively with the English language.

77In our work these aspects become relevant for the distribution of content in the Italian transposition of the selected patterns. We are going to transpose patterns from a language to another, namely from English to Italian, respecting the predominant tendency of positioning on the left side of the utterance the topic/given information and the comment/new information on the rest of the sentence. This norm can be violated for obtaining particular sense-effects and for enlightening a part of an utterance that in a normal order of the component would instead be positioned at the end of the sentence. In Italian the most used strategies for producing this effect on a written channel of communication are the inversion of the normal order of the syntactic components through focalizations such as dislocations and cleft sentences.

78Another element on which Cohen et al. (2004) dwell on are discourse markers and their use in prompts’s content design for conversational agents in English, organized into various classes of use and considered an efficient strategy for enhancing the natural flow of conversational technology. Without going deeper into the classification of discourse markers in Italian (among the others Bazzanella and Berretta ), it suffice to say that we will take them into consideration for their pragmatic impact on sentences organizations and structures during the patterns’ transposition. In Italian, they are indeed significantly important for their interactional function on both sides of the speakers, their meta-textual function as demarcations, focalizations and riformulations and finally for their cognitive functions (Bazzanella 2008). As Cohen et al. claim, it can be that discourse markers are perceived as informal or slang, even though they are classified as a functional category with no explicit formality degree.

79What instead has to do with the level of formality of an utterance and a piece of discourse in general (Clark 1996), are register and consistency. Even though the definition of register is controversial (Bazzanella 2008), we can say it involves the psychological and social rapports between the speakers, the circumstances in which the communication takes place and the adopted channel (Halliday 1994). This is one of the dimensions of variations of language. Other dimensions of variations are determined by the spatial origins and the geographical distribution of the speakers, by socio-cultural elements such as level of instruction, age and competences and channel of communication. Especially in technical or highly specialized work context, the use of jargon is an important issue to take into account. It is indeed acceptable if all the speakers do share similar background and analogous competencies in relations to the main content expressed by the conversational agent we are working on.

80The channel of communication is also an issue with significant implications on the distribution of content and therefore on relation to the drafting of the responses. Without deepening into the characterization of the language variations adopted in digital contexts of communication, we have to consider that the variant of language adopted with a text-based conversational agent is an intermediate solution between the two opposite poles of written and oral language. This variation takes place in a written form but it shares important elements with speech (Pistolesi 2018), since it is a type of writing that considers more the acoustic effect than the visual one.

81All these sociolinguistic and pragmatic aspects have to be considered as requirements for transposing the selected patterns into Italian. In the next section, we are going to outline a generic customer-service use-case and in followings there is the effective realization and transposition of the patterns on Google Dialogflow.

4.4 Implementation on Google’s Dialogflow

82We realize the selected NCF patterns on Google’s Dialogflow ES.

83The three patterns have been combined together in order to cover multiple scenarios we may face during a registration process. For doing that, we implemented fourteen intents on Dialogflow

01_reg_00_registrazione_generico	01_reg_03_registrazione_riepilogo
01_reg_01_giustificazione_nome	01_reg_03_rifiuto_email
01_reg_01_registrazione_carta	01_reg_03_spiegazione_email
01_reg_01_rifiuto_nome	01_reg_04_modifica_dati
01_reg_02_registrazione_email	01_reg_04_riepilogo_corretto
01_reg_02_spiegazione_tessera	01_reg_04_riepilogo_negativo
01_reg_02_tessera_rifiuto	01_reg_05_ringraziamenti

connected together by input and output contexts manually set and we use three system entities for the parameter extraction: (a) sys.person to extract and memorise the user’s name, (b) sys.number to extract and memorise the fidelity card’s number and (c) sys.email to extract and memorise the user’s email.

84The opening intent 01_reg_00_registrazione_generico is activated by training phrases that express the intention of starting a registration process on the website. As Figure 4 displays, the contents’ distribution of the agent’s response follows the topic/comment order described in 4.3: the first part introduces the registration process already known by the user, since he/she asked for it, and prepares the speaker to provide a set of data. In this case, the user provides the requested data and the user’s name is thus memorised through the annotation of the training phrases with the sys.person entity and the extraction of the corresponding parameter.

85As the pattern A2.7 Warrant Request & Refusal demonstrates, repetitive patterns in a process based on the data extraction, such as the outlined registration, are users’ requests of warranty and explanation concerning a specific data. We provided two intents for managing questions about the reasons of requesting a name in this context and the refusal of providing it, respectively the intent 01_reg_01_giustificazione_nome and the intent 01_reg_01_rifiuto_nome, both conncted to the hierarchical intent 01_reg_00_registrazione_generico by contexts. The intent 01_reg_01_rifiuto_nome is also subordinated to the 01_reg_01_giustificazione_nome, which means that after an explanation’s request by the user the chatbot is able to handle also a refusal of providing the requested data (see Figure 5).

86The chatbot’s responses are more extended then the examples provided by Moore and Arar . This is because we are specifically working on a text-based interface, which allows us to insert also a bit longer prompt than with only vocal conversational assistant. As we claimed in Section 4.1, the user’s name is a required information without which the registration process cannot be performed. The conversational agent offers an alternative to the user, ("Se vuoi, puoi registrarti anche dal vivo, ti basta andare alla cassa [...]") because we are in a business-oriented context, where the customer satisfaction is the most important requirement (Gnewuch, Morana, and Maedche 2017). If after the activation of the intent 01_reg_01_rifiuto_nome the user at the end provides the name, the registration process can proceed again.

87The second data to extract is the fidelity card number, which is not compulsory and it can be skipped without compromising the registration. Also in this case we have a good path (Pearl 2016) where the number is immediately provided, and two expanding intents used to manage clarification’s requests and refusal related to this topic, namely the intent 01_reg_02_spiegazione_tessera and the intent 01_reg_02_tessera_rifiuto. Since the fidelity card number is not a sensibile data, instead of a warranty we provide further explanation about the number (where it is, what it is, why it is requested) and possibility of managing refusal, especially if determined by incomprehension, impossibility to find the card or absence of the card (see Figure 6).

88In any case, if the fidelity’s card has been provided or not, the registration conversational flows proceeds to the third and last data, the user’s email, which is, like the name, a compulsory data to complete the registration, and it is therefore handled by the intents 01_reg_03_rifiuto_email and 01_reg_03_spiegazione_email with training phrases similar to ones dedicated to the name, with the difference that they are annotated with the sys.email entity.

89As displayed in the pattern A2.6 Open Request Summary, after the collection of data we have a summary of all the provided information memorised through contexts and reported in the responses of the intent 01_reg_03_registrazione_riepilogo through coding some specific variables through a specific Dialogflow syntax, as Figure 7 shows. These are the responses that the conversational agent displays once the 01_reg_03_registrazione_riepilogo is activated by the insertion of an email address. They contain parameters associated to the data extracted with entities and stored in the conversational flow through contexts, according to the form: #context-name.parameter-name. The two responses are chosen randomly from Dialogflow: it means that there is no certainty that if the user did not provide the fidelity card’s number in the flow the conversational agent will display the response with “carta fedeltà non fornita”. Our tests reveal a tendency to choose the appropriate response according to the collected data, but for more precise result we recommend to customize the response with some code. In Figure 8 we show how they appear if filled with the corresponding value.

90The summary ends with a agent’s request of confirmation before validating the inserted data. If the user confirms, the data are validated and the sequence can be closed with a sequence closer and its reception. If the users would like to modify one of the inserted data, a repair procedure can be started. The repair procedure can pass through the intermediate intent 01_reg_04_riepilogo_negativo, where the user simply acknowledges that the proposed summary is not the correct, or it can directly activate the intent 01_reg_04_modifica_dati, the core-intent of the repair. Figure 9 shows a repair procedure with the activation of both the intents 01_reg_04_riepilogo_negativo and 01_reg_04_modifica_dati.

91The repair procedure is handled by the intent 01_reg_04_modifica_dati, connected to the previous ones by contexts and whose training phrases are annotated with the system entities used before. The difference here is that the parameters associated to the entities and extracted from the user’s sentences are saved with another value. The new extracted value is used in the response with the same #context-name.parameter-name syntax, in order to overwrite the previous value with the new one. The next figures show the parameter section in Dialogflow and the responses with the extracted parameters.

92The prompts are enriched with discourse markers that functionally structure the progression of the conversational flow ("Va bene", "Ok, quindi", "Ok", "Perfetto","Ottimo, abbiamo finito").

93This demonstration does not expect to exhaust all the possibilities and expansions that a comprehensive registration process may present. The repair procedure handles for instance the modification of a single data and we did not focus on the insertion of unexpected or wrong-spelled data. The goal of our implementation is to demonstrate the effectiveness and the validity of an approach, which is supported by an high-level theoretical method and which is easily transposable into other languages and using various NLU commercial platforms.

5. Conclusions and future directions

94We have therefore obtained a formal, highly generalisable procedure for building conversational agents. Our procedure is explicitly anchored to Conversation Analysis as authoritative methodological approach for realising sequential exchanges and reproducing typical features of human conversational interactions. We additionally transposed the patterns in Italian considering the pragmatic implications of the chosen language and selected a flexible case-study that allows to easily recreate the patterns in multiple contexts and situations.

95The generalisability of the selected patterns is thus subordinated to two main factors: the cultural implications that may occur in the transposition of the patterns into another language and the level of complexity of the conversational flows to implement. The problem of the transposition of the pattern deals with the pragmatics of the language and the cultural expectations related to the customer experience that may influence the dialogue design as well. As Brandtzaeg and Følstad (2017) claim the search of productivity in chatbot use is explicitly anchored to Western culture. This aspect conditions the entire dialogue design, for example in positioning the chatbot scope within the very first conversational turns, in order to not waste time. In our experience, the adaptation of the pattern have been made between two rather culturally similar language, but it might not have been the same with, for example, an oriental culture and its language. Beyond the strictly linguistic aspects, also the project requirements and the agent personas definitions are also culturally defined (Ruane, Birhane, and Ventresque 2019; Cardinal, Gonzales, and Rose 2020). The generalisability of the patterns is also determined by the scope of the specific conversational agent that need to be implemented and by the level of granularity of the conversational flows it should have: the patterns can been seen as a base model that can be simplified or complicated depending on the individual needs.

96As future directions of our research, we are going to further investigating the applicability of this approach on the Dialogflow CX version and possibly on other commercial NLU platforms.

97CRediT author statement; ADA: Conceptualization, Methodology, Validation, Formal Analysis, Investigation, Writing (Original Draft), Writing (Review & Editing); FT: Conceptualization, Supervision, Project Administration, Writing (Review & Editing), Funding Acquisition.