Educating Dora: Teaching a conversational agent to talk

Abstract

Many organisations are witnessing the introduction of automated systems to mediate between them and their clients, often designed as dialogic user interfaces which are conversational in nature. Despite calls for the design of AI-powered conversational agents to draw on Ethnomethodological Conversation Analysis (EMCA), how an EMCA-informed conversation design can improve interactions between humans and conversational agents is yet to be empirically examined. This paper reports on a collaboration between EMCA researchers and AI conversation designers at digital health company Ufonia to develop Dora, a pre-existing ‘automated clinical assistant’ that provides telephone consultations for patients of the UK’s National Health Service. Our analysis identified differences between the conversation design of the product and conversational practices found in equivalent activities involving human clinicians, relevant to implementing changes to improve user experience. We demonstrate that users are prompted to recalibrate their practices (such as for turn-taking) in situ as they engage with a system which does not match their prior experiences of this activity type. Such insights can be leveraged to adapt system design so that it more closely approximates users’ prior experiences of such engagements.

Introduction

While text-based interfaces such as ChatGPT and other forms of chatbots are widely recognised, another form of conversational technology which is increasingly prevalent is the voice user interface (VUI). Such AI-powered conversation-framed technology enables users to engage with automated systems through talk. The most well-known VUIs are ‘command and response’ voice assistants, such as Amazon’s Alexa, Apple’s Siri and Google Assistant, with which users generally engage in short, simple, back-and-forth exchanges (user summons, assistant response, user request, assistant response). There are also, however, more complex VUI systems, typically used for telephone-based customer service transactions by large organisations such as banks. Users and organisations can use such ‘conversational agents’ to perform more complex tasks through longer stretches of spoken interaction.

At the time of writing, the talk produced by VUIs is generated through scripts written by conversation designers. Guidance generally encourages conversation designers to produce such scripts based on their imaginings of how humans might speak in that particular context, and then to test the scripts through, for example, table reads with colleagues.

Since conversation designers report an aim to emulate ‘natural conversation’ (e.g. Choi et al., 2020), and leading tech companies aspire to (for example) ‘craft conversations that are natural and intuitive for users’ (Google, 2023), we have argued that empirical examination of equivalent human-human interactions should be fed into the conversation design process (Brandt et al., 2023; Hazel and Brandt, 2023). Further, we align with others (e.g. Moore, 2018; Stokoe et al., 2021) who have argued that the social scientific approach of Ethnomethodological Conversation Analysis (henceforth EMCA), with its emphasis on the detailed examination of naturally occurring talk-(and-other-conduct)-in-interaction, is a particularly well-placed tool to enable VUIs to more accurately leverage human social interaction.

A handful of recent EMCA studies have examined how members engage with Voice User Interfaces (VUIs). Due and Lüchow (forthcoming) demonstrate how humans have to adapt their practices in the use of Google Home devices. Similarly, Avgustis et al. (2021) compared how callers to a Russian municipal call centre formulated their enquiries differently, resembling conversational input into a web search engine. Although not surprising that using a VUI is different from having a conversation, it has also been demonstrated that in some contexts people using conversational agents like Alexa do draw upon practices found in human-human interaction (Albert and Hamann, 2021; Korbut, 2023).

The research presented in this paper is part of a larger collaboration between EMCA researchers and a team of conversation designers from Ufonia, a digital health start-up. Ufonia has developed Dora, an ‘automated clinical assistant’ which provides telephone clinical consultations for patients of the United Kingdom’s National Health Service (hereafter NHS) across various counties in the UK. The aim of the collaboration is to explore the extent to which EMCA insights and analyses can inform the design of Dora’s conversational interface, and ultimately to generate some guiding principles for effective conversation design of VUIs more broadly. We focus here specifically on the openings of the telephone calls to demonstrate how EMCA insight has informed the further development of the design of Dora.

Conversation analysis and telephone openings

An EMCA approach to social interaction begins with the understanding that talk (and other conduct) is highly systematic; that interlocutors build social actions through their talk, and that these social actions are sequentially organised in ways that are recognisable and reproducible by social members.

Schegloff (1986) pointed to how the perfunctory, routine nature of telephone call openings made this a sequential environment especially interesting for study, including ‘in “artificial intelligence” studies on the production and processing of natural language use’ (p. 113, italics added). Indeed, drawing on his earlier work on call openings (Schegloff, 1968, 1979), he describes these action sequences as ‘routines that the parties “go through” in a virtually automatic or even automated fashion’ (1986, p. 113). This does not imply that all call openings are identical, but rather that there are particular ‘routine ritual[s] of conversational openings’ (Schegloff, 2007) through which speakers transition into the call.

In this sequential environment, interlocutors monitor one another’s contributions and tailor their own to establish the grounds for the call, and how to proceed. Relevant actions that might be included include greeting sequences, ‘howareyou’ sequences, speakers identifying themselves in a range of ways, displaying recognition and articulating the reason for the call (e.g. Schegloff, 1968, 1979, 1986, 2002; Stokoe, 2014; though see Mlynář and Arminen, 2023 for an historical perspective). Depending on the type of call, any of these features may also be omitted. It is through the choice of what to include that interlocutors display their understanding of the purpose of the call, and through which they subsequently can coordinate their contributions.

In institutionally oriented call openings, for example, we may find an absence of ‘howareyou’ sequences, and identification may include the name of the organisation rather than the personal name of the caller (e.g. Raymond and Zimmerman, 2016; Whalen and Zimmerman, 1987). For service providers, it is essential that the opening of the telephone call is efficient and effective. Stokoe (2014, p. 258) points out that openings are one key element in institutional telephone calls, in which ‘[e]xplaining a service one way may lead to higher client uptake; it can be the difference between winning and losing the race’. In the case of Ufonia, the Dora VUI must be sufficiently effective in the opening moments if the patient is to engage with the service and proceed through the call.

For the Conversation Analyst, one methodological challenge is that members treat their intricate coordinated efforts as simply a by-product of their participation in the event, rather than the event being the minutely coordinated assembly of their co-participation. However, for conversation designers of VUIs the problem is doubly challenging. First, they must be able to identify those finer constituent features that are oriented to as witnessable but common-sense indices of the pattern inventory for a particular action sequence. But they then must build a user interface that gives the semblance of the VUI co-participating with the human client in the production of this event.

In establishing a collaboration between analysts and designers, the current partnership sought to overcome these challenges by developing a novel set of design procedures that allowed EMCA methods and findings to form part of the product development of the conversational agent.

Method

Recordings of thousands of trial calls and calls to patients who consent to be recorded are securely stored on Ufonia’s servers for research and product development purposes. With ethical approval, the research team were given access to anonymised recordings, in which all personal information about patients was removed. A random selection of around 50 calls were transcribed, using conventions developed by Jefferson (2004) and updated for the digital transcription interface CLAN (e.g. MacWhinney and Wagner, 2010). In particular, we use arrows rather than typewriter-based punctuation markers to indicate turn-final intonation contours (i.e. to indicate falling, slight rising, strong rising intonation, etc.). This allows for greater clarity in differentiating between punctuation markers included in the conversation design scripts, and the intonation contouring of the speech production.

An equivalent data set of a small handful of calls between patients and human clinicians were also transcribed and subjected to a sequential analysis. Differences were subsequently identified between the turn formatting of the human clinicians and the Dora outputs. From here, aspects of the Dora system’s scripted prompts were revised through manipulation of text-to-speech (TTS) software (for discussion, see Hazel and Brandt, 2023).

Analysis

Here we examine how a Dora script produced by the conversation design team is converted into spoken output, and how this ultimately plays out in an engagement with the human patient. Before doing so, however, we present and consider the patient’s interactional conduct at the onset of a Dora phone call.

For the post-cataract operation calls presented here, patients are notified in their discharge letter that they will receive a Dora follow-up call three weeks later; and receive a text message a day in advance as a reminder that they will receive a call from Dora at a specified time. However, what they do not know in advance is how a call with Dora may differ from a conversation with a human interlocutor. One example of this is the timing of turn transitions. Typically, turn transition between speakers is almost immediate, with gaps of only up to 0.2 seconds between speakers (Sacks et al., 1974). However, Dora can take up to a few seconds to process a patient utterance, determine an appropriate textual response, and convert that response to talk through TTS. Especially in the initial adjacency pairs of the calls (namely a summons-response in the form of the phone ‘ringing’ and being ‘picked up’ followed by a greetings exchange), we find patients repeating their initial turns.

In Extracts 1 and 2 above, we see the patient produce a canonical response to the phone-ringing summons (Schegloff, 1986), namely they ‘pick up’ the phone and produce a greeting (line 01). The next slot is where the call taker might expect a prompt greeting and self-identification (e.g. Schegloff, 1979, 1986). Instead, due to system latency, this is not immediately forthcoming and is therefore treated as ‘noticeably absent’ (Schegloff, 1968). With this breaching an expected pattern for call openings, the patient is unable to pursue the normative trajectory for the activity, and each responds to this by repeating their greeting. This ends up being produced now in overlap with the beginning of Dora’s ‘delayed’ opening utterance.

Having encountered the system latency in the very first exchange of utterances in the call opening, patients may display a shift in the formatting of their own talk, cancelling one set of expectations (tied to a competent human interlocutor being at the other end of the ‘line’), and resetting the set of expectations according to them speaking into an automated system with unforeseen system limitations. We find that the majority of patients appear to adjust the amount of time they allow before pursuing progression through repetition of their turn. This is evidenced in the following few lines of transcript of both calls, with only the first example repeated below for brevity:

At lines 11 and 17, we see lengthy gaps between the patient’s confirmations and Dora’s follow-on utterance. These are in fact longer than the gaps we saw following the patients’ earlier summons responses. However, the patient no longer treats the ensuing silence as a potential indication of trouble, requiring repair. This suggests that there has already been a recalibration, adjusting his expectancy regarding turn-transition to the speed at which the Dora system works.

Where members have been socialised into the intricate work of a particular activity, here answering a phone call, they are able to produce this again on a next occasion. What the above shows, is that where the automated conversational agent does not perform in line with this expectation, even in terms of the timing of delivery of turns, it leads to a disruption in the flow of the activity and an orientation to the patients having to actively work at discovering the order for this interaction simulation.

The timing of interactional flow as seen here results from the latency introduced by the technological constraints of the system, and VUI designers have little to no control over this issue. However, they can inadvertently introduce silences into the interface where they use text-based punctuation symbols such as commas and full stops in the scriptwriting. We turn to this next, while considering also the ordering of the different types of utterance in these action sequences. Below is an early script for call openings to a cataract surgery follow-up call, as written by Ufonia’s conversation designers:

When the Dora system processes the text through the TTS synthesiser, it necessarily translates the written grammatical elements into features of talk. Pearl (this issue) highlights how different punctuation marks in the TTS script lead the speech synthesiser to generate pauses of varying lengths. We recognise this too, and in addition, note how they can prompt particular intonation contours that are at odds with human speech: so a comma is translated into a short pause (variably 0.2–0.4 seconds); a full-stop into a slightly longer pause (variably 0.5–0.8 seconds) while prompting turn-final falling intonation in the just prior speech, and a question mark into a similar pause along with rising intonation in the just prior speech. When processed through TTS, we observe that there are multiple places where speaker transition might be heard to be relevant (Sacks et al., 1974), in addition to the two which are accounted for and required by the system. The first is the silence triggered by the full stop following the disclosure of the purpose of the call. The second is that triggered after the self-introduction. For the Dora system as it is designed, neither of these sequential positions requires a response. But inadvertently, each of these may potentially project a confirmation of recognition, or a go-ahead response to a pre-sequence as relevant next-action on the part of the call-taker (e.g. Schegloff, 1979).

Indeed, we see this in instances of NHS patients’ engagement with Dora. Extract 4 is the opening of a pre-cataract surgery call, and displays features that we witness across the data set.

The patient produces ‘yes’ in overlap with the beginning of Dora’s identification confirmation request (line 09). Given the placement of this, it is likely to be a go-ahead in response to Dora’s prior utterance, the disclosing of the purpose of the call. Subsequently, Dora’s self-identification (line 15), produced with turn-final falling intonation, is followed by a 0.4 second silence, and the patient produces an acknowledgement, ‘yes’ (line 17). However, it is not a requirement of the system that the patient responds at this juncture, and so this turn is again produced in overlap with Dora’s next utterance.

To summarise, the use of punctuation (commas, question marks and full stops) in the design of the code for TTS introduces intonation contours and silences into the interface speech production that patients treat as indexing possible transition relevance. This leads them to produce utterances in response to Dora. However, as the design does not require these responses, it continues after the short delay, and the patients find their turns being overlapped by a system that has moved on regardless.

In addition to patients finding themselves needing to adjust to the latency of the system and responding to the speech exchange in places where this is unexpected, they may also be faced with a particular order of actions that do not align with those experienced elsewhere in their interactions with human interlocutors. Taking the above cases as an example, we note that the sequence of actions that make up the call opening follows a particular trajectory. We find one after the other: a greeting token, a self-identification, a mention of the reason for the call, a name confirmation request, an introduction (name and role) and then a permission request to proceed. While all these actions may be relevant, the order of the elements may not align with previous patient experience, and this may prompt a patient to have to adjust to a new pattern.

By comparison, here is one illustrative example of how a human clinician opens such a telephone consultation:

Although no call can ever be expected to match equivalent others in every detail, we do find similarities across the cases to which we have access, and which feature different clinicians. Comparing these calls with the Dora design highlights several differences, ones which oblige Dora’s call-takers to adjust the order of their own contributions, by diverging from the sequence organisation they might expect. For example, the clinician here confirms the patient’s identity before announcing the reason for the call. This ensures that confidential information is provided only to the relevant person and not someone else; this early Dora design, however, announces the reason for the call before ascertaining the identity of the speaker. Further, the clinician’s intonation contours follow different patterning than that of Dora. Faced with a pre-programmed conversational agent, the speech of which behaves differently from that to which they are accustomed, patients must themselves adapt their talk to fit the design of the user interface.

In studying the patterns that participants produce in similar clinician-patient interactions, we can identify areas for further development in the conversational design of these equivalent engagements. By focusing on what the patient co-constructs with the clinician we are better able to identify what each party does to produce the pattern that is recognisable to them as embodying this particular activity. This enables us to identify places of divergence and to ensure that the design is developed to further align with the patterns identified in the source data, including at the level of sequence organisation, lexical choice and intonation contouring (see also Hazel and Brandt, 2023).

Having explicated the formatting features of the human clinician calls, the following modified script can be developed:

Using the modified script, the team can trial Dora in its next iteration. An example of how the Dora call opening looks post-intervention, with the design modelled on the clinician data, is the following:

The speech synthesiser is now behaving differently, and is more closely aligned with the patterns produced by our (human) clinicians in equivalent calls. This in turn allows for greater interactional flow with the patient able to project more easily what response is required and when.

Discussion

By their very nature, engagements with AI systems are not the same as human-human interactions. For example, where in human-human calls ‘speakers constantly monitor each other’s behaviours to tailor their contributions in an appropriate and timely manner’ (Pallotti, 2007, p. 10), such complex monitoring and tailoring by the Dora system is not possible. But for the patient, we still see that they engage in the same kind of monitoring and tailoring with Dora as they would be expected to do with a human interlocutor. Avgustis et al. (2021) argue that this shows what humans expect the agents to be capable of processing: they ‘orient to their assumption about an artificial agent’s conversational competence and (in)capabilities, and they adjust their utterances to this knowledge’ (p. 170).

The argument and analysis presented in this paper supports the position that users engage with the system as they would in a human-human interaction, but that there is already a preparedness to adapt when the design of the system diverges from the organisational features of natural social interaction. In such cases, users adapt to the limitations of the system on the fly, for example during the initial moments of the engagement. If the conversation design sector, led by tech companies that advocate that we ‘craft conversations that are natural and intuitive for users’ (Google, 2023), is being tasked with providing a conversational user experience that demands less adjustment on the part of the user, then it needs to adopt a new set of design practices. Rather than imagining how conversation works, and developing our designs on this basis, we must design in a way that is informed by how people talk to each other in that particular setting (Brandt et al., 2023), allowing users to draw on their member’s knowledge of such types of encounters.

By bringing together the examination of natural talk from a human-human equivalent, understandings of EMCA, and using some basic manipulation of TTS and the use of Speech Synthesis Markup Language (SSML), VUIs can be designed to more closely simulate talk produced by humans. ¹ The extent to which this is desirable may depend on the type of VUI and the interactional setting. Where, however, EMCA-informed design can afford these calls a greater level of interactional flow, we would expect to find a greater level of acceptance from the users of the systems.