Exploring Vague Language Use In Human-Agent Interaction

1. Introduction1.1 Initial OverviewMuch of our daily lives increasingly involve digital interaction of awide variety. Computers have moved far from the stationary desks ofthe home and office alike, and digital devices now permeate a widerspace. Smartphones, tablets and other devices have saturated theworld of personal mobile computing, and our interactions with themhaving become ever more sophisticated. A large part of thissophistication comes from the research and development into agents –software that, amongst other features, displays degrees of autonomy,social capabilities and sometimes humanlike characteristics(Wooldridge and Jennings, 1995). Our collaboration with theseintelligent agents is known as human-agent interaction or HAI. This isessentially a sub-field of human-computer interaction or HCI, whichencompasses digital interactions with agents, computers and othermachines. Many of the theories and literature discussed throughoutthis thesis will often discuss HCI specifically, but there is extensivecrossover between these two fields. Because of this many of thetheories, hypotheses and ideas are transferrable from one to the other.Moreover, there is also some crossover in the specific types of agentsbeing discussed and how the results presented may influence researchinto them. This agent is focused on speech as a modality, but othersmay have multimodal capabilities. Any fundamental findings then canbe considered for these other agents too, but their multimodality andany other differentiating features also have to be considered.With the greater prevalence of human-agent interaction comes theneed to address challenges in how our relationships with agents willdevelop as their sophistication increases as they are given moreautonomy, responsibility and varied roles in collaboration (Jennings etal., 2014). One of the most salient of these roles is that of the agentinstructor. Already there are a wide variety of agents that instruct usevery day. Map based applications in smartphones and satellitenavigation systems in cars direct people across cities and the country,automated self-checkout use in supermarkets has boomed (Orel andKara, 2014), and telephone based spoken dialogue systems have beena mainstay in society for over a decade now (Nass and Moon, 2000).Because most interactions in HAI involve the agent instructing thehuman, successful communication requires that humans be open tobeing directed (Sukthankar et al., 2012), and able to engage withagents at a peer level (Maes, 1994). Agents are capable of dealing withsome types of information in quantities and complexities that wouldoverwhelm humans (Ball and Callaghan, 2011), and it is in thesesituations that they are ideally suited for a role as instructor, making11

quick decisions with vast amounts of data. Moreover, agents provide acheaper alternative than employing human beings. A lot of work hasbeen done on the role agents can play in the management of complexand information rich situations such as emergencies (Schaafstal et al.,2001) and damage control (Bulitko and Wilkins, 1999). Agents havealso been shown to be able to hold more advisory roles such as apersonal tutor (Heylen et al., 2003) or by assisting patients andmedical staff in diagnoses (Doswell and Harmeyer, 2007; Chan et al.,2008). Evidently, there is a vast area in agent instructors that are bothcurrently deployed and in development in the laboratory.Many of these agents use speech as a key mode of interaction with itsusers and signify a shift towards a greater use of natural language inagent interfaces – i.e. using language in a more natural form as itappears in conversations and interactions between humans (Cowan etal., 2015) . This presents unique challenges in understanding theeffects of spoken discourse in human-agent interaction, as speechcontains a wealth of interactional complexities that build and maintainthe way people see each other in terms of power, identity andpersonality (Goffman, 1967; Goffman, 2002; Cameron, 2001;Coulthard, 2013). Identity here is seen as the “social positioning of selfand other” (Bucholtz and Hall, 2005: p.586). With the frequency inwhich we interact with such agents and its impending rise, it isimportant to know how changes in this spoken discourse affect thehuman-agent dynamic for both users and developers alike.How human-agent interactions differ from their human-humancounterparts is particularly important. Our interactions with agents,computers and other media have been shown to be similar to that ofother humans in that the same social rules underpinning both areinstinctive in nature, and draw from the same social resources (Nass etal., 1994; Nass et al., 1995; Nass et al., 1996). This allows researchersto take inspiration from existing theories in linguistics, psychology,sociology and communication theory amongst others, and apply themto interactions with agents and other technologies. This includesadorning agents with humanlike features seen in humancommunication and interaction and seeing how users perceive themand interact with them in laboratory and real world contexts. Aspectsof this human likeness include in the specific area of verbal agentsinclude manipulating language (Clark et al., 2014; Strait et al., 2014;Rosé et al., 2008) and vocal capabilities such as the specific voice andprosodic properties being used to convey information to users(Dahlbäck and Jonsson, 2010; Tamagawa et al., 2011; Grichkovtsova etal., 2012). Despite having this wealth of human interaction to drawfrom, there is no guarantee that human-agent interaction will be thesame when these are used to design verbal agent instructors.Instruction giving in humans can be a delicate process. Being the socialactors that we are, there is a desire to not infringe upon the personal12

space and rights of others by asking them to do something, and not topresents ourselves in a negative light in the same breath (Goffman,2002). We often desire to save face in these situations. To helpmitigate these and attempt to build and maintain a rapport withinterlocutors, there are a number of linguistic strategies used tomanipulate the potential adverse effects of giving instructions, such asthose outlined in politeness theory (Brown and Levinson, 1987). Therehave been attempts to research both the general effect verbal agentinstructors have on their users in a game setting (Moran et al., 2013),as well as the effects of mitigated instructions in both pedagogicaltutoring (Wang et al., 2008; Wang et al., 2010) and task-basedscenarios (Torrey, 2009; Torrey et al., 2013; Strait et al., 2014).Although they are sometimes used successfully, for example in thepedagogical setting, but in task-based scenarios have received mixedresults. At times, this type of linguistic strategy in agents makes themseem more considerate, kind and likeable, whereas in others it asdeemed as inappropriate for the interaction.The approaches used in this type of research have provided interestingresults but there is not always a consistency in how the agent’scommunications have been designed and implemented. Politenessstrategies for example are broad and many and can include greetingsand praise, as well as face saving (Brown and Levinson, 1987). Withoutexplicit description of the language being used it, nor a consistency, theresults can be hard to put into context as to which linguistic featuresare causing the specific positive and negative points of discussion thatarise from the data.This thesis presents initial steps in creating a linguistic framework ofimplementation based on a different phenomenon – vague language(VL). VL refers to types of language that are inherently imprecise andused to achieve a variety of interactional and social goals (Channell,1994). The categories of VL that Channell refers to VL for example arevague approximators such as like, about, a bit of; vague categoryidentifiers such as and so on, and stuff; and placeholder wordsincluding thingy and thingamy. Different authors have describeddifferent categorisations of VL and these are discussed further inChapter 2. This type of language is different from vagueness that arisesfrom genuine uncertainty. VL can be used as a mark of social cohesion(Cutting, 2007), and can help towards creating an informal and lessdirect atmosphere (Channell, 1994; Cheng and O’Keeffe, 2014).There are several reasons why VL was employed in this thesis as apotentially useful linguistic strategy to be employed by verbal agentinstructors. It is a common feature of spoken interaction in particular,although it does also appear in writing (Channell, 1994; Jucker et al.,2003; Cheng and O’Keeffe, 2014). Given the agents described in thisthesis are primarily of a spoken nature, and that there are a growingnumber of such agents that people currently interact with as outlined13

previously, VL represents a good candidate for investigation in thistype of spoken interaction. Although it does not always communicateeffectively and may sometimes leads to miscommunications ininteraction (Cutting, 2007; Jucker et al., 2003), VL is also neithernecessarily good or bad. Rather, it is either appropriate orinappropriate for the context in which it is used (Channell, 1994). Howappropriate VL and other language use is within interactions, and theextent facework and politeness functions, can be affected by a varietyof factors, including gender, social status, and cultural background.One of the key research aims of this thesis is to ascertain from initialinvestigations as to whether or not VL can be appropriate or not forthe use in a verbal agent instructor.The focus in this research is to first create an explicit VL model thatoriginates from the lexical level – that is it starts with individual wordsand phrases and not broad strategies as seen in some politenessresearch in HAI. Drawing on previous literature and attempt tocategorise VL, this model presents a description of the categoriesdeemed appropriate for this research context and what lexical itemsthese contain. Previous literature is discussed in both Chapters 2 and3, while the specifics of the model are outlined in Chapter 3 alone.Chapter 3 also includes the general approach to applying VL to a HAIcontext for the research studies in Chapters 4 and 5.Primarily this research presents initial steps in creating linguisticanalysis frameworks for understanding human-agent interaction, aswell as a linguistic focused methodology for investigating thecomparative effects of vague and non-vague language, as well assynthesised and human recorded voices, on participants’ perceptionsand attitudes towards a verbal agent instructor. There are also benefitsfor the designers of these. The results presented here provide anotherinitial step into some of the preferences users display towards thesetwo variables, which may be taken into consideration for futuredevelopments of voice agent technology. Investigating the use of VL inparticular in these interactions allows for insights into whether this isa viable option to improve user experience when interacting withverbal agents, and in a broader sense assess the effects language mayhave on the way in which users perceive agents and project identitiesonto them, as well as any effects this might have on their taskperformance in a specific model assembly context.This thesis also deals with the issue of voice in human agentinteraction. Agents typically have either a synthesised voice or ahuman recording, and there are benefits and drawbacks to both.Synthesised voices can use human recordings to create a database ofnatural speech, from which the appropriate sub-word features can beused to output virtually any utterance (Black, 2002). A text-to-speech(TTS) system can accomplish this, where textual output is turned intosynthesised speech output. Human recordings, on the other hand,14

provide output from a finite list of pre-recorded utterances. The latterrepresents another shift towards human likeness, though synthesisedvoices are becoming more advanced and getting closer to the samepositive perceptions as human recordings receive (Forbes-Riley et al.,2006; Georgila et al., 2012). Both are used in verbal agents and havetheir respective benefits and drawbacks. Although human recordingsare usually of a higher quality and perceived more positively, they aremore expensive and require much more time in preparing. Text-tospeech systems on the other hand are fairly cheap and can produce thesame speech output in a fraction of the time. While comparisons havebeen made between the two there has not been an assessment as tohow it affects a linguistic phenomenon such as VL that is so ingrainedin human communication but not in HAI. Analysing both synthesisedand human speech covers one variable present in speech technologythat already has prior research, though not on the effects of VL use.Any benefits and drawbacks on the use of either type of speech inverbal agent instructors can further contribute to this research, andprovide recommendations on some combinations of voice and VL usein such contexts.Both the language used and how it is produced can influence how oneperceives a speaker, and how they in turn create different identities forthem as a result. Comparing synthesised and human voices with VLallows for the data to inform future agent designers who may explorefurther atypical agent speech.It can also be seen whether this a viable option to improve the userexperience when interacting with verbal agents. This means we cansee whether or not paying a greater attention to language caninfluence the way in which users perceive agents, whether this mattersin regards for their future interaction with them and if there is anyeffect on task performance. The latter is likely only useful in situationswhere human-agent collaboration is non-leisurely, but is a possibilityif not a probability in the future. Similarly, people will be interactingwith agents that can do more and interact with them for longer. This isin essence another continuum from early computers to humaninteraction. If something as simple as manipulating language to make itvague can improve perception of particular attributes, rapport, userexperience, efficiency, clarity etc. then it would be wise to incorporateit. This all has basis in human communication and is not merely anarbitrary inclusion of some imprecise language into human-agentinteraction.1.2 Research Aims and ObjectivesThe core aim of this thesis is to explore the use of VL, a lexical strategyand linguistic phenomenon of human interaction, in the continuouslydeveloping area of human-agent interaction. Specifically this researchlooks at the use of VL by verbal agent instructors in context of them15

guiding human users to complete model assembly tasks. This alsoincludes analysing the effects of synthesised and human voices on theVL used by these agents. This is achieved first by creating a model ofVL that can be implemented into a HAI context. The implementationinto a Lego model assembly task is analysed using a mixed methodsapproach. This approach allows for the both the analysis of attitudestowards the vague agents and the analysis of descriptive accounts ofparticipants interacting with them. Combining quantitative andqualitative approaches, this goes towards developing linguisticanalysis frameworks that can account for how agents and humansinteract in one space in a specific context of interaction. The specificaims of this thesis are as follows:1) How do users perceive and project identities towards verbalagent instructors that use VL and what contrasts can be seenwith human communication?User perception is a fairly general term, but this refers to the mixedmethods approach and analysis of quantitative questionnaires andqualitative interviews. VL exists in abundance in humancommunication but not in HAI, so this leaves a large gap in theknowledge of how successful and appropriate VL can be in thisinteraction space. This is despite research into similar linguistic areas,though there is little information

2.5.5 Language 38 2.5.6 Identity and Adaptability 39 2.6 Language, Identity & Mitigation 40 2.6.1 Mitigation in Human Interaction 40 2.6.2 Politeness in Human-Agent Interaction 44 2.7 Vague Language 46 2.7.1 Defining Vague Language 46 2.7.2 Contexts and Functions of Vague Language 48 2.7.3 General Functions of Vague Language 49