Conversational agents in language learning

4.1 General communication practice

One use of conversational agents is to facilitate general communication practice, where learners have an open conversation with agents in English. With built-in autonomous speech recognition (ASR) systems, agents are able to understand and respond to spontaneous speech input, thereby simulating interpersonal communication and serving as conversation companions for English learners.

A growing body of research has explored the potential of employing conversational agents to provide additional dialogue opportunities in second language (L2) classrooms. Moussalli and Cardoso (2020) explored the pedagogical use of conversational agents in providing extra L2 learning opportunities as part of classroom activities. In this usability study, the learners interacted with an Alexa-equipped Echo for 30 min with a set of preset questions and self-generated prompts. Then, they filled out a survey regarding their perceptions and attitudes towards the Amazon Echo and were subsequently interviewed. As indicated in the survey, all the participants enjoyed interacting with the Echo individually. Furthermore, they agreed that the Echo could function as a supplementary pedagogical tool to maximise personalised learning opportunities. The findings suggest that such conversational agents could provide learners with valuable input exposure, output practice, pronunciation feedback and authentic English conversations in a stress-free environment.

Scholars from Japan conducted a series of studies investigating English learners’ autonomous use of Alexa-based smart speakers among university students (Dizon, 2017; Dizon & Tang, 2019, 2020). For example, Dizon (2020) examined the effect of an intervention that integrated an Echo Dot into university-level English courses. Specifically, learners in the experimental group received a tutorial session on how to use the Echo Dot speaker and then received a list of possible commands that they could try out with the speaker. They were also encouraged to generate commands by themselves. The intervention, consisting of 12 min of human–agent interaction each week, lasted over 10 months. The results indicated that, compared with the control group, the experimental group demonstrated a greater gain in English speaking skills based on a speaking proficiency test developed by Payne and Whitney (2002). Nevertheless, the agent did not outperform the teacher in promoting students’ listening proficiency.

Instead of examining the use of a single intelligent speaker, Obari and Lambacher (2019) incorporated Amazon Alexa and Google Home Mini together into a Business English training programme for Japanese university students. Forty-seven undergraduates were divided into two groups. Both groups used mobile learning applications and social media other than Amazon Alexa and Google Home Mini. Specifically, Group 1 used Google Home Mini daily to improve their listening and speaking with the following programs: Best Teacher, Travel English, Let’s Play Around with English and BBC/CNN news. Group 2 used Alexa daily to improve their listening comprehension and vocabulary skills using the following programmes: Kikutan, English Quiz by Arc, Liberty English and Kindle. The post-test results based on the Test of English for International Communication (TOEIC) indicated that both groups achieved higher scores in terms of listening, speaking and total scores. In particular, Group 1 achieved a listening score almost double that of the pre-test, while Group 2 achieved higher gains in reading, which fulfilled the expectations of the initial learning goals. Based on a perception survey, students believed that speaking with a conversational agent benefitted their overall English proficiency.

On top of studies conducted in the formal learning setting, Dizon and Tang (2019, 2020 conducted two case studies that examined university language learners’ in-home use of Alexa. In the first case study (Dizon & Tang, 2019), the learners received a tutorial and then initiated the self-directed use of intelligent speakers over four weeks. Their verbal communication was retrieved through the user history page to allow for further analysis of the interactions. A series of follow-up surveys and interviews indicated that the learners considered Alexa as a favourable tool for English learning, as it provided meaningful L2 interactions and targeted pronunciation feedback and enhanced awareness of the learning gaps. However, these learners did not interact with the agent very frequently, and playing music was the most common interaction pattern (Dizon & Tang, 2019). In the second case study (Dizon & Tang, 2020), the same procedure was carried out, but the 14 participants were given two months to interact with Alexa. Likewise, the study highlighted that students perceived the agent to be fun, effective and useful for English learning. However, many of them did not demonstrate sustained use of the CA for language learning in an in-home environment.

The above studies examined the use of Alexa-equipped agents in learning English as a second language. Meanwhile, the ubiquity and mobility of Google Assistant gave rise to research studies that investigated its potential to reshape the L2 learning experience outside the classroom. Tai (2022) carried out an empirical study to understand the effectiveness of Google Assistant in improving L2 English learners’ speaking skills by comparing the agent–human interaction with human–human interaction in L2 learning. The participants, which consisted of 89 Chinese college freshmen, were divided into three groups: (1) interacting with the Google Assistant (GA group); (2) interacting with native English speakers (L1 group) and (3) interacting with other L2 English speakers (L2 group). Specifically, the GA group interacted with Google Assistant for 10 min twice a week over the course of a semester, and they had to upload the transcript of their interactions with the agent. As for the L1 and L2 groups, the participants interacted with native English speakers and L2 English speakers, respectively, within the same timeframe. The three groups took an English oral proficiency test before and after the intervention. The post-test was followed by a survey and interviews to capture learners’ perceptions and experience regarding their interactions with Google Assistant, native English speakers and L2 English speakers. The results indicated that the GA group and L1 group achieved significantly greater improvement in the oral proficiency test, and the GA and L1 group scored similarly in terms of all the subdimensions tested in the oral test: fluency, content, vocabulary, pronunciation and grammar. It was suggested that Google Assistant could function similarly as a native English speaker to help L2 learners improve their speaking skills. This study reinforced the potential of the intelligent speaker as an efficient conversation partner.

To sum up, most of the studies showed that learners generally embraced the use of CAs in practicing conversation in L2. They believed that the CA had the potential to act as a conversation companion and offer extra learning opportunities both within and outside the traditional L2 classroom. They described their learning experience with the CA as fun and motivating. Another benefit is that human–agent interaction alleviates speaking anxiety compared with talking to human beings. However, due to the short research periods in prior studies, longitudinal research examining how students’ attitudes towards L2 learning with CAs may change over time is needed. In addition, the existing research has explored the use of CAs within formal L2 learning environments. Yet, the ubiquity of CAs within households calls for more research on students’ in-home autonomous L2 learning. A handful of studies on this topic have shown that students failed to achieve sustained use of CAs at home; further research on this topic is necessary.

4.2 Task-based language learning

Another way of using conversational agents in second language learning is incorporating the agent into tasks. To complete the tasks, learners need to communicate with conversational agents in L2. Such meaning-focused communication provides learners with authentic conversation opportunities and spontaneous feedback, which ultimately benefits L2 development (for an example, see Ellis, 2003).

Researchers have explored ways to integrate conversational agents into interactive tasks in the formal English learning environment. A research team in Chinese Taiwan conducted a series of research studies to examine the effectiveness of agent-assisted task-based language learning in promoting adolescent EFL learners’ English proficiency in terms of speaking, listening and willingness to communicate (WTC) (Tai & Chen, 2020, 2022a, 2022b). The team employed Google Assistant (GA) and its associated speakers as the learning devices. They designed interactive tasks based on Google Assistant can-do lists (https://assistant.google.com/learn/). Additionally, they identified several interaction styles of the CA based on these tasks: (1) interviewer (e.g., “What’s your Zodiac sign?”); (2) narrator (e.g., some facts about space); (3) facilitator (e.g., have a movie quiz); (4) interlocutor (e.g., ask to dress right for an outfit idea); and (5) entertainer (e.g., play a song on YouTube). Learners were informed of these interaction styles during a training session before officially interacting with the agent, thereby better navigating the conversation with the agent while completing the tasks.

In the first study, Tai and Chen (2022b) investigated the impact of GA with two different feedback presentation modes on English speaking proficiency among 88 EFL learners in Chinese Taiwan. The participants were randomly assigned to three groups: (1) interacting with Google Home Hub, which provided both audio and on-screen feedback (GA-Hub group); (2) interacting with Google Home Mini (GA-Mini group), which provided only audio feedback; and (3) interacting with teachers and peers in formal classroom settings. Before the intervention, the experimental groups underwent training on how to interact with the agent. Then, all three groups were assigned identical speaking activities that engaged learners to narrate, describe, ask questions, or express opinions. In each session, the participants completed two interactive tasks. In each task, alongside verbal interaction with the agent, they completed a written worksheet. They completed the first task with their peers in a group moderated by the CA, while the second task required them to work with the agent one on one. The results indicated that both experimental groups outperformed the control group, as they received maximized L2 input. In addition, the GA-Hub group, which received both audio and on-screen feedback, exhibited better gains in speaking scores. This highlights the importance of multimodal feedback in agent-assisted language learning.

In the second study, Tai and Chen (2022a) examined the effectiveness of GA with two different media presentation modes of responses on English listening comprehension. They divided 92 adolescent English learners in Chinese Taiwan into three groups: 1) interacting with Google Home Hub with audiovisual on-screen responses (GA-Hub group), 2) interacting with Google Home Mini with the audio-only response (GA-Mini group) and 3) listening to CD players. There were 40 interactive gamified listening tasks for learners to complete. In line with the first study, the participants were assigned two games in each session. They completed a written worksheet in addition to the games. The experimental groups collaborated on the first game but worked individually on the second one, while the teacher led the whole session in the control group. A follow-up interview suggested that learners perceived gamified tasks mediated by the CA to be enjoyable. Furthermore, the statistical analysis showed that the two experimental groups scored higher in listening scores in the post-test, with the GA-Hub group exhibiting a much more significant difference than the GA-Mini group. In particular, the listening comprehension scores of the GA-Hub group improved enormously over time. These findings confirmed the advantage of the CA in resembling meaningful interaction in L2, thus promoting listening comprehension through speaking and meaning negotiation (Long, 2015). Meanwhile, interacting with the CA in different roles grants exposure to different types of input, which leads to better learning outcomes. Additionally, the outperformance of Google Hub validated that multimodal representations of responses helped bridge the missing information, strengthen the connection between form and meaning and enhance information retention.

Apart from English skills, Tai and Chen (2020) explored the impact of Google Assistant on learners’ willingness to communicate in English, an essential dimension of L2 learning and communicative competence. Compared with the two previous studies, they recruited a larger sample size of 112 eighth-grade EFL learners from a junior high school in Chinese Taiwan who were native Chinese speakers (aged 15–16 years). The participants undertook the same intervention procedure as the participants in their other studies: receiving training, completing eight different agent-assisted tasks (i.e., playing games with the agent, open conversation, trying out music commands) and participating in an interview. The findings suggested that such agent-mediated learning environments lowered learners’ affective filter and speaking anxiety, boosted their motivation and increased their WTC. In addition, most participants loved playing interactive games with GA, but not all the tasks enhanced their WTC. They preferred tasks that allowed them to express their opinions and ask questions and those contextualised in topics that overlapped with their background knowledge.

Furthermore, Galvan-Romero (2022) conducted a pilot study to understand the role of CAs in ESL learners’ WTC within the Hispanic and Latin American migrant community. During the intervention, 10 participants used Google Assistant on their mobile phones to complete tasks assigned by a teacher over 10 weeks. The tasks required learners to gather information to prepare for the upcoming class and to conduct conversations regarding daily life topics. Questionnaires were administered before and after the intervention to gain insights into learners’ perceptions of their motivation, anxiety, communication strategies and self-perceived competence. According to the preliminary findings, students reported that they would like a training session to learn more about the features of the CA. It was also reported that some features of human–human interaction were missing from the agent (e.g. asking for repetition or speech adjustment).

Likewise, Hsu et al. (2021) studied the impact of the Amazon Echo on the development of L2 speaking and listening skills among 50 college students in Chinese Taiwan. Both the control group and experimental group received the same vocabulary instruction: text-reading or movie-watching, explicit vocabulary instruction and vocabulary practice. The experimental group received an additional session in which the participant interacted with Alexa with given tasks. They received training at the beginning of this session and then engaged in the CA-assisted tasks, including obtaining information from Alexa, commanding Alexa to perform specific functions and playing games. Students were then asked to fill out a worksheet based on the assigned task. The data analysis showed that doing the tasks with Alexa significantly improved the speaking scores alone, which resonates with Dizon’s (2020) findings. The study also revealed that the agent could reduce speaking anxiety and offer rich oral interactions in L2.

Previous studies either focused on students of similar proficiency levels or overlooked students’ prior proficiency levels. Chen et al. (2020) investigated the learning experiences of students of mixed proficiency levels with Google Assistant. The participants, who were 29 college students, were asked to participate in six tasks with Google Home Hub. The tasks required learners to ask GA various questions for information, command GA to play music and tell stories, play interactive games and conduct an open conversation with GA. A follow-up interview and survey were used to further understand learners’ perceptions. In line with the studies reviewed above, the participants generally deemed that interacting with GA was enjoyable and less stressful than speaking to human beings. They also perceived GA’s utterances to be comprehensible and natural and believed in GA’s potential to improve their English listening and speaking skills. However, their perceptions varied according to their proficiency levels. Higher-level learners were more likely to achieve mutual comprehensibility with the agent, whereas lower-level learners encountered more challenges due to mispronunciation. The less capable learners suggested that more visual aids on the touch panel would help identify the linguistic gap and lead to better communication (Tai & Chen, 2022b).

Rather than using an off-the-shelf CA available on the market, Jeon (2022) designed a CA using Google Dialogflow that targeted younger English language learners in the Republic of Korea. The agent was able to perform four types of responses: corrective feedback, prompts, fallback intent and evaluation. The voice interface can be accessed through tablets. The 36 primary schoolers participated in a 16-week EFL course that incorporated the CA, with three 40-min classes covered every week. Students familiarised themselves with the chatbots over the first two weeks and learnt different topics each week throughout the remaining 14 weeks. In each lesson, students followed the interaction configuration in the following sequence: whole group, small group, and individual. Specifically, the whole class reviewed the target language points and engaged in dialogue practice in the first session. In the second session of the week, the students were given information-gap tasks and completed them collaboratively with their peers. Then, in the third session, the students completed the tasks individually and filled out a worksheet as part of the task. They shared their learning experiences with the customised CA in a follow-up interview, which was leveraged along with user log data on the tablets for data analysis. Generally, in children’s perception, there were three types of affordances the CA was able to provide: pedagogical affordance, technological affordance and social affordance. However, students’ perception of the CA, their L2 level and technology competence affected their attitudes, thus further reshaping the proportion and power of these affordances for language learning. For example, students who perceived the CA to be pedagogically valuable were less likely to be discouraged by technical limitations and more likely to actively interact with the CA, thus generating more meaningful interaction. Additionally, most students preferred to interact with the CA one-on-one, as this interaction is free from peer pressure. It is suggested that teachers adapt the task to accommodate students of different proficiency levels and encourage both human–agent interaction and human–human interaction.

Targeting a younger age group, Underwood (2017) explored the perceptions of 11 primary school–aged EFL students regarding the current and future usage of AI speakers in the classroom. In this teacher-led design-based research conducted over nine months, students frequently used a variety of smart speakers, including Alexa, Apple’s Siri and Google Assistant from the teacher’s end to support classroom activities. They all deemed this process meaningful and interesting. A co-design task was utilised to help the children communicate their ideas on what the agent looked like and how it could be used. They expected the agent to be a friend who could play with them. Unfortunately, the tasks the children wanted the agent to engage in with them were underdescribed.

Most research concentrated on ESL learners’ perceptions of using a CA through interactive tasks, opening the black box of the L2 learning experience from a cognitive and affective perspective. In contrast, Wu et al. (2020) explored the potential of CAs as conversation partners in L2 learning from a usability perspective by comparing the use of Google Assistant for non-native and native English speakers. The participants were 25 university students, of whom 13 were native Chinese speakers and 12 were native English speakers. They had to complete six tasks on a smart speaker and smartphone. The task types included 1) playing music, 2) setting an alarm, 3) converting values, 4) asking for the time at a particular location, 5) controlling device volume and 6) requesting weather information. These tasks were delivered to the participants as pictograms, prompting the speakers to conduct real-world interactions with the agent. After completing the tasks, they participated in an interview regarding their user experience. The findings revealed that the L2 speakers attributed the communication breakdowns to their own poor pronunciation and linguistic knowledge. Additionally, they preferred the smartphone-based agent, as it provided visual feedback and helped identify communication breakdowns while granting them more time to reformulate the utterances. Interestingly, L2 speakers’ demands for multimodal feedback were identified in previously mentioned studies (Hsu et al., 2021; Tai & Chen, 2022a, 2022b).

In addition to students’ perceptions, teachers’ perceptions played an integral role in the integration of CAs into task-based learning in the L2 classroom. Teachers’ beliefs about CAs shaped the ways they were implemented in the classroom. The study of Timpe-Laughlin et al. (2020) stood out among the existing literature, which mostly concentrated on learners’ perceptions. They used a survey and focus group interview to investigate teachers’ perceptions of CA-based tasks and the methods of implementing them in their instruction. The 16 participants, who were ESL teachers, first filled out a demographic questionnaire and completed four intermediate-level CA-based tasks embedded on a website, followed by an interview. The four tasks were: (1) playing a guessing game; (2) ordering at a coffee shop; (3) making a request to a boss; and (4) disputing a billing error. Specifically, the first task required the learners to produce yes/no questions to find out a target character among eight characters preselected by the system. The CA would provide feedback in the form of affirmative and negative answers, and the interaction persisted until the user got the correct answer. In the second task, the learners interacted with the cashier in a coffee shop and ordered one food item and one drink item from the menu displayed on the screen. In the third task, the learners had to call a boss to schedule a meeting and ask them to review presentation slides before the meeting. In the fourth task, learners were asked to call a service provider and dispute an error on their monthly bill. According to the results, all the teachers held positive attitudes towards the CA-based speaking tasks, and the third task turned out to be the most popular. In particular, they valued the authenticity of the critical feature of CA-based tasks and activities, where learners negotiated problems across multiple turns in the conversation. Furthermore, their perceptions were mediated by technology functionality, which directed their comments on whether the CA–human interaction was authentic or not. These teachers proposed two main ways of using the CA: (1) as supplementary practice materials outside of the classroom and (2) as potential diagnostics for formative assessments.

Overall, the majority of the studies utilised the tasks as a lens into learners’ perceptions of L2 speaking and listening with CAs. Learners tended to perceive CAs as human-like and capable of communicating. In particular, the learners preferred one-on-one interaction with CAs over collaborative use in groups, as it created a stress-free learning environment and improved their learning motivation and willingness to speak in L2. Moreover, the learning process mediated by CAs was considered enjoyable, and learners tended to talk when CAs played the role of interlocutor in the conversation. Despite the promising nature of learners’ perceptions, only one study targeted teachers’ perceptions, the findings of which also reinforced the promise of CAs and uncovered their potential uses in task-based learning.

Despite the learning benefits of perception research, the existing literature has not systematically examined actual linguistic improvement through CA-based tasks. The task types were based on the off-the-shelf functions of commercial CAs. The most common tasks assigned to learners were to try out the commands, play games and ask the CA questions, which largely reflected the current available features of CAs. Only a handful of studies have redesigned the dialogue system to fit into specific learning contexts or redesigned the tasks for specific learning purposes. Nonetheless, the tasks emphasised the authenticity of the interaction and strived to promote meaningful output from the learners.

Among the most popular CAs available in the existing literature, L2 learners prefer CAs that provide both audiovisual and screen-based feedback. This multimodal design compensated for the language gap causing communication breakdowns.

4.3 Structured pre-programmed dialogue

A third way of using conversational agents in second language learning is through structured, pre-programmed dialogue. In this work, rather than using off-the-shelf conversational agents, researchers or educators design and programme their own conversational agents to carry out dialogues on particular topics that are deemed relevant for desired learning outcomes.

Some of the most notable work in this area has been carried out by the Converse to Learn group at the University of California, Irvine and University of Michigan. This group has focused on broader goals related to young children’s language and literacy development and content learning, but a major focus of their studies is the particular effects on second language learners. The group programmes its own tailored conversational agents using the publicly available Dialogflow from Google. In one series of studies, they deployed smart speakers and e-books that help young children, aged three to six years old, learn to read by asking them dialogic questions similar to what a parent or teacher might ask (Xu, Aubele et al., 2021; Xu, Wang et al., 2021). The original and follow-up questions follow a script designed to support children’s learning of the story and vocabulary, with follow-up questions that are dependent on the conversational agents’ capacity to interpret children’s responses among a number of possible intents. The studies found that children in the AI-based dialogic condition learn much better than those without dialogue and approximately as much as those with a human asking the same questions, with the greatest benefit accruing to English learners.

In a second series of studies (Xu, Vigil et al., 2022), the team developed dialogic versions of a popular children’s science television programme that children ages 4–6 watch on a tablet or laptop. The main character pauses periodically to ask the children questions, and then, using a similar design as above, continues the conversation based on their response. As above, the children in the dialogic condition learnt more content and vocabulary and were more positively affected than children who watched the traditional show, with English language learners once again demonstrating the greatest gains.

A number of scholars have replicated these types of structured dialogue studies with a more specific focus on second language learners. For example, Lee and Jeon (2022) carried out a study with 67 nine-year-old children studying English as a foreign language in the Republic of Korea. Using a similar approach to the UCI team, they programmed Dialogflow on a smart speaker to respond to a set of restricted questions from children, such as “What colour do you like?” or “Where is the hat?”, and to ask the same question to the children, who would respond by referring to an illustrated worksheet. The conversational agent would carry on a pre-structured dialogue depending on whether the answer was correct, incorrect or not understandable. Then, to assess children’s perceptions of the conversational agent and the activity, they were asked to draw and describe the conversational agent and to discuss their drawings and descriptions in an interview. The study found that more than 70% of the children perceived the conversational agent, although disembodied, as having “human characteristics” and being capable of accurate second language communication. A number of these children, who had very little contact with native speakers of English, perceived that the agent could fulfil their need to have acquaintances who were fluent in English. As one stated, “I like the Google teacher because it would be good to have someone who is good at English around me”.

Finally, a team from Iowa State laid out pathways for designing and developing structured pre-programmed spoken dialogue systems for use in teaching (Chukharev-Hudilainen & Göktürk, 2020) and assessing (Chukharev-Hudilainen & Ockey, 2021) conversational ability among second language learners. For the purposes of instruction, Chukharev-Hudilainen and Göktürk suggested the following approach. First, the design team can create a sample instructional task and then record language learners performing the task. In their case, they recruited 9 adult non-native speakers of English to carry out a conversation about the availability of online learners. Then, they can perform linguistic analysis of this seed corpus to illuminate the kinds of conversational moves that typical conversation on the topic entails. Next, they can programme a conversational agent into a spoken dialogue system that performs similarly in response to the discussion comments, similar to how a human would respond through a set of recombinable templates. Then, they can test the system among themselves before field-testing it with end users.

Chukharev-Hudailainen and Ockey (2021) used a similar approach for developing a second language assessment system, which they field-tested with 42 adult English language learners. About half of the English learners preferred being assessed by a human, whose naturalness, body language, ease of exchanging ideas and better ability to understand their own imperfect English were appreciated. About half preferred being assessed by the conversational agent, whose slower, clear and standardised speech was appreciated.

This third area, structured pre-programmed dialogue, features the fewest papers, which is not surprising given the extensive preparation it involves. This is an expected growth area for future research and teaching.

5.1 For educators

Despite the ubiquity of conversational agents in daily life, it should not be assumed that all students have prior experience with conversational agents or have mastered the skills needed to interact with them. To ensure that everyone is on the same page, teachers are strongly encouraged to prepare a tutorial or training session for students to demonstrate interaction strategies and examples before students formally start dialoguing with CAs. If teachers choose to use a commercial agent, such as Google Assistant or Amazon Alexa, for instructional practices, they can refer to the current features and functionality of these commercial agents, which are available on their respective official websites.

Most of the research examined the use of conversational agents in the classroom setting, yielding promising results about learners’ positive beliefs and perceptions of agent-based instructional activities. Nevertheless, most studies did not explore teachers’ perceptions of CA-based tasks, thus calling for teachers’ input on the implementation of conversational agents in future research. Despite being limited in size, current research on teachers’ perceptions offers some insights insofar as the ways to use the conversational agent in instructional practices (Moussalli & Cardoso, 2020; Timpe-Laughlin et al., 2020). First, conversational agents can be used as potential diagnostic tools for formative assessment. Access to the transcript of human–agent dialogue from the device makes it possible for teachers to retrieve students’ dialogues. Therefore, teachers can use it as a reference to evaluate students’ oral proficiency and provide targeted feedback on their linguistic errors. Second, teachers can use the agent as a classroom assistant while motivating students to speak up, for example, by setting a timer, playing music, checking a fact or asking for definitions or spellings (Underwood, 2017).

Moreover, teachers can use CAs as a supplementary tool for individual practice, as they extend the L2 conversation opportunities outside the classroom. However, if learners use the conversational agent as a supplementary tool outside class, teachers should further consider how to help with sustained use of the CA. As implied by the research investigating students’ self-directed use of the agent, students tended to lose interest in the conversational agent over time (Dizon & Tang, 2020). Managing students’ outside-of-the-classroom learning experience would be challenging but meaningful. For example, teachers can monitor students’ progress by asking them to do a daily check-in and collaborating with their parents, especially with younger learners’ families. Teachers can even gamify the participation mechanics to get everyone engaged.

Moreover, given students’ favourable reactions to the use of CAs in the classroom, educators should strive to integrate these agents seamlessly into classroom tasks. The task types employed in recent research are very limited in variety and are mostly based on the available features of the commercial agents, with a focus on speaking skills. Teachers are encouraged to leverage the resources available to design tasks that force students to produce the target linguistic forms (Underwood, 2017). They should be sensitive to individual differences, such as learner proficiency levels and learning styles, when designing the tasks. According to existing research, younger learners prefer to work with the agent alone; learners of lower proficiency levels tend to participate less actively in the tasks (Jeon, 2022; Lee & Jeon, 2022). Therefore, teachers should provide personalised scaffolding and differentiated instruction in these agent-based tasks. Additionally, they should think about how to support multiple interactions in the classroom, including agent–human interaction and student–student interaction. Finally, most conversational agents do not target L2 learners and specific learning contexts. Teachers are encouraged to redesign the voice interface using the available tools to adapt to their teaching contexts.

5.2 For researchers

Notwithstanding the popularity and versatility of voice assistants, few research studies have explored the potential of conversational agents in the second language learning context. In the existing literature, the majority of the studies were case studies. The limited sample size and scope and the explanatory nature of the literature yielded mixed results. Regardless of this concern, the findings showed that a dominant proportion of learners felt less anxious and more motivated to learn with the agent. However, this might be dependent on the learners’ characteristics and task types. Younger learners prefer to work with agents alone in information gap tasks, as it relieves them from peer pressure (Jeon, 2022). This finding is divergent from Underwood’s (2017), where students spoke more when doing group work with the agent, though the task type is not specified. Adult college students who were involved in more structured tasks enjoyed working with peers because collaboration scaffolded their knowledge gaps (Tai & Chen, 2022b). Current findings highlight the importance of expanding the sample size, involving learners of diverse demographic backgrounds and proficiency levels and experimenting with more agent-based instructional practices across different learning contexts.

Furthermore, most studies examine students’ L2 learning as mediated by the agent on a cognitive and affective level, rather than as tested via learning outcomes. A number of studies imply that L2 learners’ speaking improved after the intervention of human–agent interaction, but the results were not validated by experimental evaluation. There were mixed findings about the effectiveness of conversational agents for listening skills (Dizon, 2020; Tai & Chen, 2022a). Several studies interested in the role of conversational agents in students’ willingness to communicate (Galvan-Romero, 2022; Tai & Chen, 2020) identified positive effects of the agent on this dimension of L2 learning. Future research calls for more experimental research with robust and systematic research methodologies. Apart from listening and speaking skills in general, researchers should pay more attention to other language skills (grammar, vocabulary, comprehension, etc.) and psychological factors that mediate the learning process (anxiety, motivation, self-confidence, etc.).

Most of the studies explored the use of conversational agents in the formal learning setting. Future research needs to shed light on the autonomous use of conversational agents outside the classroom and, if possible, conduct longitudinal studies on this topic. In this way, educators will be able to better locate where the intervention should come in to promote students’ sustained and active use of conversational agents. Additionally, designers can become informed about what scaffolding should be implemented for self-directed use.

Moreover, to gain holistic insights into agent-mediated L2 learning, researchers need to leverage different perspectives rather than focus on learners’ perceptions alone. Teachers’ beliefs about technology shape their instructional practices and design. Yet, there exists a wide gap in this line of research, and only one study examines teachers’ perceptions of agent-based tasks for L2 learning. When interacting with commercial conversational agents, the L2 learners also play the role of users. The findings of recent research indicate that user experience with the voice interface affects their perception along with the learning outcomes (Jeon, 2022). A possible research direction is to borrow perspectives from human–computer interaction to understand user experience.

5.3 For developers

Most commercial conversational agents, such as Amazon Alexa and Google Assistant, are designed for use by native speakers. However, non-native speakers have different interaction patterns than native speakers (Wu et al., 2020). This poses a challenge for the agent in achieving mutual comprehensibility with non-native speakers. According to recent literature, communication breakdowns persist as a barrier to successful communication, which shapes learners’ perceptions and learning outcomes (Chen et al., 2020; Dizon & Tang, 2020; Tai & Chen, 2022a, 2022b; Wu et al., 2020). On one hand, there are unavoidable errors in non-native speakers’ utterances related to syntax, pragmatics and pronunciation, among others; on the other hand, agents cannot recognise accented speech due to the limited coverage of the language models in the current dialogue system. The mutual comprehensibility between the agent and non-native speakers is highly contingent on contextual factors, such as learners’ proficiency level, age and the task type in which the agent is embedded (Chen et al., 2020). Developers should continue to extend language coverage in natural language processing systems and offer improved access to authentic interaction in English for non-native speakers. For pedagogical purposes, the agent should foster multi-turn dialogue instead of completing single-turn requests (Underwood, 2017). Ideally, a multilingual agent that can understand code-switching in both the learners’ native language and their target language might better serve the learning purposes in a language learning setting (Jeon, 2022).

Given the unique characteristics of second language learners, emergent research has begun designing conversational agents targeted specifically at L2 language learners. To personalize the L2 language learning experience, agents should ideally have multiple versions of voice databases for speakers of different proficiency levels (similar to how, for example, there is a ‘Simple English Wikipedia’). Multimodal representation of feedback can compensate for the anxiety caused by communication breakdowns, which could be especially useful for scaffolding low-proficiency learners (Jeon, 2022; Lee & Jeon, 2022; Tai & Chen, 2022a, 2022b). When learners struggle to make sense of the prompts from the agent, they can use on-screen modalities for meaning-making, thereby carrying the conversation forward. Moreover, learners prefer to interact with agents with human-like traits that provide affective feedback and reiterate their utterances (Xu & Warschauer, 2020). They also prefer interacting with agents that play the role of interlocutor in the conversation. Future studies might need to examine how different designs of embodiment and feedback modalities shape language learners’ perceptions and learning outcomes. Overall, to better customise conversational agents to act as language partners for ESL learners, developers might need to collaborate with educators to co-design pedagogical conversational agents.