Artificial intelligence technologies and applications for language learning and teaching

2.1 Natural language processing

NLP allows machines to understand human language and is used to make AI a valuable tool for language learning. It offers the utility of machine translation (MT), wherein a source language is automatically converted to a target language (see Lee, 2023, for a review of MT). Previous research on the application of NLP has focused on methods for assisting learning and learner feedback (Esit, 2011), improving the analysis of learner input for a variety of tasks and the optimal processing architecture (Amaral et al., 2011), and the potential for computer systems to automatically generate activities using methods commonly used by teachers, such as questioning (Chinkina et al., 2020). It informs researchers and educators of learning processes how various textual aspects can influence learners leading to better text selection and construction at different stages of learning (e.g., Monteiro & Kim, 2020) and how different devices can be used to improve learner outcomes (e.g., Pérez-Paredes et al., 2018).

Pérez-Paredes et al. (2018) surveyed 230 teachers in Spain and the UK to examine the teachers’ use and perceptions of NLP technologies as open educational resources (OERs). They found that the teachers’ knowledge about the technologies was generally low and the technologies were under-used while the most recognised tools were online dictionaries and spell checkers. Nevertheless, the teachers showed positive attitudes towards using OER NLP technologies. Pérez-Paredes et al. suggested that language teachers should be educated on the benefits of using NLP technologies and be supported to develop their skills for using them. Chinkina et al. (2020), on the other hand, compared the results of two studies on the use of computer-generated wh-questions. Their results showed that the computer-generated questions were of a similar quality to those designed by a human teacher.

2.2 Data-driven learning

DDL is facilitated by the use of corpora. It is encouraged through students’ own investigation of patterns naturally occurring in their target language. The corpora of language provide learners with authentic linguistic data (Pérez-Paredes, 2022). Researchers have attempted to examine how to improve DLL in practice such as using corpus data for essay writing correction (Tono et al., 2014; Wu, 2021), scientific report writing (Crosthwaite & Steeples, 2022), and extensive reading (Hadley & Charles, 2017); using big data with an inclusive approach (Godwin-Jones, 2021); incorporating DLL into mobile-assisted language learning (MALL) (Pérez-Paredes et al., 2019); and assisting teachers in incorporating DDL into their lessons (Crosthwaite et al., 2021).

Tono et al. (2014) examined the types of words and phrases that were more suited to using corpus data to assist in error corrections. They identified three main categories of writing errors their undergraduate students made: misinformation, addition, and omission. When using corpus data, misinformation errors were the most difficult to correct while addition and omission errors were efficiently resolved. Tono et al. (2014) asserted that corpora are a useful tool, but an overdependency on the tool for all forms of writing correction is not advised. In a different context, Hadley and Charles (2017) found that a DLL approach to improving the reading speed and lexicogrammatical knowledge of low proficiency students resulted in lower improvements than those students who did not use the DDL approach. They suggested a data-directed approach to DDL providing more scaffolding, structure, and assistance would be better suited to low-proficiency learners. Wu (2021) investigated the way in which seven students of English in Chinese Taiwan used a corpus to assist in identifying collocation patterns in essay writing. She found that the students varied in their ability to use various affordances available and underlined the importance of learner training.

Crosthwaite et al. (2021) investigated the methods that nine Indonesian secondary school trainee teachers used to incorporate DDL corpus tools into their lesson plans. They found that, while the teachers generally incorporated DDL into their lesson plans, the teachers did not have the required skills and instructional knowledge to use the tools for directed pedagogical purposes and tasks. They emphasised the need for experts to use corpora, especially at the school level. With a focus on scientific report writing, Crosthwaite and Steeples (2022) investigated 14–15 years old female students’ use of DDL after completing training on the use of DDL and corpus data. They found that the students’ metalinguistic knowledge was not notably developed but their productive knowledge was improved. The students were more inclined to use non-corpus applications when looking for definitions and whole passages. After three months, the students reported that they rarely used the corpora while seven students indicated their preference for and use of non-corpora online tools such as Google.

Boulton and Vyatkina (2021) identified that, between 2016 and 2019, almost 200 empirical studies on DDL were published and most studies concluded that DDL was advantageous for language learning. They pointed out the need for more theory-driven research on DDL and the lack of replication studies. In a systematic review of DDL in CALL research published in five journals during 2011–2015, on the other hand, Pérez-Paredes (2022) reported that only 4.3 % of the 759 articles published in the journals discussed DDL or corpora use in language learning, and the majority (69 %) of the articles investigated learners’ attitudes towards DDL. The review highlighted teacher training, which should be imperative to the successful uptake of DDL, and technical problems, which should be addressed to normalise DDL.

2.3 Automated writing evaluation

AWE provides students with feedback on their written work (Lee, 2020; Li et al., 2017; Link et al., 2022; Liu & Kunnan, 2016; Ranalli, 2018; Zhai & Ma, 2022). It allows students to gain valuable information on the types of errors they make (Link et al., 2014). Chukharev-Hudilainen and Saricaoglu (2016) developed an automated causal discourse analyser and examined its accuracy in evaluating essays written by seven English as a second language (ESL) students at an American university. Their findings supported the use of automated causal discourse analysers as an effective tool for assisting writers. In a study of the use of an automated causal discourse evaluation tool, Saricaoglu (2019) found that there was no improvement in 31 ESL students’ written causal explanations and pointed out the importance of pedagogical choices, teacher training, and combined feedback from the tool and the teacher. Lee et al. (2013) also found that the combination of teacher feedback and essay critiquing system feedback received more positive acceptance than essay critiquing system feedback only. In another study of 75 Turkish EFL university students’ use of AWE, Han and Sari (2022) found that their experimental group, which used both automated feedback and teacher feedback, demonstrated greater gains than those in the teacher-only feedback group.

In a study exploring postsecondary teachers’ use and perceptions of Grammarly (https://www.grammarly.com/), Koltovskaia (2023) highlighted that teacher attitudes and skills were highly influential in the use of AWE. Similarly, Link et al. (2014) found that the implementation of AWE was impacted by teachers’ perceptions, abilities, and desires to strive for best practices. They argued that a benefit for students using AWE tools outside the classroom was the improvement of student autonomy. This point was supported by Barrot (2023) who investigated the use of automated written corrective feedback (AWCF) offered by Grammarly and asserted that AWCF supported autonomous learning. Wang et al.’s (2013) study also supported the benefits of AWE for autonomous learning and improved learner accuracy. In using a mobile version of Grammarly, Dizon and Gayed (2021) found that 31 Japanese EFL students improved their grammar use and lexical richness as a result of using Grammarly although there was no real difference in their syntactic complexity or fluency.

Wilken (2018) investigated two Chinese university students’ use of AWE over a four-week period and discovered that the students valued the identification of errors with first language (L1) glossed feedback and could use the feedback to find errors on their own. She suggested that AWE developers need to build on their resources expanding selections and providing the option to use L1 or only L2. The option for self-correction was also supported by Harvey-Scholes (2018) and Godwin-Jones (2022) as a positive aspect of assisting students working without the presence of the teacher. Jiang and Yu (2022) examined a group of Chinese EFL students’ experiences with automated feedback in their writing activities and emphasised the need for developing students’ awareness of resources and strategies for using automated feedback. They found that explicit error feedback provided greater assistance than generic feedback. Chen et al. (2022) also found that their four Chinese EFL students were highly motivated to improve their AWE scores and spent time attending to language errors such as word selection and grammatical issues.

In a study of 24 Chinese EFL learners’ engagement with automated feedback using eye-tracking, Liu and Yu (2022) emphasised the importance of explicit feedback. Burstein et al. (2016) suggested the development of AWE, which breaks genres into component subconstructs focusing on the development of core competencies in writing a variety of genres. Cotos and Pendar (2016) investigated AWE feedback for discourse in research articles and asserted the need to extend AWE to consider contextual information and sequencing and to develop meaning-oriented feedback systems. Feng and Chukharev-Hudilainen’s (2022) study also focused on genre-specific writing, using a specialised corpus. Their study reported that using a genre-based AWE system improved 13 Chinese EFL students’ use of linguistic features and rhetorical moves for communicative goals in their writing.

Shi et al. (2022) examined the use of evidence in argumentative writing with 29 Chinese EFL students and the Virtual Writing Tutor (https://virtualwritingtutor.com/). They found that collaborative discourse focusing on AWE feedback helped the students improve their writing while the students appeared to pass through three main phases during the process: trustful, sceptical, and critical. Wambsganss et al. (2022) expanded the use of AWE for argumentative writing with automated feedback and social comparison nudging. They found no significant difference in their participants’ writing ability improvements between the automated feedback group and the non-automated feedback group; however, the group with social comparison nudging wrote higher quality texts containing convincing arguments. They speculated that the social comparison nudging facilitated psychological processes such as adapting to norms and comparisons.

2.4 Computerized dynamic assessment

CDA provides learners with automatic mediations (Ebadi & Saeedian, 2015) and allows learners to analyse language-related issues (Kamrood et al., 2021; Tianyu & Jie, 2018). In CDA, corrective feedback (CF) has been commonly discussed as a key topic. CF assists students in gaining feedback on their errors while assisting teachers in gaining a deeper understanding of their students’ ability levels (Ebadi & Rahimi, 2019). The ability of computers to provide appropriate and effective CF has been of interest to researchers with the added benefit of an online version of CF being able to be accessed by many students at the same time. In a small-scale study, Ebadi and Rahimi (2019) used Google Docs (https://docs.google.com/) as a writer collaboration tool in their mixed approach to online dynamic assessment with three EFL university students. Their students reported positive views of the dynamic assessment process although they showed some difficulties in writing more challenging texts.

In an intelligent CALL (ICALL) environment where insights from computational linguistics and NLP are integrated, Ai (2017) investigated the use of graduated CF (i.e., feedback progressing from general and implicit to specific and explicit) with six students learning Chinese at an American university. His ICALL system for the Chinese language could track learners’ microgenetic changes as they worked through iterations of graduated CF to complete an English-to-Chinese translation task. He found that the graduated approach to CF was effective in helping the students self-identify and self-correct a range of grammatical issues (e.g., punctuation, grammatical objects, verb complement) although there were some instances in which the ICALL system failed to provide effective graduated CF and an onsite tutor provided necessary remedies to the students.

Zhang and Lu (2019) investigated the use of a CDA listening test with 19 students learning Chinese at an American university and found that the diagnostic language assessment was successful not only for assessment but also for helping teachers facilitate more individualised support for students. The assessment offered flexibility in location and timeframe for taking the test. With a different focus on CF, Gao and Ma (2019) examined two different forms of computer-automated metalinguistic CF in drills with 117 intermediate level EFL students at a Chinese university. They reported that those who were in CF groups performed better than those who were in a no-feedback group while no significant effect of the CF was transferred to subsequent writing tasks. Yang and Qian (2020), on the other hand, conducted a study of the use of CDA as a teaching and assessment method to promote Chinese EFL students’ reading comprehension and reported that, after four weeks of learning, those who were taught using CDA performed more efficiently than those who were taught using conventional teaching methods.

2.5 Intelligent tutoring systems

ITSs are computer systems designed to provide personalised and interactive instruction to students without intervention from a human teacher. They have been the most common role of AI in language education (Liang et al., 2021). When used in an EFL context, they aim to support FL learning effectively and efficiently (e.g., Choi, 2016). They can be used as supplements to traditional approaches to education or as standalone applications for self-study. They can be used in any educational context with learners of any age (e.g., Xu et al., 2019). They leverage human obsession with digital technology to provide encapsulated learning experiences (Mohamed & Lamia, 2018). There are various types of ITSs (e.g., Bibauw et al., 2019; Heift, 2010) and some use AI and machine learning algorithms to adapt to the needs of users (Jiang, 2022).

ITSs can provide personalised experiences to users by assessing ability, detecting errors, and providing CF and deliver activities to students, which are specifically targeted at what they need to work on, such as pronunciation, vocabulary, or grammar (e.g., Amaral & Meurers, 2011; Choi, 2016). They can also provide a situational context for users. For example, they can provide cultural information related to the language being studied. Choi (2016) argued that an ICALL tutoring system can support the acquisition of grammatical concepts. Xu et al. (2019) conducted a meta-analysis to investigate the effectiveness of using ITSs for students in K-12 classrooms and found that ITSs produced a larger effect size on reading comprehension when compared to traditional instruction. Future research is encouraged to develop more advanced ITSs with more sophisticated NLP to provide more individualised targeted feedback.

2.6 Automatic speech recognition

ASR is a technology that uses AI and machine learning techniques to understand and produce spoken and written text. It is commonly used in software applications that utilise voice recognition and speech-to-text, such as intelligent personal assistants (IPAs), automatic transcribers, and notetaking apps (e.g., Evers & Chen, 2022). ASR is also used on smartphones; when a user dictates a message into a phone and the phone understands the language and performs an action using the language. ASR has progressed rapidly over the last decade, becoming more accurate and widely implemented in a broad range of industries (Daniels & Iwago, 2017). In a review of technology types and their effectiveness, Golonka et al. (2014) stated that the measurable impact of technology on FL learning largely came from studies on ASR.

ASR has generated a great deal of interest in the field of CALL (e.g., Ahn & Lee, 2016; Chen, 2011; de Vries et al., 2015; van Doremalen et al., 2016). The literature (e.g., Chen et al., 2023; Moussalli & Cardoso, 2020; Tai & Chen, 2023) shows that IPAs have great potential to be used as a tool for L2/FL learning. Learners can practice as much as they like in an anxiety-reduced environment (Tai & Chen, 2023). In a study on the evaluation of an IPA, Dizon (2020) found that the use of Alexa (https://developer.amazon.com/alexa) led to an improvement in L2 speaking proficiency. Similarly, Chen et al. (2023) claimed that Google Assistant (https://assistant.google.com/) could be useful for speaking and listening. IPAs are generally accurate at understanding users’ commands (e.g., Daniels & Iwago, 2017; Dizon et al., 2022). It is immediate feedback and natural language use, which make ASR in IPAs beneficial for L2/FL development and improvement.

The use of ASR in messaging apps, software, and websites supports the improvement of L2 pronunciation as users receive immediate, personalised, and autonomous feedback (e.g., Chen, 2011; Dai & Wu, 2023; Dizon, 2017; McCrocklin, 2016, 2019). Bashori et al. (2022) investigated two EFL learning websites that use ASR to provide different types of feedback. Compared to the control group, the treatment group, which used the ASR-based websites, improved not only their pronunciation skills but also their receptive vocabulary. Evers and Chen (2022) presented a practical approach to utilising ASR technology for pronunciation practice. Their EFL students read aloud into the notetaking app Speechnotes (speechnotes.co), which transcribed their speech into text. When they had finished transcribing, they reviewed their mistakes. They indicated that reviewing their mistakes by themselves or especially with someone else was beneficial. Evers and Chen’s study showed how a combination of peer feedback and technology feedback using ASR could improve learners’ pronunciation.

The integration of ASR into apps and software allows the learning experience to become interactive, engaging, and enjoyable, which in turn supports L2/FL motivation (Moussalli & Cardoso, 2020; Tai & Chen, 2023). IPAs such as Alexa and Google Assistant provide students with opportunities for conversation (e.g., Chen et al., 2023; Dizon, 2017). In Evers and Chen’s (2022) study, students showed positive attitudes towards using ASR-based software to work on their pronunciation. McCrocklin (2016) suggested that positive attitudes also lead to autonomous learning as students enjoy ASR-based activities and can do them by themselves. Teachers also had positive perceptions about the use of ASR-based software to improve L2 speaking performance in van Doremalen et al.’s (2016) study. In addition, ASR can be incorporated into games and simulations designed for language learning and can make the environment immersive (e.g., Morton et al., 2012). Forsyth et al. (2019) reported that their students liked interacting with an animated chatting system. When students feel comfortable communicating with an ASR system, the system can reduce the students’ anxiety, increase their willingness to communicate, and has a positive impact on their L2/FL motivation (Ayedoun et al., 2019; Chen et al., 2023; Tai & Chen, 2023).

Another benefit of ASR is that it can personalise learning content according to a learner’s needs and goals. Chen et al. (2023) found that Google Assistant was good for individualised learning as leaners could control the pace and content based on their needs. Related to accented speech, Spring and Tabuchi (2022) reported that Japanese EFL students could improve their vowel-related pronunciation as practicing with the ASR system allowed them to focus on their pronunciation mistakes and correct the mistakes. In a different context, Walker et al. (2011) showed how non-native English-speaking nurses could use a nurse-patient simulator to practice speaking English in a no-risk environment. ASR can also be useful for testing purposes. For example, Cox and Davies (2012) examined the use of oral tests that used ASR to assess the speaking abilities of EFL learners. They found that the tests could be used to predict speaking ability and could therefore be useful in specific situations such as student class placement. Forsyth et al. (2019) argued that it would be feasible to use systems based on ASR for conversation-based assessment such as an animated agent.

A few negative concerns are also noted in the literature. For example, it can be harder for low-proficiency learners to be understood by an IPA (e.g., Dizon, 2017), and, if learners have trouble communicating their command, they often give up (e.g., Dizon et al., 2022). McCrocklin (2019) also reported that some students were frustrated when ASR-based software did not understand their utterances. Even though Cox and Davies (2012) did not find any gender bias, it is possible that some ASR-based software is more accurate for L2 leaners who have specific accents compared to others. In addition, Daniels and Iwago (2017) warned about privacy concerns while explaining that it is not clear what data IPAs store, where the data are stored, and how the data are used. Researchers call for future research that examines which systems are most effective (Evers & Chen, 2022) and how a range of non-native English speakers with various accents can benefit from ASR systems (Bashori et al., 2022; Chen et al., 2023).

2.7 Chatbots

A chatbot, also known as a bot, chatterbot, dialogue system, conversational agent, virtual assistant, or virtual agent, is a software application that interacts with users via chat (Bibauw et al., 2019; Coniam, 2014; Wang et al., 2021) and stimulates human conversations by asking and answering questions via text or audio (Kim et al., 2021). Chatbots are commonly found on companies’ websites in a range of industries such as marketing, healthcare, technical support, customer service, and education, providing targeted services to website visitors (Fryer et al., 2020; Wang et al., 2021). Generally, a user asks the chatbot a question, and the chatbot interprets the input, processes the user’s intent, and then provides a programmed response to the user (Kim et al., 2021; Smutny & Schreiberova, 2020). Chatbots commonly perform form-filling tasks such as collecting information to confirm someone’s identity or information about a problem or an item they want to purchase and then directing them to an answer or preparing the information for a human to easily review.

Chatbots have been around since the 1960s when Weizenbaum (1966) developed ELIZA, a psychotherapist bot. They have been developed considerably since ELIZA. Other notable chatbots include ALICE and Cleverbot. Web-based chatbots have been utilised for several decades and are commonly integrated into messenger apps such as Facebook Messenger (https://www.messenger.com/) (Smutny & Schreiberova, 2020). Chatbots can also have human-like appearances (e.g., Replika [https://replika.com/]) that have social life-like characteristics, which can emotionally immerse users in the experience using text, audio, and other visual cues (Ayedoun et al., 2019). These days, chatbots use techniques such as NLP, pattern matching, and neural machine translation to achieve their goals (Huang et al., 2018; Smutny & Schreiberova, 2020).

Interest in chatbots is rising due to their potential to support L2 and FL learning in interesting ways (Wang et al., 2021). For learning EFL, Huang et al. (2017) developed a dialogue-based chatbot called GenieTutor to target specific areas of language learning interest, such as ordering food, or just to chat freely about any topic. For learning a range of languages, the Mondly chatbot (https://app.mondly.com/) was designed as an additional component of a language learning platform. A chatbot has unlimited patience, can instantly respond to requests using natural language, can lower learners’ anxiety, which encourages willingness to communicate and self-correction if mistakes are made, can focus on specific topics and areas of interest, and does not require a human teacher or interlocutor (Bibauw et al., 2019; Coniam, 2014; Fryer et al., 2020). Students can practice aspects of language that they might not feel comfortable practicing with a human or practice recently learnt language (Fryer et al., 2020).

Goda et al. (2014) showed that the use of a chatbot prior to a group discussion led to an increase in student output and supported the awareness of critical thinking skills. Kim et al. (2021) found positive results when using a chatbot. They specifically found that using an AI bot via text or voice prior to completing speaking tasks led to improved speaking performance. The voice-based chatbot led to greater performance than the text-based chatbot and the face-to-face condition. Ayedoun et al. (2019) argued that, if a chatbot has the ability to perform communication strategies, it can encourage willingness to communicate. In a different context, Coniam (2014) found that chatbots generally provided grammatically acceptable answers to questions. If chatbots can provide teachers with logs of conversations between chatbots and students, the teachers will be able to identify the students’ errors from the logs and plan lessons to fix the errors.

Negative results have also been reported in the literature. There is a concern about the novelty effect of using chatbots to support language learning. For example, Fryer et al. (2017) compared students’ interest in tasks in an FL course between completing a task with a human and completing a task with a chatbot. Their results suggested that a chatbot could provide initial interest due to its novelty, but students’ interest dropped quickly. Smutny and Schreiberova (2020) criticised chatbots for being too mechanic in their behaviour and lacking important communication components. Coniam (2014) also criticised a number of English language chatbots for providing answers that lacked meaning and were not grammatically accurate. Empirical studies that examine the impact of chatbots on L2 and FL learning are still lacking (Kim et al., 2021). Bibauw et al. (2019) called for studies that have more participants and occur over long periods of time. Smutny and Schreiberova (2020) suggested that future research should aim to provide guidelines for teachers to integrate chatbots into their teaching and conduct a content analysis of learners’ conversations with chatbots.

ChatGPT (https://chat.openai.com/) has recently generated great interest in various fields. It produces detailed written responses to requests for information based on vast databases. While ChatGPT has a significant issue with factual accuracy (e.g., Vincent, 2022), its impact on education is being discussed by many educators and researchers (e.g., Illingworth, 2023; Liu et al., 2023; Loble, 2023). Through a pilot study of the use of ChatGPT for writing an academic paper, Zhai (2022) reported that the text written by the AI chatbot was coherent and informative and suggested that improving students’ creativity and critical thinking should be focused on education. If carefully planned and used, ChatGPT might offer a rich opportunity for language teachers to enhance language teaching and create an engaging language learning experience for their students.