An Exploration of Using Custom GPT Tool to Teach Phrasal Verbs

Framework

This study integrates two complementary frameworks—Cognitive Linguistics (CL) and Sociocultural Theory–based Dynamic Assessment (DA)—to design and evaluate AI-mediated scaffolding for learning phrasal verbs. From a CL perspective, phrasal verbs are not arbitrary pairings but patterned expressions whose meanings are motivated by embodied cognition. Conceptual Metaphor Theory explains how abstract meanings draw on concrete experiences (e.g., MORE IS UP, STATES ARE CONTAINERS), allowing particles like up or in to be interpreted beyond literal space—prices went up; she is in trouble (Lakoff & Johnson, 1980, 1999; Kövecses, 2002). Image schemas—recurrent experiential patterns such as PATH, CONTAINER, and FORCE—provide cognitive “blueprints” that ground figurative extensions (e.g., “walked into a new phase of life”), and visualizing these schemas helps learners notice the embodied logic linking particle senses (Johnson, 1987; Lakoff, 1987; Lakoff & Johnson, 1980; Boers, 2000; Yasuda, 2010). The Principled Polysemy Model further proposes that a particle’s multiple senses radiate from a prototypical core through motivated extensions (e.g., over from spatial coverage to completion/control), giving teachers a principled way to organize meanings rather than list them (Tyler & Evans, 2003, 2004; Mahperkar & Tyler, 2015; White, 2012). Empirical work across varied L1 backgrounds supports CL-informed instruction: metaphor and schema-based approaches improve comprehension, retention, productive use, and even generativity (Yasuda, 2010; Karahan, 2015; Lu & Sun, 2017; Ansari, 2016; Alisoy, 2018; Fallah et al., 2024; González, 2010; Lin, 2024). For instance, orientational metaphors aid recognition (Yasuda, 2010; Karahan, 2015), metaphor association tasks bolster long-term retention (Lu & Sun, 2017), combining diagrams with metaphor improves comprehension and production (Fallah et al., 2024), image-schema work supports flexible creation of novel combinations (Alisoy, 2018), and cross-linguistic contrasts reveal L1-specific challenges (González, 2010; Lin, 2024). Yet these gains often depend on intensive, teacher-led scaffolding and sustained exposure—conditions difficult to maintain in typical classrooms—highlighting a feasibility gap between what works conceptually and what teachers can deliver at scale (Yasuda, 2010; Karahan, 2015; Lu & Sun, 2017; Fallah et al., 2024; White, 2012).

The sociocultural/DA framework addresses how to diagnose and promote development during instruction. DA entwines assessment with mediation to reveal learners’ emerging abilities within the Zone of Proximal Development (ZPD): rather than recording only unaided performance, it provides graduated, contingent prompts that help learners appropriate conceptual tools and strategies (Vygotsky, 1978, 1987; Feuerstein, Rand, & Hoffman, 1979; Lantolf & Poehner, 2014). Two DA orientations inform design. Interactionist DA tailors support moment-by-moment in dialogue, yielding fine-grained diagnosis and promoting self-regulation, but it is resource-intensive (Poehner & Wang, 2020; Ghonsooly & Hassanzadeh, 2019; Kushki, 2022; Nasaji, Kushki, & Rahimi, 2020). Interventionist DA uses standardized mediation hierarchies (from implicit cues to explicit explanations), enabling comparability and scale but risking a loss of sensitivity to individual conceptual paths (Marzban, 2018; Yang & Qian, 2017; Tang & Ma, 2023; Randall & Urbanski, 2022; Nasaji et al., 2020). Mixed findings suggest complementary profiles: interactionist designs can provide finer diagnosis and, in some cases, greater developmental gains, whereas interventionist formats more efficiently strengthen fluency or breadth in larger cohorts (Nasaji et al., 2020; Kushki, 2022; Asl et al., 2024). Computerized DA (C-DA) operationalizes interventionist features in software—graded hints, response/latency logging, diagnostic profiles—to inform teaching while documenting responsiveness (Poehner, Zhang, & Lu, 2015; Yang & Qian, 2017; Urbanski & Randall, 2022; Bakhoda & Shabani, 2019). Studies show immediate performance gains and signs of internalization (reliance on less explicit prompts, increased independent control), and demonstrate the value of process data for planning instruction (Poehner et al., 2015; Yang & Qian, 2017; Urbanski & Randall, 2022). However, C-DA’s efficiencies come with trade-offs: fixed hint ladders can overlook learner-specific meanings-in-progress; “mediated scores” still require human interpretation across tasks and time; and sustained implementation remains demanding (Poehner et al., 2015; Yang & Qian, 2017; Randall & Urbanski, 2022; Bakhoda & Shabani, 2019).

With the advent of generative AI, many C-DA limitations can be re-imagined. Large language models can analyze learner language in real time and generate adaptive prompts aligned with current reasoning, approximating the contingency of interactionist DA while maintaining structured pathways akin to interventionist approaches (Huang, 2023; Liu, 2024). Because learners can engage with an AI tutor frequently and over time, teachers gain richer longitudinal traces—dialogue histories, error patterns, and shifts in prompt reliance—that speak to conceptual growth rather than one-off correctness, and AI can summarize these traces into actionable diagnostics while preserving a role for teacher validation (Zhu & Wang, 2025). Building on these affordances, the present study designs a Custom ChatGPT tutor to scaffold conceptual understanding of common particles (up, out, off, in). The tutor stages learning from implicit to explicit: contextual exposure to authentic uses; guided noticing of particle behavior; metaphorical and image-schema explanations that link senses to embodied experience; categorization around prototypes; controlled practice cueing motivated extensions; and independent application with feedback calibrated to the learner’s moves (cf. Lakoff & Johnson, 1980; Tyler & Evans, 2003, 2004). In DA terms, mediation is both graduated (progressing from indirect hints to explicit explanations) and contingent (adjusted turn-by-turn to the learner’s responses), while the system logs prompt types, help-seeking, revisions, and time-on-task to build diagnostic profiles that can inform teacher follow-up (Urbanski & Randall, 2022). The theoretical synergy is deliberate: CL provides the what of conceptual content (prototypes, mappings, motivated senses), and DA provides the how of tailoring that content to developmental readiness (Lantolf & Poehner, 2014). At the same time, the design acknowledges limits observed in prior C-DA work by retaining an essential human role: teachers review logs to validate whether apparent gains reflect internalization (e.g., transfer to new contexts, fewer or weaker prompts), clarify misconceptions when AI explanations overshoot or skip steps, and decide when to fade support or re-mediate (Poehner et al., 2015; Yang & Qian, 2017; Randall & Urbanski, 2022). In sum, the framework positions generative AI as a mediational tool that can deliver concept-forward, scalable scaffolding without relinquishing the developmental lens: CL makes phrasal verb meanings systematic and inferrable, DA keeps mediation diagnostic and growth-oriented, and AI ties them together through adaptive delivery and analyzable traces, with teacher judgment anchoring interpretation and next steps (Lantolf & Poehner, 2014; Huang, 2023; Liu, 2024; Zhu & Wang, 2025).

Methods

This study performed a qualitative examination of the potential of a Custom GPT tool to scaffold phrasal verb learning among higher-intermediate MELs. The investigation was guided by two research questions: (1) How feasible is Custom GPT as a scaffolding tool for promoting conceptual understanding of phrasal verbs? and (2) How do students perceive the helpfulness and limitations of the Custom GPT tool in supporting their phrasal verb learning?

Participants and Context
This study was conducted at an intensive English program at a private university in the northwestern United States. The program offers coursework and supplemental tutoring services for international students seeking to improve their academic English proficiency. Four MELs, all placed in the higher-intermediate level classrooms, participated in the research. Two of them were native speakers of Vietnamese, and two were native speakers of Chinese. They regularly visited the tutoring center for additional support with class assignments and language development. Participants were recruited through a poster at the tutoring center and an email invitation distributed by program staff. All four chose to participate because they reported finding phrasal verbs particularly difficult to learn and admitted to frequently avoiding using them in their writing and speaking due to the polysemous and idiomatic nature of phrasal verbs.

Design of the customised GPT tool
This study leveraged a new feature called Custom GPT, released in November 2023, on the ChatGPT platform to develop an individualised scaffolding tool for phrasal verb learning. Unlike the default ChatGPT, Custom GPTs are personalised versions of ChatGPT that allow users to tailor the AI's behavior (the Custom Instructions feature) and knowledge (the Knowledge Uploads feature) for specific tasks or domains. We configured our Custom GPT as a phrasal verb tutor for intermediate-level MELs. Through the researcher-designed Custom Instructions, the Custom GPT followed an implicit-to-explicit scaffolding approach – contextual exposures to the target phrasal verb, explaining the metaphorical or special uses of the particle, detecting learner errors, and providing corrective feedback – grounded in Vygotskian sociocultural theory (Aljaafreh & Lantolf, 1994; Lantolf & Thorne, 2006). In addition, we developed knowledge documents for the Custom GPT, which explain the conceptual metaphors in the particles (e.g. in, out, up) based on existing CL scholarship (Lakoff & Johnson, 1980; Rudzka-Ostyn, 2003). For example, the document explains how 'off' conveys a sense of separation, as seen in 'take off' (departure from a surface) and 'cut off' (disconnection), and 'in' represents containment, as in 'bring in' (movement into a space) and 'sink in' (gradual absorption of information). These materials provided definitions of particle meanings and concrete example sentences as knowledge for the text generation by the customised GPT tool.

The scaffolding procedures follow a seven-step progression to guide learners from contextual exposure toward explicit conceptual explanations (see Figure 1). First, Initial Engagement invited students to locate a real-world example of the target phrasal verb to encourage authentic exposure and avoid polysemy. In the Focused Observation step, students analysed the potential particle meanings in additional sentence examples to infer the phrasal verb's meaning from patterns across contexts. In Guided Analysis, students received analogy-based hints to connect everyday experiences with the spatial or metaphorical meanings of particles. Fourth, in the Categorisation Hints step, students were asked to identify the semantic meaning of the particle from several common conceptual meanings of the particle. This task assessed learners’ conceptual understanding before moving towards concrete practices. The next Practice Activity step (a multiple-choice exercise) challenged students to apply their developing conceptual understanding by selecting the most accurate explanation for the particle meanings of the same phrasal verb across various examples. The sixth step, Conceptual Clarification, provided the most explicit scaffolding (Aljaafreh & Lantolf, 1994) by directly explaining both the meaning of the phrasal verb and the semantic function of its particle. This clarification solidifies learners’ understanding by clearly linking phrasal verb meanings to underlying conceptual metaphors. Finally, in the Independent Application step, students created one sentence using the target phrasal verb. This task evaluated whether learners internalised the conceptual meanings for independent performance (Lantolf & Thorne, 2006).

Data collection
The data collection procedures consisted of three stages: pre-interviews, five GPT-based tutoring sessions, and post-interviews. Pre-interviews were conducted individually with each participant before the tutoring sessions. These semi-structured interviews focused on two aspects: (1) participants’ prior knowledge and conceptual understanding of English phrasal verbs and (2) their previous learning experiences of phrasal verbs. Following the pre-interviews, each participant completed five one-hour GPT tutoring sessions held once a week. In each session, learners worked with the GPT tool to learn six phrasal verbs on a given database. A total of 30 phrasal verbs were studied at the end of the intervention, covering four frequently used particles, including up, off, out, and in (Rudzka-Ostyn, 2003). At the end, a semi-structured post-interview was conducted to ask participants to articulate their conceptual understanding of phrasal verb meanings and the role of particles. In addition, learners were invited to share their perceptions and attitudes toward using the Custom GPT tool, including perceived strengths, limitations, and suggestions for improvement.

Data analysis
We analysed the transcripts of the pre- and post-interviews and the GPT interaction logs of each GPT-based tutoring session. The GPT interaction logs were analysed through three interrelated criteria, adapted from holistic evaluation frameworks for AI in language education (Mittal et al., 2024) to assess the Custom GPT's feasibility and effectiveness as a scaffolding tool: (1) technical feasibility, focusing on whether the GPT followed the seven-step scaffolding procedures as designed; (2) conversational intelligence, referring to the model’s ability to provide accurate and contextually appropriate explanations using uploaded materials; and (3) student uptake, examining the accuracy of learner responses and the appropriateness of their self-generated sentences including the targeted phrasal verbs. Together, these criteria were developed to offer a systematic evaluation of how well the Custom GPT supported learners’ conceptual development and independent application of phrasal verbs. The interview transcripts were analysed through thematic analysis (Braun & Clarke, 2006) to examine students’ conceptual understanding of phrasal verbs and identify their attitudes about AI-assisted phrasal verb learning.

Results

Technical feasibility
The Custom GPT demonstrated strong technical feasibility, successfully following the instructed seven-step scaffolding procedure with an overall adherence rate of 97.6 percent. This high rate suggests that the tool reliably delivered structured support to learners across most interactions. All seven scaffolding steps exhibited robustness above 95 percent, except for Step 4, where notable inconsistencies emerged. Specifically, Step 4 occasionally failed to prompt learners to categorise the phrasal verb by particle meaning, either skipping this step entirely or giving less than three possible particle meanings as instructed. Other glitches included prematurely offering a direct definition of the phrasal verb in Step 1 before other scaffolding steps had been completed, misordering Step 2 and Step 3, and occasionally skipping Step 6, a final explanation reinforcing the phrasal verb’s meaning. Importantly, these misfunctions were not evenly distributed across participants. Two students experienced seamless tutoring sessions with no procedural violations, while the other two showed adherence rates of 97.6 percent and 92.9 percent, respectively. Closer examination of the AI–learner interactions revealed that most deviations stemmed from learner-initiated disruptions, such as requesting a new phrasal verb explanation before the previous scaffolding cycle concluded. However, a small number of glitches could not be traced to any user input or identifiable trigger, suggesting some instability in the AI’s ability to consistently retrieve and execute certain scaffolding instructions. Despite these sporadic issues, the overall high fidelity indicates that Custom GPT is a technically robust tool for delivering conceptual scaffolding steps following the researchers’ instructions.

Conversational Intelligence
The Custom GPT exhibited robust conversational intelligence—its ability to generate context‑appropriate and accurate responses at each turn – achieving 97.5 percent accuracy across 840 coded scaffolding moves in 20 tutoring sessions. The tool performed flawlessly in the first two steps – diagnosing each learner’s sentence and providing an initial semantic hint. It maintained accuracy above 98 percent in Step 3, where the AI supplied metaphor‑based visualisation hints to help learners infer the particle’s meaning, and in Step 7, where it evaluated learners’ self‑generated sentences and offered formative feedback. The only glitch in these phases was a mismatch between the explanation provided in Step 3 and the semantic hint given in Step 1. These near‑perfect scores show that GPT reliably guides learners through the instructed scaffolding procedures from initial hinting to meaning negotiation and final sentence production. The main challenges surfaced in Step 4, where GPT prompted learners to categorise the target phrasal verb under a set of particle meanings drawn from the uploaded knowledge. Though Step 4 accuracy remained at a solid 93.3 percent, 8 of the 120 Step 4 instances were problematic: four delivered an incorrect conceptual explanation after the learner selected a particle meaning, and four omitted the step altogether. These glitches show that GPT can still misapply or fail to retrieve the uploaded knowledge when it generates explanations of the particles in the phrasal verbs. These findings suggest the value of human intervention during concept‑heavy explanations. In fact, students themselves noticed these discrepancies during interaction; the researcher then negotiated the correct meaning with them, and learners ultimately achieved accurate understanding. Therefore, human mediation and ongoing monitoring of AI–student exchanges remain indispensable.

Student uptakes
The overall student accuracy rate across Steps 4, 5, and 7 was 90.1%, with individual step accuracies of 92.1%, 93.2%, and 85.1%, respectively. The high accuracy of selective responses in Steps 4 and 5 suggests that students grasped the metaphorical or spatial meanings of the particles in the phrasal verbs. No clear upward or downward trend was observed across the five sessions, indicating that GPT’s explanations were generally accessible from the outset, and no prolonged exposure is needed.

The slightly lower performance in Step 7, where students independently produced example sentences using the target phrasal verbs, revealed nuanced challenges. Out of 120 total responses, 18 were incorrect. Two primary patterns of error emerged. First, some students demonstrated an understanding of the general meaning of the phrasal verb but had difficulty applying it appropriately in context. For example, in one participant’s sentence, “Fight beat up him.” (Session 5, beat up), the intended meaning aligns with the target verb, but the usage lacks contextual appropriateness and natural phrasing. Second, a few students formed a basic understanding of the phrasal verb through GPT’s English explanations, but their comprehension was limited by the lack of support in their first language. For instance, in another participant’s sentence, “She smuggle in a textbook into the lecture class” (Session 1, smuggle in), the structure is appropriate, but the incorrect use of "smuggle" suggests a gap in her conceptual grasp of the verb, likely due to the absence of an equivalent term or insufficient L1 translation support.

MELs’ attitudes toward GPT-based phrasal verb learning
The thematic analysis of post-interview transcripts revealed that students’ perceptions of the GPT-based scaffolding tool clustered around four topics: conceptual understanding of phrasal verbs, usability of the tool, emotional responses, and user feedback. Overall, participants expressed that GPT was helpful in understanding the conceptual structure of phrasal verbs, particularly the semantic role of particles. One student reflected,
[Understanding the preposition] will help me to guess the meaning like, because when I know the meaning of the verb, but we just feel different or difficult to differentiate the preposition. So if you know the meaning of preposition, it will help you understand the whole set.
This comment highlights a key conceptual shift: the learner began to recognise the systematic contribution of particles to overall meaning rather than viewing phrasal verbs as opaque, arbitrary units. The participant's reflection illustrates an emerging strategy of analytical decoding where knowing the base meaning of the verb provides only partial understanding, and the particle plays an essential role in specifying the action's nuance or directionality. Their use of "guess the meaning" suggests that they were moving toward an inferential approach when it comes to unfamiliar phrasal verbs. This suggests that the GPT tool facilitated learners' ability to conceptualise phrasal verbs as analysable lexico-grammatical items and reflects growing metalinguistic awareness that particles are not simple add-ons but are central to the semantic unity of phrasal verbs. This shift is crucial for promoting self-directed learning strategies and greater flexibility in understanding and using new phrasal verbs in context (Qin, et al., 2023).

Students also valued the structured feedback and implicit-to-explicit nature of the scaffolding process. One participant noted, “Using GPT really gives you more contexts and information. ... you know, dictionaries are dead. They just tell you what it is. They don't tell you how and why the phrasal verb means that.” This comment illustrates learners’ appreciation for GPT’s conceptual scaffolding. Unlike descriptive definitions in the dictionary, GPT provided contexts, analogy-based hints, and explanations about the conceptual meanings of the particles, helping learners understand not just "what" but also "why." This encouraged deeper cognitive engagement and fostered learners’ ability to internalise systematic patterns rather than rely on memorisation. Three out of four students also expressed appreciation for the feedback provided by GPT in Step 7. As one student explained, "If I have the wrong meaning or an issue with the sentence, ChatGPT can help me improve the sentence and make it more appropriate." This comment highlights learners’ appreciation of GPT’s corrective and formative feedback. Rather than simply pointing out errors, GPT provided suggestions for improving the sentence or questioning the appropriateness of the phrasal verb in the sentence. The feedback supported students in making sense of phrasal verbs by helping them move beyond basic meaning recognition to using them appropriately and naturally in context.

Despite these benefits, some participants raised concerns about the cognitive load of the tool. As one student said, “I'm really confused sometimes because they give me too much.” This sentiment highlights the potential for scaffolding overload, which may stem from a mismatch between teachers' tendency to provide indirect hints and multilingual English learners’ preference for more direct, explicit feedback. Two of the four participants specifically recommended integrating L1 definitions into Step 6, suggesting that bilingual explanations would significantly aid comprehension. Additionally, several students expressed a preference for shorter, more direct explanations in each step to reduce reading time and minimise fatigue. These comments indicate a need for calibration between the depth of explanation and cognitive accessibility.

Importance

This study demonstrates that Custom GPT can serve as a feasible and effective scaffolding tool for supporting MELs in the conceptual development of phrasal verbs. Analysis of the GPT interaction logs and interviews showed that the tool successfully guided students through structured scaffolding steps and offered dynamic, adaptive feedback that was generally accurate and contextually appropriate. Learners positively perceived the tool’s ability to provide individualised support, reinforce conceptual patterns, and offer immediate corrective feedback, confirming its feasibility for integration into language learning environments. In addition, the findings also showed that students developed a deeper conceptual understanding of phrasal verbs. Learners shifted from viewing phrasal verbs as arbitrary lexico-grammatical units to recognising them as analysable combinations of verb and particle. This conceptual understanding enabled students to make informed inferences about unfamiliar phrasal verbs and use phrasal verbs more appropriately across different contexts. In this way, the Custom GPT tool facilitated immediate learning gains and fostered learners’ development of sustainable meaning-making strategies critical for long-term linguistic growth.

Several pedagogical implications emerge for integrating AI-mediated scaffolding into ESL instruction. First, instruction should emphasise conceptual understanding for systematic, context-dependent grammatical structures like phrasal verbs. Purposeful scaffolding and explicit explanations of the conceptual meanings of phrasal verbs are beneficial for recognising patterns across phrasal verbs. In addition, students should be given opportunities to reflect on and apply their conceptual understanding when they use phrasal verbs in specific contexts. Second, Custom GPT can be effectively integrated into regular or flipped classrooms where learners engage in AI-supported conceptual scaffolding activities outside of class, reserving classroom time for collaborative meaning negotiation and teacher-led applications. This blended approach harnesses AI’s capacity for individualised support while maintaining the crucial role of human mediation in guiding conceptual development. Third, it is critical to explicitly train students to evaluate AI feedback (Al-khresheh, 2024). Although GPT provided high-quality scaffolding, occasional inaccuracies highlight the need to cultivate students’ critical awareness. Students should be taught to cross-check AI outputs, question inconsistencies, and seek clarification, reinforcing the teacher’s role in ensuring conceptual precision. Finally, the success of Custom GPT in supporting phrasal verb learning suggests its potential application to other form-meaning mappings in English. Structures such as verb tense and aspect, modals, or complex prepositions, which similarly require contextual and conceptual understanding, may benefit from structured AI-mediated scaffolding. By strategically leveraging AI for conceptually rich language features, language educators can enhance learners’ autonomous meaning-making abilities while maintaining meaningful human interactions.

References

Aljaafreh, A., & Lantolf, J. P. (1994). Negative feedback as regulation and second language learning in the zone of proximal development. The Modern Language Journal, 78(4), 465–483. https://doi.org/10.1111/j.1540-4781.1994.tb02064.x
Al-khresheh, M. H. (2024). The Future of Artificial Intelligence in English Language Teaching: Pros and Cons of ChatGPT Implementation through a Systematic Review. Language Teaching Research Quarterly, 43, 54-80. https://doi.org/10.32038/ltrq.2024.43.04
Baskara, R. (2023). Exploring the implications of ChatGPT for language learning in higher education. Indonesian Journal of English Language Teaching and Applied Linguistics, 7(2), 343-358. https://doi.org/10.21093/ijeltal.v7i2.1387
Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative research in psychology, 3(2), 77-101.https://doi.org/10.1191/1478088706qp063oa
Kim, S., Shim, J., & Shim, J. (2023). A study on the utilization of OpenAI ChatGPT as a second language learning tool. Journal of Multimedia Information System, 10(1), 79-88. https://doi.org/10.33851/JMIS.2023.10.1.79
Lakoff, G., & Johnson, M. (1980). The metaphorical structure of the human conceptual system. Cognitive science, 4(2), 195-208. https://doi.org/10.1207/s15516709cog0402_4
Lantolf, J. P., & Thorne, S. L. (2006). Sociocultural theory and genesis of second language development. Oxford University Press.
Lantolf, J.P., Xi, J, & Minakova, V. (2021). Sociocultural theory and concept-based language instruction. Language Teaching. Jul. 54(3). 327-342. https://doi.org/10.1017/S0261444820000348
Rudzka-Ostyn, B. (2003). Word power: Phrasal verbs and compounds: A cognitive approach. Walter de Gruyter. https://doi.org/10.1515/9783110197235
Mittal, U., & Mittal, U. (2024). Evaluating Conversational AI Systems for Responsible Integration in Education: A Comprehensive Framework. Authorea Preprints. https://doi.org/10.22541/essoar.172433753.32174636/v1
Moorhouse, B. L., Wan, Y., Ho, T. Y., & Lin, A. M. (2024). Generative AI-assisted, evidence-informed use of L1 in L2 classrooms. ELT Journal, 78(4), 453-465. https://doi.org/10.1093/elt/ccae033
Qin, J., Wu, Z., & Zhong, S. (2023). When concept-based language instruction meets cognitive linguistics: teaching English phrasal verbs with up and out. International Review of Applied Linguistics in Language Teaching, 61(4), 1455-1480. https://doi.org/10.1515/iral-2021-0164
Vygotsky, L. S. (1978). Mind in Society: The Development of Higher Psychological Processes. Cambridge, MA: Harvard University Press.
White, B. J. (2012). A conceptual approach to the instruction of phrasal verbs. The Modern Language Journal, 96(3), 419-438. https://doi.org/10.1111/j.1540-4781.2012.01365.x
Yasuda, S. (2010). Learning phrasal verbs through conceptual metaphors: A case of Japanese EFL learners. Tesol Quarterly, 44(2), 250-273. https://doi.org/10.5054/tq.2010.219945