English as a Foreign Language (EFL) students often struggle to deliver oral presentations due to a lack of reliable resources and the limited efectiveness of instructors' feedback. Large Language Model (LLM) can ofer new possibilities to assist students' oral presentations with real-time feedback. This paper investigates how ChatGPT can be efectively integrated into EFL oral presentation practice to provide personalized feedback. We introduce a novel learning platform, CHOP (ChatGPT-based interactive platform for oral presentation practice), and evaluate its efectiveness with 13 EFL students. By collecting student-ChatGPT interaction data and expert assessments of the feedback quality, we identify the platform's strengths and weaknesses. We also analyze learners' perceptions and key design factors. Based on these insights, we suggest further development opportunities and design improvements for the education community.
Oral presentation skills are crucial for English as a Foreign
Language (EFL) students to develop their overall
communication skills [
ChatGPT 1-assisted tools have the potential to enhance
learning experiences for EFL students by providing
personalized feedback [
In this study, we explore how to efectively integrate ChatGPT into oral presentation practice for EFL students. First, we conduct a focus group interview with five EFL students to understand their needs and preferences. Using these insights, we develop CHOP, a CHatGPT-integrated platform to assist with Oral Presentation practice by providing feedback on users’ rehearsals. We test our platform with 13 students, collecting interaction data and students’ post-survey responses about their experience, then have experts evaluate the quality of the generated feedback. By analyzing the collected data, we identify the potential learning efects as well as the strengths and weaknesses of the platform. We also highlight the factors that may afect the quality and learners’ perceptions of the feedback, suggesting design improvements for further educational applications.
Previous studies have explored enhancing EFL students’
presentation skills through various methods. Workshops
and interventions have been used to increase practice
opportunities and self-confidence [
ChatGPT, a large language model-based chatbot powered
by OpenAI, has significantly advanced language learning
[
ISSN1613-0073 on users’ oral performance according to a specific rubric. Furthermore, we shed light on oral presentations rather than casual conversations, which require deep contemplation of unique aspects in oral presentations, including but not limited to the content and organization of the presentation. To the best of our knowledge, our work firstly explores how ChatGPT can be integrated into oral presentation practice to provide holistic feedback on presentation rehearsals.
We conducted a focus group interview with five EFL
students in South Korea to gain deeper insights into their need
for oral presentation assistance. The interview details are in
Appendix B. The main challenges for EFL students when
giving oral presentations are limited vocabulary, nervousness
or lack of confidence, appropriate formality, and correct
pronunciation. We also discover that they prefer direct,
specific, and negative feedback, which aligns with the
findings of [
Given that PPT-based oral presentation can enhance EFL
students’ essential soft skills [
Based on established presentation rubrics [
We use the gpt-4-turbo-preview model with a temperature of 0 to generate rubric-based feedback. Details on the system prompt for ChatGPT are in Appendix D. In the system prompt (Table 4), we provide ChatGPT’s role as a presentation aid, the presentation context (e.g., where and to whom the presentation is being presented, and summarized slide notes. We then utilize Whisper 4 to transcribe rehearsal audio and integrate this into the feedback prompt (Table 5), along with the presentation rubric, feedback style, and an example to ensure consistent format. For feedback on the delivery component, we use SpeechSuper 5 to assess the rehearsal audio on elements, such as pace, fluency, and pronunciation. It outputs scores (0-100) on each element and word- and sentence-level pronunciation feedback. We provide all interim results to ChatGPT for feedback generation on delivery.
Our platform ofers two feedback settings to the students:
partial rehearsal feedback and full rehearsal feedback.
Partial rehearsal feedback, a default setting of our platform
(Figure 1), allows students to practice any specific segments
of their presentations and immediately receive feedback.
According to previous studies, students prefer receiving
feedback with less delay as it allows them to interpret the
feedback while the context is still fresh in their working
memory [
However, applying the same detailed feedback approach to longer rehearsals, i.e., full rehearsals, proved impractical
Grammar Vocabulary Content
Delivery Organization 7.0 6.5 6.0 5.5 5.0 4.5 4.0 Level of Detail Accuracy Relevance
Helpfulness
due to the overwhelming specificity and volume of
comments. To resolve this issue, we separately create a full
rehearsal mode, where we prompt ChatGPT to aggregate
common errors and emphasize the types of errors rather
than every single instance. This metalinguistic feedback
allows users to receive concise yet informative guidance
that supports long rehearsals [
We carry out an experiment spanning two weeks in which 13 EFL students utilize our platform for their oral presentation practice. The details on the students’ backgrounds are in Appendix A. Throughout this period, we collect all interaction data, including rehearsal audio, ChatGPT-generated feedback, user ratings, chatlogs, and platform logs. We then conduct a post-survey to further our understanding of the users’ experiences and satisfaction with the platform. Details on the post-survey are in Appendix C.
To assess the quality of the generated feedback, we recruit 10 English education professionals, each holding a Secondary School Teacher’s Certificate (Grade II) for English Language, licensed by the Ministry of Education in Korea. They evaluate the feedback based on accuracy, level of detail, relevance, and helpfulness, using a 7-point Likert scale. Throughout the experiment, a total of 289 feedback samples were generated for each presentation criterion. We randomly select two partial rehearsals and one full rehearsal from each participant, resulting in 39 feedback samples to be evaluated. Each sample is assessed by three experts, taking the average of three as the final score.
In this section, we assess the efectiveness of our platform with feedback quality (§4.1) and learners’ perceptions (§4.2). We then gain deeper insights by analyzing the influencing factors (§4.3).
Figure 2 shows English experts’ evaluation results of the feedback quality. Across all presentation criteria, the feedback excels in accuracy and relevance, scoring a minimum of 5.66/7 and 5.8/7, respectively. This indicates that the platform is capable of accurately addressing errors that are relevant to the given presentation rubric.
In particular, our platform is particularly efective in providing feedback on vocabulary, as perceived by both experts and students. It achieves the highest scores from experts for accuracy (6.17/7) and helpfulness (6.02/7). Students rate the feedback on vocabulary as the most helpful out of all criteria according to the post-survey (5.25/7). Students also rate 5.5/7 on average regarding the platform’s efectiveness in enhancing their overall vocabulary skills.
Nonetheless, our feedback system shows its limitations in the delivery component, which received noticeably lower scores for level of detail (4.81/7) and helpfulness (4.81/7). Both experts and students report instances of ambiguity within the delivery feedback. Students (S2, S12) find it dififcult to understand the exact rationale behind difering delivery scores across rehearsals. This is due to the limitations on the interpretability of black-box SuperSpeech API, which lacks details in the rationale of the scores. This indicates that the quality of feedback on delivery could be enhanced via improvements in the speech assessment tool’s transparency and capability to extract more granular information, such as pinpointing exact moments in the rehearsal when the pace was too fast.
Feedback on the content and organization components shows comparable scores to other criteria in terms of quality. Nevertheless, students (S6, S11) report that feedback on content and organization components requires greater efort to interpret and is more challenging to implement. This shows that the degree of dificulty and time needed to apply feedback can vary across presentation criteria, as perceived by learners. Thus, the feedback quantity or delivery mechanism should be flexibly adjusted for each criterion to prevent information overload.
We gain further insights into learners’ perceptions of the platform and potential learning efects from the post-survey results. Students rate the platform’s impact on improving their confidence and reducing nervousness at an average of 5.26 out of 7. S2 mentions that revising notes based on the feedback received increases their credibility, thereby boosting their confidence. This highlights the platform’s potential to enhance learners’ confidence, addressing a significant challenge for EFL students. Regarding the feedback’s role in helping them recognize their strengths and weaknesses, students provide an average rating of 5.27 out of 7. S3 points out that a major benefit of the platform is making students aware of unrecognized habits. This suggests the platform’s efectiveness in developing self-evaluation skills through a cycle of practice, feedback, and reflection. Furthermore, students rate how closely the full rehearsal mode resembles a real online presentation environment and its efectiveness in practicing online presentations, with average ratings of 4.92 and 6.33 out of 7, respectively. This indicates that the platform can potentially serve as an efective tool for online presentation practice, which is becoming increasingly common in today’s environment.
While expert evaluations demonstrate the platform’s
potential to provide quality feedback, we further explore factors
that afect feedback quality and learners’ perceptions by
analyzing usage patterns and post-survey results.
4.3.1. Usage Patterns by Presentation Note Type
Preliminary Analysis Presentation notes are crucial for
helping students efectively deliver their content when
giving oral presentations. Proficient English speakers,
including EFL students, often utilize brief keyword notes to avoid
reliance on a full manuscript [
These observations underscore several implications for
platform design and feedback mechanisms. The platform
should ofer flexible feedback tailored to each student’s note
type. For instance, in our current platform, key points users
must remember specific vocabulary and grammar feedback
and recall it from memory for future rehearsals. This is
because of the dificulty of incorporating such detailed
feedback into their notes. Therefore, we can enhance their
learning experiences by providing feedback memory-eficiently,
making it easier to incorporate into their subsequent
rehearsals. This could involve using metalinguistic feedback
to help learners engage more cognitively with the feedback,
thereby enhancing its retention [
Manuscript users (S5, S11), on the other hand, show interest in separate features for note-specific feedback (e.g., “Give me feedback on my notes.”). This is because in our current platform, in order to receive additional feedback on their notes after revision, the student must rehearse it again, which is ineficient. This suggests that creating separate feedback features for the notes and rehearsal delivery could lead to more eficient learning for manuscript users.
Partial Rehearsal
Full Rehearsal 0Vocabulary Grammar partial rehearsals for detailed segment-by-segment refinement, then finish with full rehearsals for a comprehensive review. According to the post-survey, numerous students (S2, S4, S6, S9) recognize the value of partial rehearsal feedback in honing minor details thanks to its specificity, while full rehearsal feedback is praised for providing comprehensive insights, given its wide-ranging context (S2, S6). These ifndings underline the distinct yet complementary roles of each feedback mode.
Implications and Design Suggestions Challenges emerge concerning the context interpretation of full rehearsal feedback. Despite the comparable quality of feedback in both modes as evaluated by experts, students prefer partial over full rehearsal feedback in ratings from both the platform usage experiment and post-survey (Figure 3). Further, students (S3, S6, S12) report dificulties in pinpointing the exact context of errors addressed in the full rehearsal feedback. Students struggle to apply the given feedback due to these challenges in understanding their context. This issue likely arises due to two factors: (1) the lengthy nature of the rehearsal audio which leads to dificulties in locating the error, and (2) the metalinguistic approach of addressing error types rather than each specific instance, initially incorporated to minimize feedback quantity. Therefore, to improve the efectiveness of feedback for long rehearsals, the platform design should consider the appropriate amount and detail of feedback and ways to help the learner interpret the context of feedback. Potential UI enhancements include linking feedback to specific timestamps in rehearsal audio or transcripts, linking error types with their instances, and using various forms of media feedback, such as audio feedback.
Meanwhile, students (S3, S4, S11) report that partial rehearsal feedback sometimes contains too much trivial information. S3 claims that it even leads to distraction and impacts their confidence. This is likely due to the iterative nature of presentation practice on our platform, which involves multiple rehearsal and feedback cycles, implying that the feedback volume could afect the learner’s practice flow. This suggests the need to carefully manage both the volume of feedback and the significance of errors, especially for partial rehearsals, which require more frequent iterations. For instance, dynamically adjusting the threshold of error severity for which feedback is provided, or providing feedback by order of significance could help ensure the feedback is more helpful and relevant to the learner. 4.3.3. Repeated Feedback Requests Another issue emerges when a student requests feedback from the same slide multiple times. Students commonly practice this to rehearse the slide repeatedly and to request confirmatory feedback on the slide notes after making revisions. Students report instances where the feedback appears inconsistent with previous feedback (S2, S7) or remains unchanged despite revisions being applied (S4). Such inconsistencies could be mitigated by enhancing ChatGPT’s context awareness of previous feedback for specific slides and optimizing prompting methods. This will enable ChatGPT to reference earlier feedback better and thereby ofer more consistent feedback. 1–8 4.3.4. Presentation Topic Complexity Students report a challenge in ChatGPT’s inability to grasp the technical depth of presentation topics (S1, S6, S9, S11), all of their topics pertaining to science and engineering ifelds. Upon feedback sample analysis, the issue manifests in two ways: technical terms being incorrectly transcribed by Whisper and ChatGPT’s inherent lack of expertise. For instance, ChatGPT may advise the student to replace an irreplaceable technical term (e.g., “The word ‘sampling’ is overused. Consider replacing the second instance to ‘collecting’ or ‘gathering.’ ”). Enhancements could include supplying the speech recognition model and LLM with additional expert knowledge and ensuring more accurate feedback in technical domains.
In this section, we address the limitations of our work and opportunities for future improvement. We test the platform on a relatively small sample of 13 EFL students, potentially afecting the generalizability of our findings. Additionally, the platform only utilizes the user’s presentation notes and rehearsal transcript to generate feedback using ChatGPT. In other words, it ignores other useful input signals, including but not limited to the raw rehearsal audio, PPT file, or the presenter’s visual cues such as eye contact and gestures, which could ofer valuable non-linguistic features. Incorporating such information to provide more comprehensive feedback could be an avenue for future improvement. Lastly, as ChatGPT is a black-box language model, the feedback generated by our platform lacks transparency and clear rationale. We recognize the need for further research to develop models capable of producing more explainable feedback.
In this paper, we introduce CHOP, a ChatGPT-based interactive platform for EFL students, to provide personalized feedback on oral presentations. CHOP provides students with improvements in confidence, vocabulary, self-assessment, and online presentation skills. The feedback generated from our platform is perceived as helpful in vocabulary, content, grammar, and organization by both English experts and students. CHOP addresses various students’ needs in feedback mechanisms based on their practice patterns. We identify design considerations for personalized feedback needs of students to efectively integrate ChatGPT into EFL education. This study contributes to EFL education by providing insights into learners’ perceptions and key design factors, which is invaluable for the further development of ChatGPTbased platforms for oral presentation practice.
1–8
This section shows the demographic and educational backgrounds of the students who participated in the experiments. All students are from the Republic of Korea, and their average age is 24.5. Table 3 describes the demographic details of the participants.
In this section, we describe the design of the preliminary focus group interview conducted to find the needs and preferences of EFL students with oral presentations. Students S1, S2, S3, S4, and S5 participated in the interview. The following paragraphs describe the questions used in the interview.
Experience with Oral Presentations 1. How did you prepare for your oral presentations? 2. What do you think are the key elements that contribute to an efective oral presentation? 1–8 3. What challenges did you encounter while preparing for and delivering oral presentations? 4. How did you attempt to overcome these challenges? Feedback Preferences 1. Below are feedback samples with diferent styles: one is direct and the other is indirect. Which do you prefer and why? 2. Below are feedback samples with diferent styles: one is positive and the other is negative. Which do you prefer and why? 3. Below are feedback samples with diferent styles: one is vague and the other is specific. Which do you prefer and why? 4. Below are feedback samples with diferent styles: one is straightforward and the other is polite. Which do you prefer and why? 5. Below are feedback samples with diferent styles: one is immediate and the other is delayed. Which do you prefer and why?
B.2.1. Dificulties Encountered in Oral Presentations Students face various dificulties in preparing for oral presentations. S1, S3, and S4 state that a common issue is the overuse of filler words (e.g., “however” and “that”) when presenting. This results from limited vocabulary, nervousness leading to unexpected delivery errors (e.g., speaking too fast and mispronouncing words), and dificulties in composing lengthy, well-structured sentences. S2, S5 also struggle with choosing appropriate language that matches the formality of the presentation context. When preparing their script, due to a lack of expert resources or reference materials, S2, S3, and S4 have trouble determining the appropriateness of their word choice, sentence structure, and overall script quality. S3 and S4 claim that feedback from translation tools often appears unnatural, yet they have no reliable means to verify or correct this.
B.2.2. Approaches to Address Challenges S1, S2, and S5 address these issues by reviewing recordings of their own presentations and learning from them. They watch online videos to understand native and natural expressions. S4 and S5 create synonym banks from blog posts to enhance their vocabulary. However, they note that these resources are not always reliable as they are often produced by other non-native English speakers. To compensate for the lack of expert advice, S2, S3, and S4 consult peers who are more proficient in English. While S2 and S4 have tried using ChatGPT, they consider it time-consuming and inefective, as it requires detailed explanations of the presentation context to get useful feedback.
B.2.3. Criteria for an Efective Presentation We asked students what they perceive as an efective presentation and what it consists of. S1 and S2 highlight delivery aspects to be important factors, such as natural pronunciation and maintaining a fluent pace. S2 and S3 believe a good presentation goes beyond correct pronunciation or delivery; it must convey information efectively and engagingly to the audience. Unlike essays, presentation scripts do not have to be fancy; instead, they should be clear and concise. B.2.4. Desired Assistance S4 and S5 expressed their need for comprehensive assistance in writing the script and detailed feedback on delivery aspects like pace, accent, and body language. S1 and S3 also wanted help improving the content of their presentations, such as ensuring logical flow and completeness of information. Furthermore, S2 and S4 wished for a tool that could accurately understand the nuances of both Korean and English to provide more accurate and relevant feedback.
In this section, we describe the design of the post-survey conducted on the 13 platform users. Students are asked to answer a 7-point Likert scale and open-ended questions related to their perceptions and experience of using the platform. The following paragraphs describe the questions used in the post-survey.
Student’s Experience with the Platform 1. How efective was the feedback in the following areas: Vocabulary, Grammar, Content, Delivery, Organization? 2. How efective was the feedback during partial/full rehearsal? 3. How reliable did you find the feedback? 4. How well did ChatGPT understand the content of your presentation? 5. What aspects did you like or dislike about the partial/full rehearsal feedback? 6. What was the feedback particularly helpful for? 7. What was the feedback particularly unhelpful for? Potential Learning Efects 1. To what extent did the platform improve your confidence and reduce nervousness in presenting? 2. To what extent did the platform help enhance your vocabulary? 3. To what extent did the platform help you identify your strengths, weaknesses, and habits in presenting? 4. To what extent did the full rehearsal practice page resemble an online presentation environment? 5. How useful was the platform for practicing online presentations?
This section describes the prompts used to guide ChatGPT’s responses.
Table 4 shows the system prompt, which includes ChatGPT’s role and the context of the presentation. It also includes the summarized notes of each slide to provide ChatGPT with knowledge of the presentation content, necessary for providing accurate feedback on the content and organization components.
Table 5 shows the feedback prompt. It contains the rehearsal information, transcript, presentation rubric, and a description of the feedback style. This description guides ChatGPT in providing feedback that is specific, negative, and direct for partial rehearsals and groups common error 1–8 types together for full rehearsals. The feedback prompt for the delivery component is separately designed to incorporate the speech assessment results.
Table 6 shows the presentation rubric provided to ChatGPT
for feedback generation. We take the established
presentation assessment criteria [
You will be given the user’s presentation rehearsals, and you will provide feedback on them. The user is presenting at a <presentation location> in front of <target audience>. Here are the summarized contents of each slide of the presentation. ###Summarized slide contents: <Summarized slide contents> Grammar, Vocabulary, Content, Organization Delivery The following is the transcript of a rehearsal from slide <start slide> to <end slide>: The following is the transcript of a rehearsal from slide <start slide> to <end slide>: ###transcript: <transcript> ###transcript: <transcript> Using the transcript, provide feedback on this rehearsal according to the following rubric: The following are the assessment results of the delivery of the rehearsal: ###Presentation rubric: <presentation rubric> ###Assessment results: <Delivery scores and brief explanation> ###Feedback style: <feedback style> The final output should be in the following format: ###Output example: <Output example> Description