Written answers to open-ended questions can have a higher long-term effect on learning than multiple-choice questions. However, it is critical that teachers immediately review the answers, and ask to redo those that are incoherent. This can be a difficult task and can be time-consuming for teachers. A possible solution is to automate the detection of incoherent answers. One option is to automate the review with Large Language Models (LLM). They have a powerful discursive ability that can be used to explain decisions. In this paper, we analyze the responses of fourth graders in mathematics using three LLMs: GPT-3, BLOOM, and YOU. We used them with zero, one, two, three and four shots. We compared their performance with the results of various classifiers trained with Machine Learning (ML). We found that LLMs perform worse than MLs in detecting incoherent answers. The difficulty seems to reside in recursive questions that contain both questions and answers, and in responses from students with typical fourth-grader misspellings. Upon closer examination, we have found that the ChatGPT model faces the same challenges.

Predicting long-term student achievement is a critical task for teachers and for educational data mining. However, most of the models do not consider two typical situations in real-life classrooms. The first is that teachers develop their own questions for online formative assessment. Therefore, there are a huge number of possible questions, each of which is answered by only a few students. Second, online formative assessment often involves open-ended questions that students answer in writing. These types of questions in online formative assessment are highly valuable. However, analyzing the responses automatically can be a complex process. In this paper, we address these two challenges. We analyzed 621,575 answers to closed-ended questions and 16,618 answers to open-ended questions by 464 fourth-graders from 24 low socioeconomic status (SES) schools. Using regressors obtained from linguistic features of the answers and an automatic incoherent response classifier, we built a linear model that predicts the score on an end-of-year national standardized test. We found that despite answering 36.4 times fewer open-ended questions than closed questions, including features of the students’ open responses in our model improved our prediction of their end-of-year test scores. To the best of our knowledge, this is the first time that a predictor of end-of-year test scores has been improved by using automatically detected features of answers to open-ended questions on online formative assessments.

Arguing and communicating are basic skills in the mathematics curriculum. Making arguments in written form facilitates rigorous reasoning. It allows peers to review arguments, and to receive feedback about them. Even though it requires additional cognitive effort in the calculation process, it enhances long-term retention and facilitates deeper understanding. However, developing these competencies in elementary school classrooms is a great challenge. It requires at least two conditions: all students write and all receive immediate feedback. One solution is to use online platforms. However, this is very demanding for the teacher. The teacher must review 30 answers in real time. To facilitate the revision, it is necessary to automatize the detection of incoherent responses. Thus, the teacher can immediately seek to correct them. In this work, we analyzed 14,457 responses to open-ended questions written by 974 fourth graders on the ConectaIdeas online platform. A total of 13% of the answers were incoherent. Using natural language processing and machine learning algorithms, we built an automatic classifier. Then, we tested the classifier on an independent set of written responses to different open-ended questions. We found that the classifier achieved an F1-score = 79.15% for incoherent detection, which is better than baselines using different heuristics.

Felipe Urrutia, Cristian Calderon, Valentin Barriere. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

The research focus on teacher questions is justified by previous research emphasizing the essential role of questions in facilitating meaningful learning in science. Analysis of teacher questions has traditionally been based mainly on manual coding, which is extremely labour intensive. In this study we explore how both machine learning and large language models can be used for this purpose. Whereas machine learning approaches involve supervised training with extensive data, pre-trained large language models operate through prompt engineering. The automatic speech recognition text outputs of 23 physics lessons on the same topic from 23 science teachers were analysed with variety of techniques. The results revealed that the large language model approaches improved with few-shot approaches compared to zero-shot ones. Furthermore, few shot approaches outperformed the supervised machine learning approaches, yet human- and hand-crafted approaches continue to demonstrate their relevance. Implications for science teaching and learning are discussed.

Felipe Urrutia, Cristian Buc, Roberto Araya, Valentin Barriere. Proceedings of the 20th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2025). 2025.