Natural language processing (NLP), a branch of artificial intelligence, is unlocking groundbreaking advancements in the learning experience. This technology tailors itself to the unique needs of every student and also identifies their emotional cues to personalize the learning experience. This article deliberates on the increasing importance of NLP in education, its applications, and recent developments.
Image credit: Miha Creative/Shutterstock
Importance of NLP in Education
In education, NLP primarily focuses on improving and developing the learning process. This is an effective approach for educators, students, and teachers for analysis, writing, and assessment procedures. Thus, this technology is integrated extensively in diverse educational contexts like evaluation systems, e-learning, linguistics, and research. NLP has also yielded positive outcomes in other educational settings, like universities, higher educational institutes, and schools.
NLP provides the theoretical basis for developing effective techniques and approaches to assist in scientific learning. It promotes language learning and betters the academic performance of students. This technology facilitates the development of educational strategies and software systems, like e-rater and Text Adaptor, which assist in using natural languages for education. An efficient system to manage linguistic input in natural settings through different texts, sentences, and words can be realized using NLP.
Moreover, NLP also utilizes different linguistic approaches and grammatical rules, like tenses, morphemes, corpus, lexicon, and semantic system, in educational settings to ensure that students can better understand the educational curriculum and material.
Although search engines currently provide sufficient information to students, language constraints hinder the process of language learning through online material available on the internet and electronic sources for most students. This issue is effectively addressed using NLP-based approaches.
In e-learning, the application of NLP assists students in developing a general understanding of psychological and cognitive perspectives that have a critical role in language acquisition. Question answering (QA), question construction (QC), automated assessment (AA), and error correction are the key NLP applications for education.
Question Answering
Textbook question answering (TQA) and math word problem (MWP) solving are the two common QA examples in education. TQA refers to a task that requires a system to comprehensively understand the multi-modal information from the textbook curriculum, including diagrams, images, and text documents.
In TQA, the key challenge is understanding the multi-modal domain-specific contexts and the questions, then identifying the important information in the questions. From a technical standpoint, TQA resembles VQA's reliance on question encoding, image understanding, and information fusion.
Conventional VQA studies utilize recurrent neural networks (RNN) to encode the question and convolutional neural networks (CNN) to encode the image. Subsequently, the multi-modal information is combined to understand the questions. Additionally, methods using bilinear pooling schemes, compositional strategies, and spatial attention can bolster the VQA performance. A novel pre-trained machine reader has been developed as a retrofit of pre-trained masked language models (MLMs), which addresses the discrepancy between downstream fine-tuning and model pre-training of specific domain MLMs.
Graph-based parsing methods that extract the concepts from diagrams by converting a diagram to a diagram parse graph were developed to comprehend the tables and diagrams. Optical character recognition (OCR) identifies the chart-specific answers from the charts.
Recently, large language models (LLMs) have displayed significant application potential in different NLP tasks, including TQA, with GPT-3 being the first LLM applied to solve the TQA in 2022. Another study augmented queries for LLMs on TQA to inject general knowledge from the LLMs into specific domains by retrieving evidence from textbooks. A multi-modal model that connects an LLM and a vision encoder for general-purpose language and visual understanding was realized by extending the instruction tuning to a multi-modal field.
MWP solving is a specific type of QA that converts a narrative description to an abstract expression. This task is challenging due to the significant semantic gap in parsing human-readable words into machine-understandable logic. The advent of large-scale datasets has facilitated the extensive adoption of deep learning (DL)--based methods for MWP solving. Most of the methods are based on an encoder-decoder framework, which initially encodes the narratives into a dense vector space and then creates the mathematical expression token as the output sequence. Several methods were also proposed to express a mathematical expression as an abstract tree.
Multiple studies have employed reinforcement learning (RL) to solve MWPs owing to the weak supervision caused by the absence of annotated expressions. Knowledge distillation is also applied to MWP-solving tasks, which learn a smaller model from a pre-trained large generic model. Prompting LLMs with effective planning strategies is an effective approach for MWP solving under the zero-shot/few-shot scenarios.
Question Construction
QC involves constructing questions automatically from a given context, and thus, has a key role in education. Multiple-choice questions (MCQs) are the most common types of questions for a test or quiz. MCQ construction primarily consists of question generation (QG) and distractor generation (DG). QG generates questions depending on a given context, while DG generates distractors to complement the correct ones.
Early general QG methods mostly relied on rule matching. The Seq2Seq model was first applied in 2017 for automatic QG. Based on this framework, methods like RL, leveraging multi-task learning, training a multi-modal QG model, and integrating linguistic features have been augmented to optimize the QG models. Moreover, the LLM GPT-3 and BERT model have been applied to QG with the emergence of LLMs.
Early DG studies were feature- and rule-based, which depended on linguistic rules and specific features such as word frequency or different similarity measurements between correct answers and distractors. Fueled by advancements in DL, distractor creation has evolved into two distinct approaches, including ranking and generation.
Generation-based methods generate distractors token by token automatically, while ranking-based methods consider DG as a ranking task for a pre-defined distractor candidate set. For instance, a hierarchical model was proposed with dual attention for word relevance and sentence relevance and irrelevant information filtering to generate distractors that are traceable in and semantically consistent with the context.
Automated Assessment
In the education domain, AA is an extensively investigated task as it reduces the burden on teachers. AA is primarily categorized into automated code scoring (ACS) and automated essay scoring (AES). AES involves developing a system to score an essay automatically. These systems take the written text as the input and summarize the text quality as a score.
The early AES systems were based on linear regression, sequential minimal optimization, support vector regression, logistic regression, Bayesian networks, and support vector machines. Later, neural networks with various architectures, such as long short-term memory (LSTM), CNN, and attention mechanisms, were proposed to encode the essay and predict the scores.
ACS grades a code snippet's score on multiple dimensions. This is more complicated than essay grading as it involves comprehending the long-dependency inside a code snippet and the relationship between identifiers. Machine learning (ML) algorithms were also applied to ACS, with neural networks, breadth-first search, and clustering approaches being used to track the logic errors in code.
Recent Developments
A study published in the International Journal of Electrical and Computer Engineering (IJECE) presented a pilot prototype chatbot to serve as a learning assistant for the Scratch subject. Scratch is a graphical utility utilized to teach programming concepts to schoolchildren. Student queries were input to the chatbot using a Slack-based UI and an open-source NLP/natural language understanding (NLU) library to obtain the sought explanation as the answer.
The information retrieval and NLP extraction performance of the chatbot were evaluated through a two-stage testing process. The results obtained after the second iteration of testing demonstrated that the prototype chatbot successfully matched and captured the intent and entity for 72% of the test stimuli.
Overall, NLP revolutionizes education by personalizing learning, analyzing writing, and automating assessments. Key applications include question answering and construction and automated grading, with recent advancements like chatbots for programming assistance. However, overcoming challenges like data scarcity and bias, explainability, accessibility, and understanding of nuance and context is essential to exploit the potential of NLP for education fully.
References and Further Reading
Mathew, A. N., Rohini, V., Paulose, J. (2021). NLP-based personal learning assistant for school education. International Journal of Electrical and Computer Engineering, 11(5), 4522. https://doi.org/10.11591/ijece.v11i5.pp4522-4530
Lan, Y., Li, X., Du, H., Lu, X., Gao, M., Qian, W., Zhou, A. (2024). Survey of Natural Language Processing for Education: Taxonomy, Systematic Review, and Future Trends. ArXiv. https://doi.org/10.48550/arXiv.2401.07518
Alhawiti, K. M. (2014). Natural language processing and its use in education. International Journal of Advanced Computer Science and Applications, 5(12). https://doi.org/10.14569/IJACSA.2014.051210