PaperlayerView exact revision on arXiv

Source-linked AI summary

A Review of the Trends and Challenges in Adopting Natural Language Processing Methods for Education Feedback Analysis

Thanveer Shaik, Xiaohui Tao, Yan Li, Christopher Dann, Jacquie Mcdonald, Petrea Redmond, Linda Galligan

arXiv:2301.08826v1cs.CL

TL;DR

Educational institutions need ways to analyze student feedback, while the evidence base for adopting NLP methods in education remains distributed across methodologies and applications. This review synthesizes NLP approaches, applications, trends, and challenges relevant to education feedback analysis. It concludes that methods from NLP and other application domains can be adapted to education, while challenges such as domain-specific language, sarcasm, ambiguity, emoticons, and aspect-based sentiment analysis require attention.

  • Problem

    Manual analysis of student feedback is time-consuming, creating a need to examine NLP methods for analyzing educational opinions and feedback.

  • Method

    The paper reviews NLP methodologies and applications including feature extraction, feature selection, topic modelling, summarization, annotation, and knowledge graphs.

  • Results

    The review identifies education-relevant NLP trends and discusses methodologies used to address domain-specific language, sarcasm, ambiguity, emoticons, and aspect-based sentiment analysis.

  • Takeaways & Limitations

    NLP approaches from existing education and industry applications can be adapted for student feedback analysis and related educational applications.

Abstract

from arXiv · show

Artificial Intelligence (AI) is a fast-growing area of study that stretching its presence to many business and research domains. Machine learning, deep learning, and natural language processing (NLP) are subsets of AI to tackle different areas of data processing and modelling. This review article presents an overview of AI impact on education outlining with current opportunities. In the education domain, student feedback data is crucial to uncover the merits and demerits of existing services provided to students. AI can assist in identifying the areas of improvement in educational infrastructure, learning management systems, teaching practices and study environment. NLP techniques play a vital role in analyzing student feedback in textual format. This research focuses on existing NLP methodologies and applications that could be adapted to educational domain applications like sentiment annotations, entity annotations, text summarization, and topic modelling. Trends and challenges in adopting NLP in education were reviewed and explored. Contextbased challenges in NLP like sarcasm, domain-specific language, ambiguity, and aspect-based sentiment analysis are explained with existing methodologies to overcome them. Research community approaches to extract the semantic meaning of emoticons and special characters in feedback which conveys user opinion and challenges in adopting NLP in education are explored.

I. INTRODUCTION

The review frames NLP as a way to analyze student feedback and surveys existing AI and NLP opportunities, methodologies, trends, and education-specific challenges.

  • AI in Education: AI can process structured and unstructured educational data to support prediction, classification, and personalized learning experiences.Examples include learning management systems, deep learning, transfer learning, and NLP-based processing of student feedback.
  • NLP for Feedback: NLP is used for student opinion mining because it interprets textual feedback and can analyze perceptions with limited human intervention.Its applications include feedback analysis, topic modelling, and summarization.
  • Motivation: Manual monitoring of quantitative and qualitative student feedback is time-consuming, motivating NLP-based annotation and summarization.The review connects feedback analysis to educational infrastructure, teaching practices, and learning environments.
  • Research Questions: The review examines generic NLP challenges, current education trends, and methods from other disciplines that could be adapted to education.These questions organize the paper’s investigation of NLP adoption in the education domain.
  • Methods: Deep learning methods such as CNNs, RNNs, LSTMs, and GRUs are presented as techniques used in text classification and NLP.The review also notes continuous-bag-of-words and skip-gram models as deep-learning concepts supporting NLP.
  • Contributions: The paper contributes a synthesis of NLP methodologies for annotating student views and an exploration of adoption trends and challenges.Its stated scope includes feature extraction, feature selection, topic modelling, summarization, knowledge graphs, and annotation.

II. METHODOLOGY

The methodology review covers feature extraction, feature selection, embeddings, classification, and text evaluation methods used to prepare and analyze student feedback.

  • Preprocessing: Feature extraction and selection transform student feedback into quantitative representations before machine-learning or topic-modelling analysis.Examples include BoW, TF-IDF, and word embeddings.
  • Feature Extraction: BoW ignores word order, whereas TF-IDF weights terms by their importance and corpus rarity.These differences affect how textual context and term significance are represented.
  • Word Embeddings: Word embeddings represent similar meanings numerically, reduce dimensionality, and include Word2Vec, GloVe, Doc2Vec, and BERT.BERT uses bidirectional context, unlike fixed word representations described for Word2Vec.
  • Comparative Evaluations: Word2Vec performed better as a feature extractor in one comparison, while random forest was the stronger classifier in that study.The comparison used logistic regression and random forest with 3-fold stratified cross-validation.
  • Comparative Evaluations: TF-IDF outperformed TF-IDF-LSA and TF-IDF-LDA on larger datasets, while TF-IDF and TF-IDF-LSA had similar accuracy on smaller datasets.TF-IDF-LDA had difficulty classifying text because it failed to reduce noise.
  • Deep Learning: A compared fusion model achieved 94.13% accuracy, exceeding LSTM at 86.27%, LSTM+ATT at 87.49%, and multi-head ATT at 90.03%.The study evaluated proposed methods using GloVe and contextualized vectors with dropout-rate experiments.
  • Sentiment Analysis: The review describes fine-tuned BERT sentiment analysis for student course feedback with intra-domain training, a conditional random field layer, and double attention layers.These additions target grammatical constraints and sentiment analysis of student feedback.

B. FEATURE SELECTION

Feature selection reduces the dimensionality of feature representations to simplify models while maintaining or enhancing performance.

  • Feature Selection: Feature selection reduces feature dimensionality, potentially simplifying model complexity while maintaining or enhancing algorithmic performance.With n features, the number of possible feature subsets is 2^n, making exhaustive modelling infeasible as feature count increases.
  • Evaluation Methods: Feature-subset evaluation considers stability, robustness, non-contributing features, and subset size through filter, wrapper, and embedded methods.These categories organize alternative criteria for selecting useful feature subsets.

1) Filter Methods

Feature-selection methods reduce or rank feedback features before modeling, balancing predictive performance, computational cost, and model complexity. The reviewed studies report gains from selected-feature classifiers, including ensemble and SVM approaches.

  • Filter Methods: Filter methods rank features by representative value and remove low-importance features before model training.They use predictive-power or dependence measures such as mutual information and χ2 statistics.
  • Wrapper Methods: Wrapper methods select feature subsets with a predefined classifier and evaluate each subset by classifier performance.Sequential forward selection adds features iteratively from the full feature pool.
  • Embedded Methods: Embedded methods combine filter- and wrapper-like selection during modeling to address filter-method accuracy and wrapper-method computation limitations.Regularization is a common embedded approach that penalizes model complexity to reduce overfitting or variance.
  • Applications and Results: 87.16% accuracy was achieved by voting ensemble learning integrated with Chi-Square feature selection for student feedback classification.The study reduced the attributes to 18 before comparing ID3, J48, Naïve Bayes, and the ensemble.
  • Applications and Results: 85.17% accuracy was obtained by an SVM with a radial kernel, while feature selection raised another sentiment-analysis SVM result to 85.9%.The reported pre-selection random forest accuracy was 81.6%.
  • Applications and Results: Feature-selection approaches can resist noise and avoid irrelevant data during modeling.The review also discusses newer methods compared with existing approaches.

C. TOPIC MODELLING

Topic modelling uses unsupervised techniques to group documents and identify covered topics, with probabilistic and non-probabilistic approaches represented in education-related applications. The reviewed studies frequently use LDA, including for student feedback, while word embeddings outperform TF-IDF in a broader clustering evaluation.

  • Topic Modelling: Topic modelling automatically groups corpus words and documents into topics without requiring training, using probabilistic or non-probabilistic models.Non-probabilistic approaches include LSA and NMF, which operate on bag-of-words representations.
  • Probabilistic Models: LDA models latent topic structures by assigning documents different topic proportions under a predefined number of topics.PLSA instead detects semantic co-occurrence and estimates topic and document-topic probabilities.
  • Applications: LDA was used to identify topic trends in mobile learning, organize extended-reality course dimensions, and extract misunderstood computer-science topics.Applications included technology, learning, learners, and eight StackOverflow question categories.
  • Applications: 20 topics were selected from 59,662 international-student reviews, covering language skills, accommodation, weather, academic burdens, and courses.The topics were ordered by their weight in the review collection.
  • Student Feedback: LDA extracted multiple topics from a single student comment, whereas cosine-based clustering assigned one topic per comment.The same tool reported TextBlob recall of 96.17%, precision of 67.47%, and F-score of 79.30% for sentiment analysis.
  • Evaluation: Word-embedding representations outperformed traditional TF-IDF representations in document-clustering evaluations on Twitter and Reddit benchmarks.The evaluation used NMI, AMI, and ARI across feature representations and clustering methods.
  • Student Feedback: Student feedback mining systems applied clustering and text analytics to categorization, entity extraction, sentiment analysis, and document summarization.One system extracted ten topics and represented each topic with its top five to ten words.
  • Review Findings: LDA was the most frequently used topic-modelling technique because its generative process disambiguates words and aligns keywords with topics.The review reports this pattern across numerous cited studies.

D. TEXT EVALUATION

The paper identifies text summarization, document categorization, text annotation, and knowledge graphs as NLP applications discussed for education.

  • D. TEXT EVALUATION: The text-evaluation subsection discusses text summarization, document categorization, text annotation, and knowledge graphs.

1) Text Summarization

Text summarization condenses expanding student-feedback corpora while preserving critical information. The review distinguishes extractive, abstractive, and hybrid approaches and surveys representative methods and results.

  • Motivation: Text summarization reduces large student-feedback corpora into more digestible summaries without losing critical information.The review frames summarization as a response to the effort required to consolidate rapidly growing feedback.
  • Approaches: Extractive summarization ranks and selects significant original sentences using representations such as BoW, N-grams, and graphs.Selected sentences are retained verbatim after importance-based ranking.
  • Approaches: Hybrid summarization paraphrases top-ranked sentences selected by an extractive component.It combines extractive selection with abstractive generation.
  • Applications: CourseMIRROR clustered extracted feedback phrases and reranked them by student coverage, producing better results than LexRank.The approach targeted significant feedback items and students’ difficulties or misunderstandings.
  • Approaches: Abstractive summarization generates shorter paraphrased representations that preserve the corpus’s information and overall meaning.It produces summaries rather than reproducing the original sentences.
  • Results: ROUGE-1 reached 34.9% and ROUGE-2 reached 17.8% for an LSTM-CNN abstractive system.These scores represented increases of 4.4% and 1.6%, respectively, over existing models.

2) Document Categorization

Document categorization assigns predefined labels to entire documents based on content, intent, and sentiment. Reviewed applications include aspect-based student-feedback analysis and hybrid deep-learning classification.

  • Concept: Document categorization annotates an entire document with a single predefined label based on its content, intent, and sentiment.This differs from entity linking, which labels individual words or phrases.
  • Student feedback: Aspect-based opinion mining evaluated teaching performance across pedagogy, behaviour, knowledge, assessment, experience, and general aspects.Student feedback was labelled by aspect and represented with academic-domain word embeddings before LSTM processing.
  • Deep-learning classification: Li et al. combined LSTM and CNN models for Chinese text classification and compared the hybrid model with a simple LSTM.The experiments used Chinese benchmark datasets and an English BBC news dataset.
  • Results: Accuracy was 91.73% and 94.88% for the compared Chinese settings, and 91.11% and 96.23% for the English settings.The passage reports these paired accuracies for the simple LSTM and proposed BLSTM-C, respectively.

3) Entity Extraction

Entity extraction identifies entities and relations in educational text, while knowledge graphs organize extracted information for integration and learning-path recommendation.

  • Entity annotation: Entity annotation locates named entities, parts of speech, and key phrases using predefined labels.Entity linking can connect extracted entities to larger repositories, with disambiguation linking names to knowledge data.
  • Entity and relation extraction: A proposed architecture detected six entity types and defined seven relation types for scientific information extraction.Entity types included task, method, material, metric, scientific-term, and generic; relations included compare, part-of, and used-for.
  • Knowledge graphs: Knowledge graphs integrate NLP-extracted information from multiple sources and can provide domain knowledge as machine-learning input.They can combine human, automated, and semi-automated data.
  • Learning-path recommendation: A multidimensional knowledge-graph framework stored learning objects in separate classes and defined six semantic relationships between them.Its figure distinguishes inter-class and intra-class relationships and uses node colors for different learning-object classes.
  • Learning-path recommendation: The reviewed framework describes the knowledge graph as able to map different objects and support a well-organized, diverse learning path.The passage presents this as the framework’s reported conclusion.

III. CHALLENGES

Adopting NLP for educational feedback requires handling teaching-context dependence, sarcasm, and ambiguity. The review describes domain-specific training and rule-based, statistical, and deep-learning responses to these challenges.

  • Domain-specific language: Classifying academic datasets and student feedback requires understanding core teaching-context factors.Without domain-specific understanding or training, NLP may fail to uncover latent semantic meaning.
  • Domain-specific language: A domain-specific NER system extracted technology skills from job postings, course descriptions, and MOOCs for personalized course recommendation.Its annotated corpus included StackOverflow and GitHub data.
  • Sarcasm: Sarcasm detection is critical because it affects interpretation of student opinions about course structure and educational infrastructure.Reviewed approaches included rule-based, statistical, and deep-learning methods.
  • Sarcasm: Statistical sarcasm detection can use punctuation, sentiment lexicons, unigrams, embedding similarity, rare-word frequency, and sentiment flips.These features are reported as indicators for distinguishing sarcastic text.
  • Ambiguity: Natural-language ambiguity may be structural, syntactic, or lexical because interpretation depends on context and user perception.The review presents ambiguity in language processing as a central implementation challenge.

D. EMOTICONS AND SPECIAL CHARACTERS

Emoticons and special characters convey emotion in student feedback but create processing challenges for NLP-based opinion mining. Existing multimodal approaches combine text and image processing, while aspect extraction remains an important challenge for fine-grained sentiment analysis.

  • Emoticons and Special Characters: Emoticons and special characters express student emotions but are difficult for NLP systems to process and label with appropriate emotion tags.Prior work addressed emoji processing, emoji anticipation, and query-by-emoji using deep learning and multimodal methods.
  • Emoticons and Special Characters: A multimodal method combined Bi-LSTM text-to-emoji and CNN image-to-emoji models to process emojis and determine sentiment.The approach converted textual and visual emoji information into sentiment-related representations that could be adapted to student feedback.
  • Aspect-Based Sentiment Analysis: Most student-feedback studies classify document-level positive or negative sentiment rather than identifying opinions about different aspects.Aspect extraction surveys separate the problem into aspect extraction, aspect sentiment, and related subtasks.
  • Data Challenges: Data imbalance biases NLP classification because educational datasets often lack massive manually labelled collections from domain experts.When one class substantially exceeds another, deep-learning performance can reflect distribution discrepancies rather than balanced opinion coverage.

IV. DISCUSSION

The discussion reviews AI opportunities and NLP methods that could be adapted to education, especially for analyzing student feedback. It highlights emerging challenges including domain-specific language, sarcasm, ambiguity, emoticons, data imbalance, and aspect-based sentiment analysis.

  • AI Opportunities in Education: AI applications in education include decision-support systems, personalized learning, tutoring, assessment, teacher training, and analysis of multimodal institutional data.The reviewed opportunities span structured and unstructured educational data, including textual feedback, video, and audio.
  • AI Opportunities in Education: AI in education is framed as a layered progression from substitution and augmentation to modification and redefinition, rather than simply using robot teachers.This framework distinguishes broader educational transformation from narrower educational-technology changes.
  • Implementation: Python, Java, and R are widely used for NLP, with language choice depending partly on programming expertise and available package support.Python is described as offering packages for topic modelling, embeddings, classification, and sentiment annotation, while Java supports clustering, tagging, and information extraction.
  • NLP Methodologies: NLP methods reviewed for educational adaptation include preprocessing, feature extraction and selection, machine and deep learning, topic modelling, summarization, annotation, categorization, and knowledge graphs.The review discusses both probabilistic and non-probabilistic topic models, with LDA identified as commonly used for unsupervised topic extraction.
  • Challenges: Education-domain NLP challenges include domain-specific language, sarcasm, ambiguity, data imbalance, emoticons, special characters, and aspect-based sentiment analysis.The reviewed responses include NER, rule-based, statistical, deep-learning, BERT, Unicode conversion, image processing, and multimodal approaches.
  • NLP Methodologies: LDA was the most commonly used probabilistic topic-modelling technique for extracting unsupervised topics from a corpus.The review also identifies VADER, SentiWordNet, query expansion ranking, and product-review analysis as approaches potentially adaptable to student feedback.

V. CONCLUSIONS

The article reviews AI’s educational impact and synthesizes NLP methodologies that can be implemented or adapted for student-feedback analysis. It concludes that the review is methodologically focused and leaves data, interpretability, pedagogy, and privacy challenges for further work.

  • Conclusions: The study aims to explore existing NLP methodologies adaptable to education, assess AI opportunities, and synthesize methods for processing and annotating student feedback.The literature review used Google Scholar and searched multiple bibliographic databases, with results manually checked for relevance.
  • Conclusions: Most references were published within the last five years, and more than 90% of the citations were journal articles or conference papers.The year-wise reference distribution and source composition are reported as characteristics of the reviewed literature.
  • Conclusions: The review covers feature extraction, feature selection, topic modelling, text summarization, annotation, categorization, and knowledge graphs for education-related NLP applications.It also examines challenges in adopting NLP methodologies for educational feedback analysis.
  • Limitations: The study is confined to AI implementation methodologies and gives less attention to pedagogy concepts.This scope boundary limits the review’s treatment of educational-pedagogical issues.
  • Limitations: Data scarcity, class imbalance, deep-learning interpretability, and privacy-preserving feedback extraction are identified as areas requiring further investigation.The authors note that data-specific challenges and black-box interpretation were not explored and propose privacy-aware future work.
Loading 2301.08826v1…