Source-linked AI summary
Reading.help: Supporting EFL Readers with Proactive and On-Demand Explanation of English Grammar and Semantics
Sunghyo Chung, Hyeon Jeon, Sungbok Shin, Md Naimul Hoque
TL;DR
EFL readers lack affordable, on-demand support for understanding complex English texts. Reading.help combines interpretable detection with on-demand and proactive explanations, and the evaluation found 87% average recall when recommending three elements per dimension.
Problem
EFL readers lack affordable, on-demand and situational support for identifying and understanding difficult parts of English texts.
Method
Reading.help combines interpretable models for detecting difficult words and sentences with explanations of vocabulary, grammar, context, and paragraph summaries.
Results
87% average recall was achieved when recommending three elements per dimension, while evaluation involved 5 EFL readers and 2 English-education experts.
Takeaways & Limitations
Reading.help could potentially support EFL readers’ independent English learning when external support is unavailable.
Takeaways & Limitations
The tool provided comprehension support only for selected text rather than the whole document, assuming users could identify where they needed guidance.
Abstract
from arXiv · showhide
A large portion of texts is written in English, but readers who see English as a Foreign Language (EFL) often struggle to read texts accurately and swiftly. EFL readers seek help from professional teachers and mentors, which is limited and costly. In this paper, we explore how an intelligent reading tool can assist EFL readers. We conducted a case study with EFL readers in South Korea. We at first developed an LLM-based reading tool based on prior literature. We then revised the tool based on the feedback from a study with 15 South Korean EFL readers. The final tool, named Reading.help, helps EFL readers comprehend complex sentences and paragraphs with on-demand and proactive explanations. We finally evaluated the tool with 5 EFL readers and 2 EFL education professionals. Our findings suggest Reading.help could potentially help EFL readers self-learn English when they do not have access to external support.
1 Introduction
Reading.help addresses the need for accurate, efficient English reading by providing EFL readers with on-demand, situational support for independently interpreting texts. The tool combines interpretable detection models with LLM-generated explanations and was refined and evaluated through studies in South Korea.
- Motivation and problem: English reading remains crucial because English dominates communication and the production of academic, professional, and social texts worldwide.
- Motivation and problem: Online courses, electronic dictionaries, and translators do not fully provide the on-demand, situational support EFL readers need while reading articles.
- Tool contribution: Reading.help supports independent EFL reading by detecting challenging words and sentences and providing explanations and paragraph summaries.Users can also manually highlight text to request an explanation; a secondary LLM evaluates and refines the information.
- Tool contribution: The tool targets university-educated EFL readers willing to independently interpret English texts and uses natural language processing techniques.
- Evaluation context: In South Korea, the researchers piloted Reading.help with 15 EFL readers, revised it from their feedback, and evaluated it with 5 readers and 2 English-education experts.
2 Background and Related Work
EFL reading difficulties arise from vocabulary, grammar, text complexity, reading strategies, fluency, and domain knowledge, while existing intelligent reading tools largely prioritize reading speed rather than language learning. Reading.help addresses this gap by unifying proactive and on-demand support for vocabulary, grammar, and comprehension using NLP heuristics and LLMs.
- Difficulties for EFL Readers: EFL reading is shaped by fluency, vocabulary breadth, and domain knowledge, while text complexity and reading strategies can cause misinterpretation or slower reading.Text complexity includes difficult vocabulary, complex grammar, and lengthy sentence structures; reading strategies involve inferring context, topics, and takeaways.
- Computational Approaches to Assist Academic Reading: Reading.help unifies vocabulary, grammatical, and comprehension support by proactively recommending potential issues and providing on-demand help when EFL readers encounter problems.The design targets self-learning as well as immediate reading assistance.
- Computational Approaches to Assist Academic Reading: Existing intelligent reading tools improve reading experiences but generally do not address EFL readers’ need to learn syntax and semantics, leaving a gap between reading assistance and language learning.Prior work includes AI-generated questions, critical-thinking prompts, highlighting, visualization-based support, and eye-movement analysis for confusion or mind wandering.
- Computational Approaches to Assist Academic Reading: Existing English-learning tools address proficiency-tailored comprehension questions, real-time vocabulary integration, or critical reading, while Reading.help extends this work with proactive explanations [71].The cited tools include ChatPRCS [71], VocabEncounter, and CriTrainer.
- Work that Augments Writing: Intelligent writing tools motivate this work because reading and writing are related, yet how NLP and LLMs can assist EFL readers remains largely unknown.The work argues that carefully designed NLP heuristics combined with an LLM can capture a wide range of reader problems and provide proactive and on-demand explanation.
3 Design Requirements
The authors iteratively developed Reading.help from preliminary design requirements grounded in prior literature. These requirements specify proactive and on-demand guidance, explanations, hierarchical detail, adaptive support, and visual emphasis for EFL readers.
- DR1 Proactive and On-Demand Guidance: The tool should proactively identify potentially difficult text while also providing guidance on demand, helping readers avoid overlooked complexity or incorrect interpretations.
- DR2 Explain vocabulary, grammar, and semantics: It should explain unknown vocabulary, complex grammar, core topics, and main takeaways to improve comprehension.
- DR3 Drill-down to the Details: Feedback should be hierarchical, initially offering a high-level summary before allowing users to drill down into details [62] [63].
- DR4 User-centered adaptive support: Support should adapt to readers’ differing English proficiency, formal education, and social exposure to English.
- DR5 Visual Emphasis: Visual emphasis such as word highlights should direct attention to salient text and reduce the time needed to find specific keywords [29].
4 Reading.help-lite
Reading.help-Lite is an initial LLM-based prototype that analyzes user-selected text for vocabulary, grammar, and comprehension difficulties. It combines proactive recommendations with interactive explanations and proficiency-sensitive support to inform the final tool’s design.
- Implementation rationale: The prototype uses an LLM analytical pipeline to support preliminary user studies and guide development of the final tool.The authors selected an LLM because of its accessibility and analytical capability.
- Interactive interface: Users can request recommended help or manually invoke tools, then inspect detailed vocabulary, comprehension, and grammar explanations.Features include context-relevant definitions with Korean translations, main ideas and paraphrased examples, phrase-based grammar analysis, and subject-specific keyword tooltips.
- Core analysis: Reading.help-Lite analyzes uploaded text across vocabulary, grammar, and comprehension to identify and suggest potential sources of difficulty.Vocabulary targets unfamiliar keywords or phrases, comprehension situates text in the broader passage, and grammar addresses grammatical issues.
- Prompting and personalization: The system provides explanations through OpenAI’s GPT-4o using role-based, structured prompts that incorporate user proficiency and exact response formats.The prompting design addresses unpredictable responses, hallucination, and verbosity while supporting explanation and recommendation templates.
5 Pilot Study
The pilot study evaluated Reading.help-Lite with 15 South Korean EFL readers to assess its effectiveness and identify limitations. Feedback was positive overall, while usage patterns, module preferences, and trust and comprehension gaps informed the revised tool.
- Participants: The study recruited 15 South Korean adults with bachelor’s degrees who reported difficulty reading English and completed guided tasks and post-interviews.Sessions lasted around 60 minutes, and participants received approximately KRW 15,000 compensation.
- Results: Participants gave positive overall feedback and saw Reading.help-Lite as valuable for personal growth and improvement.The pilot study used Reading.help-Lite as a probe to evaluate effectiveness and identify limitations for improving the tool.
- Usage patterns: Participants sought help most often for vocabulary, averaging 4.2 explanations per passage, followed by comprehension at 2.1 and grammar at 0.9.They used the general help button 1.6 times per selected text to identify issues.
- Post-study interview: All participants found LLM recommendations beneficial despite expecting occasional errors; comprehension and vocabulary were the preferred modules, selected by 9/15 and 8/15 participants.The comprehension module helped participants decode lengthy, complex sentences.
- Limitations: Participants considered LLM responses insightful but difficult to trust, and the tool lacked adequate comprehension support beyond selected text.These challenges informed the revised tool in § 6.
6 Reading.help: Updated Version · 6.1 Interpretable CEFR Prediction Models
Reading.help was revised to address transparency and severity-calibration concerns by introducing interpretable CEFR-based predictors for challenging words and sentences. These specialized models outperform GPT-4o on both classification tasks while providing instance-specific explanations and substantially faster inference.
- 6 Reading.help: Updated Version: Reading.help addresses pilot-study concerns about opaque recommendations and difficulty judging recommendation severity by adopting CEFR as the standard for challenging words and sentences.The revised system was developed to address issues identified in Reading.help-Lite.
- 6.1.1 Lexical Complexity Estimator: The Lexical Complexity Estimator classifies words using frequency, character length, number of senses, and Qwen3-0.6B [72] embeddings, while dedicated expert networks, dynamic gating, feature fusion, and a classifier produce CEFR levels.The four linguistically meaningful features were inspired by prior work [17].
- 6.1.2 Sentence Readability Estimator: The Sentence Readability Estimator applies the same expert-network, alpha-prediction, fusion, and classification design to sentence length, average syllables per word, and Qwen3-based sentence embeddings.Its three features were inspired by prior work [4] [53].
- 6.1.3 Interpretability: Dynamic alpha weights make both predictors interpretable by showing which linguistically meaningful features contributed to each individual CEFR prediction.This provides both feature-level and instance-specific explanations because the gating is data-dependent.
- 6.1.6 Model Evaluation: 65.2% vs. 51.8% average F1 and 64.8% vs. 52.9% overall accuracy show the Lexical Complexity Estimator outperforming GPT-4o on word-level CEFR prediction.Grouped accuracy was also higher at 71.1% vs. 62.3%.
- 6.1.6 Model Evaluation: 75.0% vs. 59.6% average F1 and 87.4% vs. 64.0% overall accuracy show the Sentence Readability Estimator outperforming GPT-4o on sentence-level CEFR prediction.The estimator achieved higher F1 on five of six CEFR levels, while GPT-4o was superior only at C2; grouped accuracy was 89.7% vs. 69.9%.
- 6.1.7 Inference efficiency: 0.05 s vs. 0.611 s for words and 0.052 s vs. 0.592 s for sentences make the proposed models approximately 12× and 11× faster than GPT-4o.Measurements used batch size 1 on a single NVIDIA A100 GPU, with GPT-4o latency measured as end-to-end API roundtrip time.
6.2 Automated Summary
The updated system automatically generates short, paragraph-aligned summaries upon document upload to improve text comprehension. These summaries highlight topic sentences and key claims while remaining anchored to their source paragraphs for coherent skimming and review.
- 6.2 Automated Summary: Reading.help automatically produces short, paragraph-aligned summaries as soon as a document is uploaded.This feature addresses DR2, improving text comprehension (Figure 5A).
- 6.2 Automated Summary: The summaries foreground topic sentences and key claims so EFL readers can quickly grasp each paragraph’s main idea before detailed reading.
- 6.2 Automated Summary: Each summary is anchored to its source paragraph, minimizing context switching and preserving local coherence during skimming and review.
6.3 Validation
Reading.help adds a validation mechanism to improve the reliability of LLM-generated explanations. A second LLM checks each explanation for grounding, assistance-type alignment, and linguistic correctness before visual validation in the interface.
- The validation mechanism was added because LLMs generate grammar explanations despite the CEFR model improving identification-task interpretability.
- A validator LLM checks whether each explanation is grounded in the selected text, matches the requested assistance type, and is linguistically correct and non-contradictory.It returns a binary valid/invalid decision with a brief rationale.
- Validated explanations are visually indicated in the user interface (Figure 6D).
6.4 UI and Usability Improvements
Reading.help’s interface was refined to support different proficiency levels, proactive guidance, and adjustable explanation detail. It also preserves on-demand explanations while making grammar validation and reading-support components more visible.
- UI and Usability Improvements: The interface supports different proficiency levels through recommended CEFR difficulty settings and adjustable summary detail, ranging from concise to detailed.
- UI and Usability Improvements: The system provides vocabulary, comprehension, and grammar components for EFL academic reading, including keyword meanings, contextual main-idea analysis, and grammar support.
- UI and Usability Improvements: Uploaded documents automatically generate paragraph-aligned summaries in a fixed sidebar, while potentially difficult words and sentences receive color-coded grammar, vocabulary, or comprehension highlights.
- UI and Usability Improvements: Grammar explanations were prompted to remain brief, and second-LLM validation is shown with green icons whose hover text reveals the decision rationale.
- UI and Usability Improvements: Reading.help-Lite features remain available, including highlighting text spans to obtain detailed explanations on demand.
6.5 System Architecture and Implementation
Reading.help combines a web-based reading interface with a backend of CEFR models and LLMs. Its pipeline preprocesses documents, performs batched CEFR inference, filters predictions by target level, and renders validated annotations with rationales when available.
- 6.5 System Architecture and Implementation: Reading.help consists of a web-based reading-assistance interface and a backend containing CEFR models and LLMs.The architecture and analysis pipeline are described in Figure 7.
- 6.5 System Architecture and Implementation: LCE estimates word-level complexity and SRE estimates sentence-level readability to assign CEFR levels across uploaded documents.Both interpretable models are implemented in PyTorch and exposed through a FastAPI server.
- 6.5 System Architecture and Implementation: The five-stage workflow segments documents, splits sentences and tokens, runs batched LCE and SRE inference, reconstructs document structure, and filters predictions by target CEFR level.The target level may be user-specified or inferred from history.
- 6.5 System Architecture and Implementation: The interface displays validated annotations using an indicator and, when available, the validator’s rationale.The validator returns a JSON object containing valid and rationale fields, which the decision logic parses.
7 Evaluation
An evaluation with five South Korean EFL readers examined Reading.help’s usage, CEFR-based recommendations, and LLM validation, supplemented by two EFL education experts. Participants most often used vocabulary modules, CEFR gating focused attention on challenging items, and human and LLM validation showed similar behavior despite minor errors.
- Usage patterns: Vocabulary modules were used most often, followed by comprehension and grammar modules, across 101 challenging text segments highlighted by five participants.Participants’ comments indicated that the comprehension module aided understanding of uncertain sentences.
- CEFR-based recommendation: At B2 ≥, recall averaged 87.64% with 18.47% precision and 30.38% F1, while C1 ≥ raised average precision to 37.77% and F1 to 40.82%.The B2 ≥ threshold favored coverage, whereas C1 ≥ provided greater specificity; the peak C1 ≥ F1 was 72.22% on R4.
- CEFR-based recommendation: CEFR recommendations captured rare terms and idioms but missed nonce forms and contextually tricky lower-CEFR words, indicating a need for sense-aware cues.The analysis identified rare or invented terms, idiomatic or figurative expressions, and polysemous common words as three dominant confusion patterns.
- Validation: Human experts rated vocabulary 88% valid, comprehension 93% valid, and grammar 90% valid, while LLM self-validation rated them 93%, 97%, and 96%, respectively.Comprehension showed the highest agreement with human judgments, and the findings indicated similar validation behavior overall.
- Expert feedback: Two EFL education experts reported that Reading.help mitigated text complexity, supported comprehension of details and overall structure, and recommended challenging topics.They also noted that the tool primarily suits highly motivated self-study learners and needs greater accessibility for less motivated users.
8 Discussions
The discussion shows that EFL readers frequently need vocabulary and occasional comprehension, grammar, and summary support, motivating proactive guidance while raising validation concerns. It proposes adaptive, lightweight assistance that preserves productive struggle, supports immersive reading, and augments rather than replaces teachers.
- 8.1 Interpreting the Results: Readers sought vocabulary help 5–6 times on average while reading a 400-word script, plus occasional comprehension, grammar, and paragraph-summary support, showing the effort required for fluent contextual understanding.These usage patterns support proactive guidance for unfamiliar language and comprehension difficulties.
- 8.1 Interpreting the Results: The tool achieved approximately 37.77% precision when recommending a few items, but experts and participants still identified concerns about recommendation validation and response factuality.The system was assessed for adherence to human recommendations and factual correctness; the passage indicates that recommending more items could improve capture, but does not provide the complete result.
- 8.2 Lessons Learned: More adaptive support could personalize explanation language, organization, feedback type, and timing based on accumulated learner skills, accelerating reading and building confidence.Experts envisioned faster proactive guidance for unfamiliar word meanings, grammar, and comprehension.
- 8.2 Lessons Learned: Excessive explanation can increase mental load and cause users to disengage from detailed grammar information, so guidance should be brief and lightweight by default.Participants were overwhelmed by the amount of grammatical text and later did not revisit that section.
- 8.2 Lessons Learned: To prevent over-reliance, reading tools should preserve some productive struggle rather than minimize user effort, because their goal is for tool reliance to decrease as English proficiency grows.One proposed policy is to withhold hints for some terms, though the passage is truncated before specifying the full example.
- 8.3–8.4 Future Directions: Reading assistants should augment teachers by providing adaptive, anytime support while leaving humans responsible for progress, motivation, encouragement, care, and social dimensions of language learning.The discussion also envisions immersive, distraction-free environments with just-in-time clarification and future assessment of interest attraction versus distraction-free immersion.
9 Conclusion
Reading.help is an LLM- and interpretable-neural-network-based tool developed to support EFL readers’ reading processes. Its development followed a pilot study of Reading.help-Lite, and the experiment found that CEFR models can recommend unfamiliar content while LLM abilities remain limited.
- 9 Conclusion: Reading.help was developed with LLMs and interpretable neural networks to support EFL readers’ reading processes.
- 9 Conclusion: A pilot study of Reading.help-Lite informed development of the Reading.help tool, which effectively helps EFL readers.
- 9 Conclusion: CEFR models can effectively recommend parts of texts that EFL readers do not know, while LLMs possess only limited abilities.