Reading.help: Supporting EFL Readers with Proactive and On-Demand Explanation of English Grammar and Semantics

TL;DR

Reading.help tackles the challenge of English reading for EFL learners by integrating proactive, on-demand explanations with LLM-driven support. The work combines a lite prototype and an updated, CEFR-guided system that provides paragraph-aligned summaries and interpretable predictors for lexical and sentence complexity, validated by human experts and a dual-LLM validation scheme. Across pilot and follow-up studies in South Korea, Reading.help demonstrates potential to enable self-guided English learning, especially for vocabulary and comprehension, while highlighting challenges in trust, customization, and grammar explanations. The approach offers practical impact for scalable, learner-centered reading assistance, with implications for adaptive feedback, trust signals, and augmentation of teaching rather than replacement of educators.

Abstract

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.

Figures (7)

• Figure 1: Reading.help-lite interface. (A) A control panel that appears upon selection of a text. The user can choose to get help, or manually look for what they want by pressing the button 'tools.' (B) On selecting 'help', Reading.help-Lite displays potential challenging issues that the user might be interested in. A user can select any of the issues to see it in a detailed view (C, D, E). By toggling the 'subject-specific vocabulary,' (F) the user can view subject-specific keywords underlined in the text.
• Figure 2: Reading.help-Lite usage during pilot study. (A) shows the average number of modules (vocabulary, comprehension, and grammar) used by participants while conducting a unit task. (B) compares the number of times 'help' and manual manipulation, or 'tools' are accessed while conducting a unit task.
• Figure 3: Overview of our interpretable CEFR prediction models. Left: Lexical Complexity Estimator (word-level). Right: Sentence Readability Estimator (sentence-level). Each feature is processed by a small expert network $f_i$; a data-dependent router predicts weights $\alpha$ that gate the experts and form a weighted sum, which a classifier maps to a CEFR level (A1-C2). Word features: frequency, character length, number of senses, and a contextual embedding from a LoRA-adapted Qwen3 0.6B qwen3. Sentence features: number of words, average syllables per word, and a contextual embedding from the same model. The gray callouts show the prompts used to obtain contextual embeddings. The learned $\alpha$ weights provide instance-specific, explanations of which features drove each prediction. The word- and sentence-level CEFR results are combined, then filtered to include only items at or above the user-selected CEFR level, and presented as the final recommendations.
• Figure 4: User Actions and System Pipeline for Summarization and Proactive Recommendations. Users first upload a document (①); then choose a summary style, Concise or Detailed (②); set a target CEFR level (e.g., B2 or C1) (③); and run the CEFR analysis (④). In the system, the backend endpoint POST /summarize generates paragraph-aligned summaries (A), which are displayed in a fixed left sidebar at the selected detail level (B). Interpretable CEFR predictors, the Sentence Readability Estimator (SRE) and the Lexical Complexity Estimator (LCE), perform batched inference and serve results (C). The passage view is annotated with CEFR difficulty labels at the token and sentence levels (D), and proactive recommendations surface words and sentences that exceed the user’s threshold (E, F). A help panel lists recommended items that are likely to be difficult for the user (G). When the user requests a CEFR decision, an interpretable feature analysis presents level probabilities and compact feature-contribution bars (H).
• Figure 5: Detailed explanation of components to help EFL academic reading. We provide three components to help EFL researchers in academic reading. They are: (A) vocabulary (lexical terms), (B) comprehension, and (C) grammar. In (A), we present the definition of the selected keyword, or the idiom, usage, as well as its meaning in Korean. In (B), we help users understand the text by analyzing the main idea of the user-selected text with respect to the entire context. In condition (C), we provide grammatical explanations about the user-selected text at the phrase level.