Paper deep dive

MuTSE: A Human-in-the-Loop Multi-use Text Simplification Evaluator

Rares-Alexandru Roscan, Gabriel Petre1, Adrian-Marius Dumitran, Angela-Liliana Dumitran

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 33

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%

Last extracted: 4/14/2026, 1:47:55 AM

Summary

MuTSE is a human-in-the-loop web application designed to streamline the evaluation of LLM-generated text simplifications. It features an asynchronous concurrency model for multi-model inference, a tiered semantic alignment engine using a linearity bias heuristic (λ), and a customizable annotation framework for educational and NLP research purposes.

Entities (5)

Large Language Models · technology · 100%MuTSE · software-application · 100%Intelligent Tutoring Systems · domain · 95%Linearity Bias Heuristic · algorithm · 95%Together AI · organization · 90%

Relation Signals (3)

MuTSE → evaluates → Large Language Models

confidence 100% · MuTSE, an interactive human-in-the-loop web application designed to streamline the evaluation of LLM-generated text simplifications

MuTSE → utilizes → Linearity Bias Heuristic

confidence 95% · integrating a novel tiered semantic alignment engine augmented with a linearity bias heuristic (λ)

MuTSE → supports → Intelligent Tutoring Systems

confidence 90% · MuTSE extends this tradition by enabling educators to actively generate and comparatively evaluate simplified texts

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:As Large Language Models (LLMs) become increasingly prevalent in text simplification, systematically evaluating their outputs across diverse prompting strategies and architectures remains a critical methodological challenge in both NLP research and Intelligent Tutoring Systems (ITS). Developing robust prompts is often hindered by the absence of structured, visual frameworks for comparative text analysis. While researchers typically rely on static computational scripts, educators are constrained to standard conversational interfaces -- neither paradigm supports systematic multi-dimensional evaluation of prompt-model permutations. To address these limitations, we introduce \textbf{MuTSE}\footnote{The project code and the demo have been made available for peer review at the following anonymized URL. this https URL, an interactive human-in-the-loop web application designed to streamline the evaluation of LLM-generated text simplifications across arbitrary CEFR proficiency targets. The system supports concurrent execution of $P \times M$ prompt-model permutations, generating a comprehensive comparison matrix in real-time. By integrating a novel tiered semantic alignment engine augmented with a linearity bias heuristic ($\lambda$), MuTSE visually maps source sentences to their simplified counterparts, reducing the cognitive load associated with qualitative analysis and enabling reproducible, structured annotation for downstream NLP dataset construction.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

33,054 characters extracted from source content.

Expand or collapse full text

MuTSE: A Human-in-the-Loop Multi-use Text Simplification Evaluator Rares , -Alexandru Ros , can 1 and Gabriel Petre 1 and Adrian-Marius Dumitran 1,3,4[0009−0005−3547−5772] and Angela-Liliana Dumitran 2[0009−0003−3590−9441] 1 University of Bucharest, Faculty of Mathematics and Computer Science, Academiei 14, 010014, Bucharest, Romania marius.dumitran@unibuc.ro, radu.dita@gmail.com 2 Universitatea Cres , tină "Dimitrie Cantemir" angela.dumitran@gmail.com 3 Cu Drag si Sport SRL, Bucharest, Romania 4 Softbinator Technologies, Bucharest, Romania Abstract. As Large Language Models (LLMs) become increasingly prevalent in text simplification, systematically evaluating their outputs across diverse prompting strategies and architectures remains a critical methodological challenge in both NLP research and Intelligent Tutor- ing Systems (ITS). Developing robust prompts is often hindered by the absence of structured, visual frameworks for comparative text analysis. While researchers typically rely on static computational scripts, educators are constrained to standard conversational interfaces — neither paradigm supports systematic multi-dimensional evaluation of prompt-model per- mutations. To address these limitations, we introduce MuTSE 5 , an interactive human-in-the-loop web application designed to streamline the evaluation of LLM-generated text simplifications across arbitrary CEFR proficiency targets. The system supports concurrent execution ofP× M prompt-model permutations, generating a comprehensive comparison matrix in real-time. By integrating a novel tiered semantic alignment engine augmented with a linearity bias heuristic (λ), MuTSE visually maps source sentences to their simplified counterparts, reducing the cog- nitive load associated with qualitative analysis and enabling reproducible, structured annotation for downstream NLP dataset construction. Keywords: Text Simplification· Large Language Models· Intelligent Tutoring Systems· Human-in-the-Loop· Semantic Alignment. 1 Introduction The adaptation of complex texts into accessible reading materials—known as text simplification—is a core component of Intelligent Tutoring Systems (ITS) 5 The project code and the demo have been made available for peer review at the following anonymized URL.https://osf.io/njs43/overview?view_only= 4b4655789f484110a942ebb7788cdf2a arXiv:2604.08947v1 [cs.CL] 10 Apr 2026 2Rares , -Alexandru Ros , can and language learning applications. The rapid advancement of Large Language Models (LLMs) has provided educators and researchers with robust generative frameworks capable of tailoring texts to specific educational levels (e.g., CEFR A2, B1). Nevertheless, optimizing the interplay between LLM architectures and prompting strategies remains a significant methodological challenge. Currently, evaluating LLM outputs for text simplification is a fragmented and labor-intensive undertaking. Researchers and applied linguists frequently depend on standalone scripts or computational notebooks for text generation, approaches that lack scalability when systematically comparing multiple models against diverse prompts. This configuration yields a high-dimensional evaluation matrix ofP× M(wherePdenotes the number of prompts andMthe number of models). Furthermore, although existing methodologies employ conventional automated metrics (e.g., BLEU, SARI) for output evaluation, they typically lack the interactive, human-in-the-loop interfaces required for nuanced qualitative assessment. Consequently, tracking how a specific complex sentence is structurally altered across numerous generated variants imposes a substantial cognitive load on researchers. To address these methodological constraints, we introduce MuTSE, an inter- active, human-in-the-loop framework developed for the systematic evaluation of LLM-generated text simplifications. The system mitigates the operational complexity associated with model inference and prompt orchestration, enabling users to conduct parallel generation tasks across arbitrary permutations of local or cloud-based LLMs. The primary contributions of this paper are: –Parallel Comparative Workflow: A system architecture capable of ex- ecuting multi-dimensional (P× M) text simplification tasks concurrently, presenting the results in a unified, side-by-side visualization. – Interactive Semantic Alignment: A novel, tiered alignment engine fea- turing a linearity bias heuristic (λ) that visually maps original sentences to their simplified counterparts across all selected models in real-time. –Educational Annotation Framework: An integrated suite of readability metrics (e.g., Flesch-Kincaid) and a customizable manual annotation system, enabling educators and linguists to score texts based on personalized criteria and export structured data for further analysis. By bridging the gap between raw LLM capabilities and practical qualitative assessment, MuTSE provides an accessible environment for educators to select optimal texts and for NLP researchers to build high-quality annotated datasets. 2 Related Work 2.1 Text Simplification Tools and Evaluation Interfaces While automated metrics drive large-scale benchmarking, the qualitative analysis of simplified text requires specialized software interfaces. At the programmatic MuTSE: A Human-in-the-Loop Multi-use Text Simplification Evaluator3 level, toolkits like EASSE [2] have successfully standardized the extraction of text simplification metrics (e.g., SARI, compression ratio) and the generation of automated evaluation reports. However, such packages remain strictly developer- centric, operating via command-line or code interfaces without visual alignment capabilities. On the visual interface front, contributions such as TS-ANNO [15] have introduced dedicated human-in-the-loop web environments for manually aligning, rating, and annotating text simplifications. While highly effective for post-hoc corpus creation, tools like TS-ANNO are primarily designed for manual annota- tion rather than live, concurrent multi-model generation. Furthermore, a recent comprehensive review of automatic text simplification tools [5] highlights a signifi- cant gap in the current ecosystem: a pervasive lack of publicly accessible resources that offer high customization and empower users to control the simplification process without technical barriers. MuTSE directly addresses these limitations. By combining the concurrent generation capabilities of modern LLMs with the automated metric extraction of programmatic toolkits and the visual semantic mapping of annotation platforms, MuTSE provides a unified, highly customizable, and accessible environment for rigorous text simplification research. Within Intelligent Tutoring Systems, the problem of matching text complexity to learner proficiency has a long history. Systems such as REAP [4] selected reading materials calibrated to individual vocabulary gaps, while Project LIS- TEN [12] demonstrated the pedagogical value of automated text-level assessment in supporting struggling readers. MuTSE extends this tradition by enabling educators to actively generate and comparatively evaluate simplified texts across proficiency targets such as CEFR A2 and B1, rather than merely selecting from pre-existing corpora. 2.2 LLMs and the Evolution of Simplification Benchmarks The paradigm of automated text simplification has shifted dramatically toward the deployment of Large Language Models (LLMs). Recent studies [6] demonstrate that LLMs possess strong zero-shot capabilities for adapting complex texts. To systematically evaluate these capabilities, comprehensive benchmarks such as BLESS [10] have been introduced, assessing dozens of off-the-shelf LLMs across multiple domains. While such benchmarks confirm that LLMs perform on par with state-of-the-art baselines, they also highlight the necessity of manual, qualitative analysis to truly gauge the diversity of edit operations and the educational quality of the generated text. To evaluate these generative models computationally, researchers often rely on sophisticated, embedding-based automated metrics like BERTScore [17], which compute semantic similarity using contextual embeddings rather than exact lexical overlap. However, while metrics like BERTScore are excellent for static benchmarking, they aggregate performance into single numerical values. They lack the granular interpretability required by educators to visually map and validate multi-reference transformations. MuTSE bridges this methodological 4Rares , -Alexandru Ros , can gap; it complements large-scale benchmarks by providing a real-time, human-in- the-loop environment where researchers can qualitatively analyze LLM outputs through visual semantic alignment, moving beyond opaque aggregate scores. 3 System Architecture and Methodology To facilitate the near real-time evaluation of text simplification, MuTSE is built upon a decoupled, asynchronous client-server architecture. The backend is powered by Python and FastAPI, providing a high-performance REST API that orchestrates parallel LLM generation tasks via Together AI’s serverless endpoints. On the frontend, a Vue.js 3 interface leverages reactive state management to handle multi-dimensional comparisons and client-side alignments. This infrastructure empowers researchers to systematically benchmark multi- ple LLMs and prompting strategies via an interactive, human-in-the-loop visual- ization framework. By automatically identifying and highlighting semantically corresponding sentences—utilizing a core NLP pipeline based on spaCy, sentence- transformers, and scikit-learn—the platform alleviates the challenges of tracking complex transformations. Furthermore, a lightweight, local JSON-based persis- tence layer ensures maximum portability for educators without requiring complex database configurations. The following subsections detail the concurrency frame- work designed to computeP× Mevaluation matrices efficiently, along with the alignment engine driving the comparative interface. 3.1 Asynchronous Concurrency Model for Multi-Model Inference A fundamental constraint in multi-model evaluation is the latency overhead inherent to sequential network requests. To ensure the responsiveness required for an uninterrupted human-in-the-loop workflow, the architecture implements a highly concurrent execution paradigm. Upon request initiation, the orchestration layer provisions independent, dis- tributed tasks for each prompt and model permutation. These processes are evaluated asynchronously, effectively restraining the overall computational la- tency to approximatelyO(max(t i )), wheret i denotes the inference time of the slowest individual model iteration. However, an inherent characteristic of this con- current approach is that the system’s absolute responsiveness is strictly bounded by the highest latency within the queried batch. For instance, evaluating a highly optimized, low-latency model (e.g., Llama 3.3 Turbo) alongside a computationally heavy "reasoning" model will inevitably throttle the final matrix resolution time to match the latter’s prolonged inference speed. To maintain concurrency safety, the architecture strictly isolates state during the parallel execution phase. Each task independently manages text generation, semantic alignment computations, and metric extraction. Computationally inten- sive procedures, such as similarity matrix derivations, are strategically delegated to prevent blocking primary execution threads. The subsequent aggregation of the generated corpora and metadata is performed strictly after all concurrent MuTSE: A Human-in-the-Loop Multi-use Text Simplification Evaluator5 tasks have reached completion, ensuring structural integrity without distributed state conflicts. asyncio.gather (Concurrent Execution) Input: Source Text + P Prompts + M Models Orchestrator: Generate Coroutines LLM API Call Alignment & Metrics LLM API Call Alignment & Metrics . . . LLM API Call Alignment & Metrics Assemble & Output JSON Task 1 Task 2 Task P× M Fig. 1. Asynchronous parallel orchestration (P× M) pattern demonstrating the con- current execution of prompts across multiple LLMs. 3.2 Semantic Alignment Engine A central contribution of the system is the Semantic Alignment Engine, designed to visually map original sentences to their simplified counterparts. While text simplification predominantly exhibits a monotonic structural progression [1,16], standard cosine similarity measures applied to sentence embeddings frequently yield false positive alignments—often associating a simplified construct with a semantically congruent but positionally distant original sentence. To mitigate this alignment degradation, we propose a hierarchical fallback strategy augmented by a positional penalty heuristic. Hierarchical Embedding Strategy To guarantee operational resilience and hardware agnosticism, sentence embeddings are derived through a multi-level sequential cascade: 1.Primary Semantic Level: Utilizes a condensed, 384-dimensional multilin- gual transformer-based model (paraphrase-multilingual-MiniLM-L12-v2) [13] to compute the foundational cosine similarity matrix. 2.Secondary Lexical Level: In scenarios where semantic embedding fails or yields null vectors, the algorithm reverts to a hybrid TF-IDF representation [14]. This incorporates both word-level and character-level n-grams to capture foundational cross-lingual morphological similarities. 3.Tertiary Positional Level: A definitive structural fallback that bypasses lexical similarity entirely, establishing alignments based strictly on normalized sequence positioning within the text documents. 6Rares , -Alexandru Ros , can Linearity Bias Assignment To align sentences accurately while avoiding the false positives inherent in purely semantic matching, we introduce a linearity factor (λ). The alignment score between an original sentenceiand a simplified sentence j is calculated as: Score(i,j) = CosineSim(S (i) orig ,S (j) simp )− Pos (i) rel − Pos (j) rel × λ(1) Where: – S (i) orig andS (j) simp represent the original and simplified sentences, respectively. – Pos (i) rel = i max(N orig −1,1) andPos (j) rel = j max(N simp −1,1) represent the normal- ized relative positions within their respective texts. – λis the linearity factor, adjustable in real-time by the user between 0 and 2, with a default of 0.5. The algorithm uses a many-to-one assignment, mapping each simplified sen- tence to the original sentence that yields the highest Score(i,j). Computational Advantages of the Linearity Heuristic The integration of the linearity bias functions as a structural regularizer, effectively offsetting the reduced discriminative capacity inherent to heavily compressed embedding models. Whereas unpenalized semantic alignment (λ= 0) often mandates large-parameter models and dedicated GPU acceleration for optimal precision, our heuristic enables lower-dimensional embeddings to yield highly robust alignments using standard CPU architectures. This approach significantly lowers computational prerequisites, facilitating local deployment for educators and researchers without requiring specialized infrastructure. Moreover, by offloading the final alignment graph computation to the client interface using the pre-calculated similarity matrix, modifications to theλparameter are resolved instantaneously on the frontend without invoking redundant server-side processing. 4 Educational Features and Interactive Workflow The primary objective of MuTSE is to facilitate the qualitative evaluation of gen- erated texts for educational purposes. To bridge the gap between algorithmic text generation and pedagogical utility, the system incorporates an interactive visual- ization interface, automated readability assessment models, and a customizable manual annotation framework. 4.1 System Modularity: Prompts, Models, and Custom Criteria A core design principle of MuTSE is high modularity, accessible directly through the user interface’s Settings module. This architecture permits researchers to dynamically configure the theoretical bounds of their evaluation matrix without altering the underlying application code: MuTSE: A Human-in-the-Loop Multi-use Text Simplification Evaluator7 Fig. 2. The impact of the Linearity Bias (λ) on the semantic alignment visualization. Top: Pure semantic matching (λ= 0.00) showing scattered, positionally distant alignments. Bottom: Strict positional penalty applied (λ= 2.00), resolving false positives and enforcing monotonic alignment. 8Rares , -Alexandru Ros , can Input: Original & Simplified Sentences Tier 1: SBERT Valid vectors? Tier 2: TF-IDF Valid vectors? Tier 3: Positional Zero Matrix Base Cosine Matrix Apply Linearity Bias Penalty (λ) Final Alignment Graph (Many-to-one) No No Yes Yes Fig. 3. The 3-tier semantic alignment cascade showing hierarchical fallback logic and the application of Linearity Bias (λ). – Prompt Management: Users can define, persist, and retrieve custom in- structional prompts. To facilitate immediate experimentation, the platform is pre-loaded with a curated set of 5 pre-defined prompts. While not neces- sarily state-of-the-art implementations, these foundational prompts serve as practical starting points for standard language proficiency objectives (e.g., A2, B1), encouraging standardized and reproducible research. –Model Integration: The framework leverages cloud-based serverless in- frastructure (via Together AI) to provide zero-provisioning access to diverse open-weight LLMs (e.g., Llama 3, DeepSeek V3, Qwen 2.5). Expanding the evaluation matrix is highly intuitive: users can dynamically add new models directly through the interface by simply inputting the model’s official string identifier (e.g.,openai/gpt-oss-20b) exactly as listed on the provider’s pub- lic registry. This approach eliminates complex local hardware configurations while allowing users to instantly evaluate newly released models without altering the underlying codebase. –Customizable Annotation Framework: To meaningfully capture the qualitative nuances of educational simplification, MuTSE features a highly configurable inline scoring system. Human evaluation in Natural Language Generation traditionally assesses dimensions such as Fluency and Mean- ing Preservation. However, recent literature highlights a pervasive lack of standardization and extreme diversity in evaluation methodologies and ter- minology across the field [9]. Furthermore, an ongoing methodological debate exists regarding rating scales: while discrete Likert scales (e.g., 1-5 points) are widely used, their application frequently deviates from best practices [3], prompting researchers to advocate for continuous measurement scales to capture finer quality distinctions [8]. Rather than enforcing a rigid evalua- MuTSE: A Human-in-the-Loop Multi-use Text Simplification Evaluator9 tion paradigm, MuTSE operates as an evaluation-agnostic framework with zero predefined metrics. Through the Settings module, evaluators build their custom assessment criteria entirely from scratch. For each dimension, users establish their own custom numerical rating scale—supporting any arbi- trary range from strict binary evaluations (e.g., a 1-to-2 scale representing True/False) up to continuous 100-point sliders—and assign a relative impact weight (from 0.1 to 10.0). During analysis, the system normalizes and ag- gregates these manual inputs to compute a weighted overall performance percentage for each generated column. 4.2 Interactive Simplification and Visual Alignment The primary analytical interface of MuTSE is designed to mitigate the substan- tial cognitive load involved in evaluating multiple simplified derivations of a source text. Conventionally, tracking structural modifications, fragmentations, or semantic omissions across M generated variants is a visually taxing process. MuTSE addresses this through a full-screen, side-by-side comparative layout, which directly visualizes the results of the multi-model generation process. To further manage visual complexity within high-dimensional comparisons, the interface features a dynamic filtering header. Evaluators can selectively toggle specific prompting strategies or LLM architectures on and off, instantly hiding or revealing their corresponding text columns. This functionality empowers users to compute a massiveP× Mevaluation matrix in a single execution, yet systematically analyze the outputs in smaller, focused subsets (e.g., isolating a single prompt across two specific models) according to their preferred analytical workflow without re-triggering generation tasks. The interface then augments this focused view with an interactive cross- document highlighting mechanism governed by the pre-computed semantic align- ment matrix (detailed in Section 3.2). Upon user interaction (hover or click) with a specific sentence in any document (source or generated), the system synchronously highlights the corresponding semantically aligned sentences across all active text columns. This alignment traversal operates natively within the client environment, ensuring near-instantaneous visual feedback and allowing evaluators to rapidly trace specific semantic units across disparate architectural outputs. Crucially, the interface augments this qualitative visual tracking with rigorous, quantitative linguistic metrics and inline alignment diagnostics displayed directly alongside each variant column. Specifically, the system features a dedicated Alignment visual toggle that embeds the exact cosine-similarity alignment score inline with each generated sentence. This granular visualization allows evalua- tors to immediately identify sentences with low correspondence to the original text, significantly accelerating the manual detection of unwanted LLM conver- sational artifacts (e.g., "Here is the simplified text:") or unprompted semantic hallucinations. Furthermore, rather than relying exclusively on subjective visual inspection, researchers can continuously cross-reference human annotations with the system’s 10Rares , -Alexandru Ros , can calculated statistics. MuTSE computes and visually badges real-time textual diagnostics including: –Textual Statistics: Word frequency, total sentence count, and average sentence length. –Compression Ratio: The proportion of retained lexical content relative to the source text, indicating structural expansion or reduction. –Flesch-Kincaid Grade Level: An estimation of the US educational grade level required for comprehension [11]. – Flesch Reading Ease (FRE): A quantitative assessment of overall reading accessibility on a 0-100 scale [7]. By unifying these immediate, automated benchmarks—calculating metrics like syllable densities dynamically via language-specific hyphenation libraries—with the visual semantic mapping and inline custom scoring panels, MuTSE signifi- cantly reduces the evaluation time overhead while enhancing the consistency of multidimensional comparative analysis. Fig. 4. The integrated statistics module computing real-time educational metrics (e.g., Flesch-Kincaid, Reading Ease) and structural diagnostics across multiple prompt-model permutations. 4.3 Structured Data Export for NLP Research To support the creation of high-quality, human-in-the-loop datasets for future Natural Language Processing (NLP) research, all evaluation sessions are persis- tently stored locally. Crucially, researchers can export the entirety of an evaluation session—encompassing the source texts, generated outputs, semantic alignment links, readability metrics, and the weighted manual annotations—in standardized, structured formats (JSON and CSV). This export functionality is essential for downstream applications. The struc- tured CSV data is optimized for rapid Exploratory Data Analysis (EDA) and statistical significance testing, while the nested JSON format preserves the com- plex relational mapping of sentences and annotations, making it directly suitable for fine-tuning subsequent generations of language models or training specialized reward models. MuTSE: A Human-in-the-Loop Multi-use Text Simplification Evaluator11 Backend JSON Response Render Side-by-Side Comparison Path B: Adjust Linearity Bias (λ) Client-side Graph Recomputation Update Visual Sentence Highlights Path A: Toggle Prompts/Models Instantly Hide/Reveal Columns Path C: Input Manual Scores Calculate Weighted Final Percentage Export (JSON/CSV Dataset) Fig. 5. The Vue.js frontend workflow demonstrating client-side interactivity, zero-latency recomputations, and parallel user interaction paths. 12Rares , -Alexandru Ros , can 5 Limitations and Future Work While MuTSE significantly streamlines the evaluation of text simplification, the current architecture presents certain limitations. First, the reliance on a local JSON-based persistence layer, while maximizing portability for individual researchers, does not scale effectively for concurrent, multi-user deployments in large laboratory settings. Transitioning to a robust relational database is required for collaborative annotation campaigns. Second, while the integration of cloud-based LLM providers (e.g., Together AI) eliminates the need for local GPU clusters, users still face initial deployment friction regarding environment configuration (e.g., Python, Node.js dependencies). For future work, the modularity of the semantic alignment engine presents significant opportunities beyond monolingual text simplification. The primary tier of our alignment cascade already utilizes a robust multilingual embedding model (paraphrase-multilingual-MiniLM-L12-v2) capable of processing over 50 languages, while the secondary TF-IDF tier captures cross-lingual morpholog- ical roots via character n-grams. Consequently, a natural extension of MuTSE is adapting the framework for Machine Translation (MT) evaluation and cross- lingual summarization. While the current implementation is highly optimized for simplification, verifying MT outputs would require recalibrating the linearity bias heuristic (λ), as cross-lingual syntactic restructuring may not strictly adhere to the monotonic sentence order typically observed in monolingual simplification. Expanding the interface to support these diverse Natural Language Generation evaluation branches remains a primary objective. 6 Conclusion The systematic evaluation of Large Language Models in educational contexts requires tools that bridge the gap between automated generation and qualitative human assessment. In this paper, we introduced MuTSE, a human-in-the-loop web application designed to accelerate the comparative analysis of text sim- plification. By coupling an asynchronous multi-model generation pipeline with a novel, tiered semantic alignment engine, the system mitigates the cognitive load associated with tracking complex textual transformations. Furthermore, the integration of customizable annotation scales and automated readability metrics provides educators and applied linguists with an accessible, highly configurable environment. Ultimately, MuTSE not only streamlines the selection of optimal texts for language learners but also serves as a robust foundation for building high-quality annotated corpora for future NLP research. References 1.Alva-Manchego, F., Martin, L., Bordes, A., Scarton, C., Sagot, B., Specia, L.: ASSET: A dataset for tuning and evaluation of sentence simplification models with multiple rewriting transformations. In: Jurafsky, D., Chai, J., Schluter, N., MuTSE: A Human-in-the-Loop Multi-use Text Simplification Evaluator13 Tetreault, J. (eds.) Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. p. 4668–4679. Association for Computational Lin- guistics, Online (Jul 2020).https://doi.org/10.18653/v1/2020.acl-main.424, https://aclanthology.org/2020.acl-main.424/ 2. Alva-Manchego, F., Martin, L., Scarton, C., Specia, L.: EASSE: Easier automatic sentence simplification evaluation. In: Padó, S., Huang, R. (eds.) Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP): System Demonstrations. p. 49–54. Association for Computational Linguistics, Hong Kong, China (Nov 2019).https://doi.org/10.18653/v1/D19-3009,https: //aclanthology.org/D19-3009/ 3. Amidei, J., Piwek, P., Willis, A.: The use of rating and Likert scales in natu- ral language generation human evaluation tasks: A review and some recommen- dations. In: van Deemter, K., Lin, C., Takamura, H. (eds.) Proceedings of the 12th International Conference on Natural Language Generation. p. 397–402. As- sociation for Computational Linguistics, Tokyo, Japan (Oct–Nov 2019).https: //doi.org/10.18653/v1/W19-8648, https://aclanthology.org/W19-8648/ 4.Brown, J., Eskénazi, M.: Retrieval of authentic documents for reader-specific lexical practice. In: Proc. InSTIL/ICALL 2004 Symposium on Computer Assisted Learning, paper 006 (2004), https://api.semanticscholar.org/CorpusID:6480264 5.Espinosa-Zaragoza, I., Abreu-Salas, J., Lloret, E., Moreda, P., Palomar, M.: A review of research-based automatic text simplification tools. In: Mitkov, R., Angelova, G. (eds.) Proceedings of the 14th International Conference on Recent Advances in Natural Language Processing. p. 321–330. INCOMA Ltd., Shoumen, Bulgaria, Varna, Bulgaria (Sep 2023), https://aclanthology.org/2023.ranlp-1.36/ 6. Feng, Y., Qiang, J., Li, Y., Yuan, Y., Zhu, Y.: Sentence simplification via large language models (2023), https://arxiv.org/abs/2302.11957 7. Flesch, R.: A new readability yardstick. Journal of Applied Psychology 32(3), 221–233 (1948). https://doi.org/10.1037/h0057532 8. Graham, Y., Baldwin, T., Moffat, A., Zobel, J.: Continuous measurement scales in human evaluation of machine translation. In: Pareja-Lora, A., Liakata, M., Dipper, S. (eds.) Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse. p. 33–41. Association for Computational Linguistics, Sofia, Bulgaria (Aug 2013), https://aclanthology.org/W13-2305/ 9. Howcroft, D.M., Belz, A., Clinciu, M.A., Gkatzia, D., Hasan, S.A., Mahamood, S., Mille, S., van Miltenburg, E., Santhanam, S., Rieser, V.: Twenty years of confusion in human evaluation: NLG needs evaluation sheets and standardised definitions. In: Davis, B., Graham, Y., Kelleher, J., Sripada, Y. (eds.) Proceedings of the 13th International Conference on Natural Language Generation. p. 169–182. Association for Computational Linguistics, Dublin, Ireland (Dec 2020).https://doi.org/10. 18653/v1/2020.inlg-1.23, https://aclanthology.org/2020.inlg-1.23/ 10. Kew, T., Chi, A., Vásquez-Rodríguez, L., Agrawal, S., Aumiller, D., Alva- Manchego, F., Shardlow, M.: BLESS: Benchmarking large language models on sentence simplification. In: Bouamor, H., Pino, J., Bali, K. (eds.) Proceed- ings of the 2023 Conference on Empirical Methods in Natural Language Pro- cessing. p. 13291–13309. Association for Computational Linguistics, Singa- pore (Dec 2023).https://doi.org/10.18653/v1/2023.emnlp-main.821,https: //aclanthology.org/2023.emnlp-main.821/ 11. Kincaid, P., Fishburne, R.P., Rogers, R.L., Chissom, B.S.: Derivation of new readability formulas (automated readability index, fog count and flesch reading 14Rares , -Alexandru Ros , can ease formula) for navy enlisted personnel (1975),https://api.semanticscholar. org/CorpusID:61131325 12. Mostow, J., Aist, G.: Evaluating tutors that listen: An overview of project LISTEN. In: Intelligent Tutoring Systems in e-Learning Environments: Design, Implementa- tion and Evaluation. Information Science Publishing (2001) 13.Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks (2019), https://arxiv.org/abs/1908.10084 14.Salton, G., Buckley, C.: Term-weighting approaches in automatic text re- trieval. Information Processing and Management 24(5), 513–523 (1988). https://doi.org/https://doi.org/10.1016/0306-4573(88)90021-0,https:// w.sciencedirect.com/science/article/pii/0306457388900210 15. Stodden, R., Kallmeyer, L.: TS-ANNO: An annotation tool to build, annotate and evaluate text simplification corpora. In: Basile, V., Kozareva, Z., Stajner, S. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computa- tional Linguistics: System Demonstrations. p. 145–155. Association for Compu- tational Linguistics, Dublin, Ireland (May 2022).https://doi.org/10.18653/v1/ 2022.acl-demo.14, https://aclanthology.org/2022.acl-demo.14/ 16.Xu, W., Napoles, C., Pavlick, E., Chen, Q., Callison-Burch, C.: Optimizing statistical machine translation for text simplification. Transactions of the Association for Computational Linguistics 4, 401–415 (2016).https://doi.org/10.1162/tacl_a_ 00107, https://aclanthology.org/Q16-1029/ 17.Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert (2020), https://arxiv.org/abs/1904.09675

Instant research discovery