Simplifications are Absolutists: How Simplified Language Reduces Word Sense Awareness in LLM-Generated Definitions

Georg Groh; Lukas Ellinger; Miriam Ansch\"utz

REVIEW 2 major objections 5 minor 15 references

Simplifications are Absolutists: How Simplified Language Reduces Word Sense Awareness in LLM-Generated Definitions

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Simplified definition requests make large language models routinely drop all but one sense of ambiguous words; explicit prompting or small-scale preference fine-tuning mostly restores the lost senses.

desk verdict A solid, useful empirical study showing simplification strips homonym senses in English; the multilingual claim needs judge validation before it carries weight. read the letter →

arxiv 2507.11981 v1 pith:E662FPQX submitted 2025-07-16 cs.CL

Lukas Ellinger , Miriam Anschütz , Georg Groh This is my paper · ORCID

classification cs.CL

keywords homonymspolysemytextsimplificationdefinitiongenerationwordsenseawarenessLLM-as-a-judgedirectpreferenceoptimizationmultilingualevaluation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that when users ask large language models for simplified or child-friendly definitions, the models systematically stop acknowledging that a word has multiple meanings. Across five LLMs and five languages, 'simple language' and 'explain like I am 5' prompts cut homonym definition completeness from 68–84% (normal prompts) to as low as 1.3% (ELI5), because the model lists a single sense and no longer signals that other senses exist. To measure this, the authors introduce Helpful Sense Awareness (HeSA): a response counts as complete if it either lists all senses or explicitly tells the user that more meanings exist or asks for context. They further show that appending a 'words can have multiple meanings' reminder to the prompt, or fine-tuning Llama 3.1 8B with direct preference optimization on 116 preference pairs, largely closes the gap. If the paper is right, the default simplification behavior of LLMs misleads exactly the audience simplification is meant to serve, namely children and language learners.

What carries the argument

The load-bearing object is Helpful Sense Awareness (HeSA): a definition of a homonym is judged complete if it either enumerates all the word's senses or explicitly signals that it is not exhaustive, by remarking that additional meanings exist or by asking the user for the intended context. A relaxed variant, Sense Awareness (multiple definitions listed or HeSA), is used where no dictionary data exists. The measurement pipeline around this notion has three moving parts: an LLM-as-a-Judge (GPT-4o mini) that extracts listed definitions and flags HeSA signals, a sentence-embedding matcher that maps extracted definitions onto WordNet coarse-grained sense clusters to score coverage, and a DPO fine-tuning step that trains Llama 3.1 8B to prefer complete over incomplete responses. HeSA is what carries the argument, because it is precisely the acknowledgment of ambiguity—'other meanings exist' or 'tell me the context'—that simplification removes.

What would settle it

Take a random sample of non-English ML-WiC responses (for example, 200 Russian and 200 Chinese responses from Qwen3-30B A3B and DeepSeek v3 under Simple and ELI5 prompts) and have human annotators apply the same categorization rubric. If human agreement with the judge falls well below the English validation (93.33% category accuracy, kappa 0.86), or if human raters find that simplified responses list multiple senses or signal incompleteness as often as Normal responses do, the cross-lingual version of the central claim would be an artifact of the evaluator rather than a property of the models.

Watch

Extended reading notes

Core claim

The paper's central claim is that stylistic constraints aimed at simplification drastically degrade homonym definition quality in current LLMs by erasing word sense awareness. On the HoWN dataset of 164 English homonyms, completeness under the Normal prompt ranges from 68.63% (DeepSeek v3) to 83.66% (Qwen3-30B A3B); under Simple it falls to 22.22–37.25%; and under ELI5 it collapses to 1.31–15.69%. A response counts as complete only if it covers all senses or exhibits HeSA, and both the number of listed senses and the rate of HeSA fall from Normal to Simple to ELI5, while Flesch–Kincaid scores confirm that the models are genuinely following the simplicity instruction. The same ordering appears in Sense Awareness on the multilingual ML-WiC dataset across Arabic, Chinese, English, French, and Russian. Two correctives are shown to work: Multi-Sense-Aware prompting ('keep in mind that some words have more than one meaning') raises completeness and sense coverage in every prompt type, and DPO fine-tuning of Llama 3.1 8B on just 116 preference pairs improves completeness on unseen words by 11.22 to 25.51 percentage points, letting the 8B model outperform all five larger baselines. The authors conclude that LLMs are capable of complete homonym definitions but are steered away from them by expected output style.

Load-bearing premise

The automated judge used to score all responses correctly counts how many senses a definition gives and whether it signals incompleteness, in every language it was applied to, even though it was validated only on 450 English responses.

Editorial extensions

If this is right

Asking any of the five tested LLMs for a simple or child-friendly definition of an ambiguous word will typically yield exactly one sense with no warning that other senses exist; under the ELI5 prompt, completeness falls to 1.31–15.69%.
Adding one sentence to the prompt ('keep in mind that some words have more than one meaning') substantially recovers completeness and sense coverage in every prompt type, with the largest gains under Simple and ELI5.
A small 8B model fine-tuned with DPO on 116 preference pairs—all from Simple prompts—outperforms all five base models on homonym completeness, including on words it never saw during training and across all five tested languages.
The Normal > Simple > ELI5 decline in sense awareness appears in Arabic, Chinese, English, French, and Russian, indicating the effect is not specific to English.
Readability metrics (FKGL and FRE) show the models do comply with the simplification constraints, so the lost senses are a side effect of style compliance rather than a failure to simplify.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct extension the paper motivates but leaves untested is per-language validation of the GPT-4o mini judge: the judge was validated on 450 English HoWN responses only, and human annotation of non-English ML-WiC samples would tell whether the cross-lingual trend in Figure 4 is a property of the models or an artifact of the evaluator.
If the effect is as general as claimed, a concrete design consequence follows for dictionary apps and educational chatbots: 'simple definition' requests should trigger explicit sense enumeration or a disambiguation question, rather than leaving completeness to the model's default behavior.
The DPO result implies a scalable data recipe that the paper does not spell out: incomplete base-model responses paired with complete responses from other models, labeled only by the automated judge, can generate preference data for more languages and models without new human annotation.
The proposed mechanism—simplification compresses away polysemy—predicts the same completeness collapse in other simplification settings with ambiguous content, such as simplified legal or medical explanations; running the HoWN protocol on those domains would test whether the effect is specific to word definitions or general to information compression under style constraints.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The headline result is real: across five models, completeness of homonym definitions drops monotonically from Normal to Simple to ELI5, and the effect size is big. The HeSA construct is sensible—it distinguishes responses that enumerate all senses from those that at least tell the user more senses exist. HoWN is a grounded resource with coarse-grained WordNet senses, and the DPO mitigation, though small-scale, actually works and generalizes to unseen words. This deserves a serious look.

What is new: HeSA as a metric, HoWN as a dataset, and a filtered multilingual ML-WiC set. The core finding clearly extends prior work on simplification-induced information loss to the specific case of context-free homonym definitions. The English result is independently validated on 450 human-annotated responses, and the judge's category accuracy (93%) and kappa (0.86) are credible.

Soft spots, in proportion. The main one is the cross-lingual layer. The LLM judge is validated only on English, then applied to Arabic, Chinese, French, and Russian with the same English-written prompts. There is no per-language validation, so the multilingual trend in Figure 4 is suggestive, not demonstrated. This should be fixed by either validating the judge per language or downgrading the multilingual claim to a preliminary observation. It does not sink the English finding.

Minor points: the human validation uses one annotator, so the kappas are judge-vs-author rather than inter-annotator reliability; that is acceptable but worth flagging. The DPO section is a single run with 116 preference pairs and no validation set, but the paper is transparent about that, and the gains are large. The sense-matching threshold is a free parameter, yet the monotonic trend is too consistent across models and prompt types to be explained by threshold choice.

Who this is for: anyone working on LLM-based education, text simplification, or evaluation of definition quality. The paper is a competent empirical study with a clear, safety-relevant finding. I would send it to peer review and ask for either per-language judge validation or a re-scoped multilingual claim. Engage with it; the central English result is robust and the resources will be useful.

Referee Report

2 major / 5 minor

Summary. The paper studies whether simplification prompts (Simple, ELI5) reduce LLMs' awareness of multiple senses when defining homonyms. The authors construct two datasets—HoWN (English, WordNet-annotated homonyms) and ML-WiC (multilingual word-in-context)—and evaluate five LLMs under Normal, Simple, ELI5, and Multi-Sense-Aware prompts. They introduce a metric, Helpful Sense Awareness (HeSA), use GPT-4o mini as an automatic judge with human validation on 450 English HoWN responses, and fine-tune Llama 3.1 8B with DPO. The paper reports that definition completeness decreases monotonically from Normal to Simple to ELI5 across all models in HoWN and that Sense Awareness declines across the ML-WiC languages, with DPO improving performance across prompt types and languages.

Significance. If the findings hold, they provide a concrete, practically important demonstration that style constraints aimed at simplification can trade away crucial polysemy information in LLM definitions, with direct implications for educational NLP, accessible interfaces, and users who rely on simplified outputs. The paper's strengths include the release of code and datasets, evaluation of five diverse model families, human validation of the judge on an English sample, and a DPO intervention that improves completeness and generalizes to unseen words. The main caveat is that the cross-lingual and ELI5-specific evaluation components are not yet validated at the level of the central claim; the core English HoWN result, however, is robust across models and consistent with human judgments.

major comments (2)

[§4.3, Figure 4, Appendix F] The central cross-lingual claim rests on an evaluation instrument that is validated only for English. The GPT-4o-mini judge's agreement is reported only against 450 English HoWN responses (Appendix F, Table 5), yet the same English-prompted judge is then used to score Arabic, Chinese, French, Russian, and English ML-WiC responses (Section 4.3, Table 8). No non-English human validation is provided. If the judge systematically undercounts senses in simplified non-English outputs, the cross-lingual trend in Figure 4 could be an artifact of the evaluation instrument rather than a property of the LLMs. Please add per-language human validation (or a language-matched judge) on a sample of ML-WiC responses, or narrow the claim to English until such validation exists.
[§3.4, Appendix D, Figures 9–10] The definition-extraction prompt is not held constant across the conditions being compared. The ELI5-specific prompt (Figure 10) explicitly instructs the judge to treat metaphors, examples, and playful elaborations as one definition and to avoid splitting in a wide range of cases, while the Normal/Simple prompt (Figure 9) uses a different, less elaborate criterion set. This asymmetry can mechanically reduce the number of 'Multiple' classifications in ELI5 responses relative to Normal/Simple responses, potentially inflating the reported decline in sense awareness. The human validation reports high accuracy for ELI5 category classification, but the single annotator may have internalized the same prompt-specific criteria. Please provide a robustness check using an identical extraction prompt across all three conditions, or otherwise demonstrate that the prompt asymmetry does not drive the main result.

minor comments (5)

[§3.5] The 0.4 cosine-similarity threshold for WordNet sense matching is presented without sensitivity analysis; please report how coverage numbers change with the threshold or explicitly treat it as a limitation.
[Appendix F, Table 5] The HeSA kappa for ELI5 is 0.17 despite 94.67% accuracy, which the text attributes to distributional imbalance; please provide the raw confusion matrix or a prevalence-adjusted agreement measure so readers can judge the reliability of HeSA in the condition that drives the strongest decline.
[Section 7 (Limitations)] The statement that 'differences observed between prompt types far exceed any potential error margin' would be more convincing with confidence intervals or a paired significance test (e.g., McNemar) on the HoWN completeness data; the paper currently reports only aggregate percentages.
[Table 6] The caption 'Duplicate WordNet Table' appears to be unfinished; please replace it with a descriptive caption and clarify what the reported percentages represent.
[Section 4.3, Figure 4] The acknowledged exceptions in Chinese (Llama 3.1 8B and Llama 4 Maverick having higher Sense Awareness for Simple than for Normal) are discussed only in text; since Figure 4 presents a cross-lingual trend, please mark these exceptions visually or address them in the figure caption.

Circularity Check

0 steps flagged · score 1.0 of 10

The central claim is an empirical measurement, not a derivation: simplification-induced drops in definition completeness are observed via human-validated judging, not defined into existence. Only a non-load-bearing anonymous self-citation and an English-only-validated multilingual judge are flagged; neither forces the result.

full rationale

The paper's derivation chain is: define completeness/Helpful Sense Awareness (Section 3.3), construct HoWN and ML-WiC datasets (Section 3.1), prompt five LLMs under Normal/Simple/ELI5 (Section 3.2), score responses with a GPT-4o-mini judge validated against 450 English human labels (Section 3.4, Appendix F), and report the decline (Sections 4.1 and 4.3). None of these steps reduces by construction to its inputs. 'We define a response as complete if it either provides all meanings of a word or exhibits Helpful Sense Awareness (HeSA)' is an evaluation criterion, not a fitted parameter renamed as a prediction; the completeness decline is measured on responses generated independently of the judge, and it appears in all five models, including the four that are not the judge (Table 1). The human evaluation on HoWN (Table 5) mirrors the judge scores, and the DPO result is evaluated on 98 unseen words (Section 4.2), so the central claims have independent content. Two passages deserve flags but do not change the verdict. First, the Limitations cites 'Additionally, prior work suggests LLMs underperform in simplified settings, which may affect automated evaluation reliability (Anschütz et al., 2024; Anonymous, 2025)'; the anonymous ARR submission's topic matches the authors' own line of work, but it is used only to acknowledge that automated evaluation may underperform in simplified settings, which weakens rather than forces the paper's conclusion, so it is not load-bearing. Second, the GPT-4o-mini judge is validated only on English HoWN responses and is then applied to Arabic, Chinese, French, and Russian ML-WiC responses (Section 4.3), with an ELI5-specific extraction prompt (Figure 10); this is a genuine measurement-validity risk, since if the judge undercounts senses in non-English simplified outputs the Figure 4 trend could be an artifact, but it is an evaluation-reliability concern, not circularity, because the cross-lingual outcome is not defined in terms of the judge's design. The paper is a self-contained empirical benchmark study; the score reflects only the mild non-load-bearing self-reference and the unvalidated cross-lingual judge application, neither of which reduces the central claim to its inputs.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central measurement rests on a hand-set cosine threshold (0.4), a SemCor frequency filter, and a normative definition of 'complete' (all senses or HeSA). The judge's cross-lingual reliability and WordNet's sense inventory are assumed. No physical entities are invented; HeSA is a scoring construct with internal annotation agreement but no external validation.

free parameters (4)

WordNet sense-match cosine threshold = 0.4
Section 3.5: extracted definitions are mapped to WordNet glosses only when cosine similarity is at least 0.4 using all-MiniLM-L6-v2. No sensitivity analysis is reported, so the 'Covered' and sense-distribution metrics depend on this hand-set cutoff.
SemCor frequency filter = at least 2 occurrences with different senses
Appendix A: only words with at least two SemCor occurrences annotated with different senses are kept in HoWN. This hand-set threshold selects for common polysemous nouns and shapes dataset difficulty.
DPO beta = 0.1
Appendix E, Table 4: the DPO temperature parameter chosen for the fine-tuning run; affects the preference strength in the secondary DPO claim.
LoRA rank and alpha = r=64, alpha=16
Appendix E, Table 4: LoRA configuration for DPO fine-tuning; chosen without a validation split.

assumptions (4)

domain assumption The GPT-4o-mini judge accurately replicates human judgments of definition count and HeSA in all evaluated languages.
Validated only on English HoWN responses (Appendix F: overall kappa 0.86 for category, 0.71 for HeSA), yet applied to Arabic, Chinese, French, Russian, and English responses in Section 4.3.
domain assumption WordNet coarse-grained sense clusters from Proietti et al. (2024) are the correct gold-standard set of senses for homonyms.
Used to define Full, Complete, and coverage metrics in Sections 3.5 and 4.1; if the clustering omits or merges real senses, completeness is miscalibrated.
domain assumption The three prompt templates (Normal, Simple, ELI5) are representative of real user requests, and readability formulas measure language simplicity validly across languages.
The paper verifies that outputs become simpler via FKGL and FRE (Tables 1 and 9), but FRE is an English-tuned formula applied here to French and Russian, and prompt phrasing is known to affect LLM outputs.
ad hoc to paper A correct context-free definition should either enumerate all senses or explicitly signal that senses are missing (HeSA).
This normative completeness target is introduced by the authors in Sections 1 and 3.3. It is not established by external evidence that users require this, so the metric embeds the authors' value judgment.

invented entities (1)

Helpful Sense Awareness (HeSA)
purpose: Scoring criterion that treats a response as complete when it signals unlisted senses or asks for context, even if it does not enumerate every sense.
Introduced in Section 1 and operationalized in Section 3.3. It has internal annotation agreement (kappa 0.71 on 450 responses from one annotator) but no external validation linking HeSA to user comprehension or learning outcomes.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Simplifications are Absolutists: How Simplified Language Reduces Word Sense Awareness in LLM-Generated Definitions." pith.science (2026). https://pith.science/paper/E662FPQX

@misc{pith2026250711981,
  author       = {Pith},
  title        = {Pith review of: Simplifications are Absolutists: How Simplified Language Reduces Word Sense Awareness in LLM-Generated Definitions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E662FPQX}},
  note         = {Machine review of arXiv:2507.11981}
}

read the original abstract

Large Language Models (LLMs) can provide accurate word definitions and explanations for any context. However, the scope of the definition changes for different target groups, like children or language learners. This is especially relevant for homonyms, words with multiple meanings, where oversimplification might risk information loss by omitting key senses, potentially misleading users who trust LLM outputs. We investigate how simplification impacts homonym definition quality across three target groups: Normal, Simple, and ELI5. Using two novel evaluation datasets spanning multiple languages, we test DeepSeek v3, Llama 4 Maverick, Qwen3-30B A3B, GPT-4o mini, and Llama 3.1 8B via LLM-as-Judge and human annotations. Our results show that simplification drastically degrades definition completeness by neglecting polysemy, increasing the risk of misunderstanding. Fine-tuning Llama 3.1 8B with Direct Preference Optimization substantially improves homonym response quality across all prompt types. These findings highlight the need to balance simplicity and completeness in educational NLP to ensure reliable, context-aware definitions for all learners.

Figures

Figures reproduced from arXiv: 2507.11981 by the authors.

**Figure 2.** Definition completeness for each model under [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Distribution of coarse sense coverage across model outputs. The x-axis shows synsets coverage, while the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Sense Awareness across Languages: Llama 3.1 8B (blue line, circle), DPO Llama 3.1 8B (orange line, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: illustrates an example of a complete response, which includes all definitions (Full) as well as both key characteristics of Helpful Sense Awareness (HeSA): (i) a remark indicating that not all possible meanings may be covered, and (ii) a request for additional conte…

**Figure 6.** Figure 6: Frequency distribution of covered WordNet sense ranks across model-generated definitions. The x-axis [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

**Figure 7.** Figure 7: System Prompt used to evaluate Helpful Sense Awareness [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

**Figure 8.** Figure 8: User prompt layout used in all requests. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]

**Figure 9.** Figure 9: System prompt used to extract definitions in the Simple and Normal prompting settings [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]

**Figure 10.** Figure 10: System prompt used to extract definitions in the ELI5 prompting setting [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 11 canonical work pages

[1]

there are other meanings

**remark_not_all_listed** -> Set this to True **only if ** the response explicitly signals that the list of definitions is incomplete. - Set to True when the response includes phrases like: - "there are other meanings" - "this is just a selection" - "some examples include" - "among others" - "etc." - "this is not an exhaustive list" - Set to False when: -...

work page
[2]

if you have a specific context in mind, please tell me

**context_clarification_request** -> Set this to True **only if ** the response explicitly asks the user to provide more context, clarification, or disambiguation. - Set to True when the response includes phrases like: - "if you have a specific context in mind, please tell me" - "can you specify the context the word is used in?" - Set to False when: - The...

work page
[3]

clip" like I am 5 years old. Response: Ooh, let me explain

In general usage: ... The core idea involves deception or directness, depending on context. Does this make sense? **Expected Output **: {’remark_not_all_listed’: False, ’context_clarification_request’: False, ’explanation’: ’remark_not_all_listed is False because the response lists definitions without signaling incompleteness. context_clarification_reques...

work page
[4]

Keep in mind that some words have more than one meaning

Automatically Generated Definitions and their utility for Modeling Word Meaning. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14008–14026, Miami, Florida, USA. Association for Computational Linguistics. Lorenzo Proietti, Stefano Perrella, Simone Tedeschi, Giulia Vulpis, Leonardo Lavalle, et al. 2024. Ana-...

arXiv 2024
[7]

In geography or geology:

work page
[9]

Read the response carefully

work page
[10]

- Treat it as **one definition ** if the response elaborates, gives examples, describes subtypes, or mentions alternative forms **within the same core meaning **

Identify the distinct meanings or senses of the word being defined. - Treat it as **one definition ** if the response elaborates, gives examples, describes subtypes, or mentions alternative forms **within the same core meaning **. - Split it into **multiple definitions ** only if the response clearly and explicitly explains **separate meanings or senses *...

work page
[11]

March means to walk in a formal, organized way as part of a group, often with music and banners

Provide three outputs: - ’definitions’: a list of distinct definitions (if any). - ’category’: either ’One’ or ’Multiple’. - ’explanation’: a brief reason for your categorization. Example 1 (single meaning with details): Input: "March means to walk in a formal, organized way as part of a group, often with music and banners." Output: definitions: - To walk...

work page

Show all 15 references

[12]

Note that child-friendly responses often include metaphors, playful language, or multiple examples to engage young audiences.\n\n

**Read the Response Carefully **: - Identify the word being defined and all meanings or senses provided. Note that child-friendly responses often include metaphors, playful language, or multiple examples to engage young audiences.\n\n
[13]

- Multiple examples or subtypes (e.g., ’paper clips, hair clips’ for a clip)

**Extract Definitions **: - **One Definition **: Count as one if the response describes a single core meaning, even if it includes: - Imaginative metaphors (e.g., ’a big hug from the water’ for a bay). - Multiple examples or subtypes (e.g., ’paper clips, hair clips’ for a clip...
[14]

- **One**: One distinct definition extracted

**Assign Category **: - **None**: No definitions provided. - **One**: One distinct definition extracted. - **Multiple**: Two or more distinct definitions extracted.\n\n
[15]

- ‘category‘: ’None’, ’One’, or ’Multiple’

**Provide Outputs **: - ‘definitions‘: A list of distinct definitions, summarizing each sense concisely. - ‘category‘: ’None’, ’One’, or ’Multiple’. - ‘explanation‘: A brief reason for the categorization, addressing why details are treated as elaborations or distinct senses in...
[2018]

InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 266–271, Melbourne, Australia

Conditional Generators of Words Definitions. InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 266–271, Melbourne, Australia. Association for Computational Linguistics. Aparna Garimella, Abhilasha Sancheti...

2022 arXiv
[2023]

ArXiv:2303.15621 [cs]

ChatGPT as a Factual Inconsistency Evaluator for Text Summarization. ArXiv:2303.15621 [cs]. Federico Martelli, Najla Kalach, Gabriele Tola, and Roberto Navigli. 2021. SemEval-2021 Task 2: Mul- tilingual and Cross-lingual Word-in-Context Dis- ambiguation (MCL-WiC). InProceeding...

2021 arXiv
[2024]

How We Went beyond Word Sense Inventories and Learned to Gloss

Simpler Becomes Harder: Do LLMs Exhibit a Coherent Behavior on Simplified Corpora? InPro- ceedings of the Workshop on DeTermIt! Evaluating Text Difficulty in a Multilingual Context @ LREC- COLING 2024, pages 185–195, Torino, Italia. ELRA and ICCL. Michele Bevilacqua, Marco Mar...

2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

there are other meanings

**remark_not_all_listed** -> Set this to True **only if ** the response explicitly signals that the list of definitions is incomplete. - Set to True when the response includes phrases like: - "there are other meanings" - "this is just a selection" - "some examples include" - "among others" - "etc." - "this is not an exhaustive list" - Set to False when: -...

work page

[2] [2]

if you have a specific context in mind, please tell me

**context_clarification_request** -> Set this to True **only if ** the response explicitly asks the user to provide more context, clarification, or disambiguation. - Set to True when the response includes phrases like: - "if you have a specific context in mind, please tell me" - "can you specify the context the word is used in?" - Set to False when: - The...

work page

[3] [3]

clip" like I am 5 years old. Response: Ooh, let me explain

In general usage: ... The core idea involves deception or directness, depending on context. Does this make sense? **Expected Output **: {’remark_not_all_listed’: False, ’context_clarification_request’: False, ’explanation’: ’remark_not_all_listed is False because the response lists definitions without signaling incompleteness. context_clarification_reques...

work page

[4] [4]

Keep in mind that some words have more than one meaning

Automatically Generated Definitions and their utility for Modeling Word Meaning. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14008–14026, Miami, Florida, USA. Association for Computational Linguistics. Lorenzo Proietti, Stefano Perrella, Simone Tedeschi, Giulia Vulpis, Leonardo Lavalle, et al. 2024. Ana-...

arXiv 2024

[5] [7]

In geography or geology:

work page

[6] [9]

Read the response carefully

work page

[7] [10]

- Treat it as **one definition ** if the response elaborates, gives examples, describes subtypes, or mentions alternative forms **within the same core meaning **

Identify the distinct meanings or senses of the word being defined. - Treat it as **one definition ** if the response elaborates, gives examples, describes subtypes, or mentions alternative forms **within the same core meaning **. - Split it into **multiple definitions ** only if the response clearly and explicitly explains **separate meanings or senses *...

work page

[8] [11]

March means to walk in a formal, organized way as part of a group, often with music and banners

Provide three outputs: - ’definitions’: a list of distinct definitions (if any). - ’category’: either ’One’ or ’Multiple’. - ’explanation’: a brief reason for your categorization. Example 1 (single meaning with details): Input: "March means to walk in a formal, organized way as part of a group, often with music and banners." Output: definitions: - To walk...

work page