wikillm

Concept · 2 papers · updated 2026-10-25

Text simplification

Rewriting dense text so a reader without specialist training can follow it. Two papers so far. In a radiology study, 15 radiologists rated 45 chatbot rewrites: mostly correct, but with errors and dropped results. A service robot restates answers in German easy language through an external API, with no measurements.

text-simplificationmedical-nlpllmhuman-evaluationhuman-robot-interaction

Markdown5 min read

AI-compiled overview, updated 2026-10-25, not yet reviewed by a human. Every claim cites its source; check it when a detail matters.

Overview

Text simplification rewrites a text so that a reader without specialist training can follow it, while the content stays the same. Two papers with cards apply it in different settings. In the radiology study, the content that must survive is the medical content: a simpler text that changes a finding has failed, however readable it is. The study checks this with radiologist ratings of factual correctness, completeness and potential harm 2212.14882 §3.3. In the service-robot paper, the rewrite restates official information in German easy language on request. The paper does not measure whether the restated text is understood or stays accurate 2412.05022 §6 §7.

Variants

Prompted chat model, expert-rated

One fixed instruction, asking the model to explain the report to a child in simple language, is sent with each report to a general-purpose chat model. Each input is run 15 times, and every run starts a new chat so that earlier turns do not influence the output 2212.14882 §3.2. The prompt was picked from six informal trials on a different fictitious report and is not claimed to be the best prompt 2212.14882 §3.2 §5.3. Ratings: the median factual correctness and completeness were both 2 (agree), and the median potential harm was 4 (disagree) 2212.14882 tab. 1. The harm statement describes the risk of wrong conclusions, so the low harm median means raters did not see much risk. Card: ChatGPT Makes Medicine Easy to Swallow.

Easy-language service inside a robot dialogue

A humanoid service robot answers a customer's question from a stored knowledge base. When the customer asks, the stored answer is sent over the internet to an external easy-language service (SUMM AI), and the result is spoken and shown on a tablet 2412.05022 §6. The paper names no prompt, model or service settings, and the service's version is not stated 2412.05022 §6. Easy language is offered for German only 2412.05022 §6. Translation is a separate mode, through DeepL, and only Danish was tested [arxiv:2412.05022 §6]. The paper reports no quantitative results. Workshop feedback with public-service staff is described as positive, without numbers 2412.05022 §7. Card: Get It Right: Improving Comprehensibility with Adaptable Speech Expression of a Humanoid Service Robot.

Other variants

None else on record. Other simplification methods exist, but no paper with a card describes one, so this page cannot compare them.

When to use which

The two studies cover one model in one interface, and one service on one robot. The rows state only what their results support.

SituationUseWhy
Output goes to a patient with no clinician in the loopDo not rely on this approach aloneRaters found misread terms, invented findings and missed results. The authors recommend physician oversight 2212.14882 §4.2 §5.2
Output is checked by a clinician before releasePrompted chat model, with a review stepMost raters judged the rewrites correct, complete and not harmful 2212.14882 tab. 1, and the authors recommend that a radiologist checks them 2212.14882 §5.2
Results must be reproduced exactlyStore the outputs; do not plan on re-running themThe interface had no temperature or seed control, and the authors call exact reproduction "difficult, if not almost impossible" 2212.14882 §3.2 §5.2
Output length mattersMeasure length against the sourceMedian word counts rose for knee MRI (222 to 414) and head MRI (117 to 150), but were similar for oncology CT (241 to 249) 2212.14882 tab. 4
Citizen-facing answers in German, from a fixed knowledge baseEasy-language service, with your own accuracy checkThe authors show the flow works in a demo. The paper reports no accuracy or comprehension measure 2412.05022 §6 §7
Target language is not GermanTranslation service, tested in your languageEasy language covers German only. Translation is shown for Danish only, though the service lists 27 languages 2412.05022 §6
Customer data would be sent to a third partyCheck the data flow firstCustomer utterances and answers pass through external services. The paper names data protection as open and describes no safeguards 2412.05022 §6 §7
Spoken input to a robotChat dialogue mode over keyword spottingThe authors chose Chat because keyword spotting handles only known phrases. The gain is described without numbers 2412.05022 §6.2
A baseline simplifier is neededNot answerable from these studiesNo other system was compared 2212.14882 §3.2 §4; 2412.05022 §6
A rating design is neededThree Likert statements plus free-text highlightsThis is the design the radiology study used 2212.14882 §3.3

Implementation sketch

The radiology study has no training step, so the first sketch covers the generation and rating procedure only. The robot paper gives no algorithm, so the second sketch restates its request flow in our own words, with the numbers it omits marked as missing. Both are written in our own words.

Pseudocode
# Generate candidate rewrites for one source text  [arxiv:2212.14882 §3.2]
# Input:  source text, instruction string, repeats = 15
# Output: list of candidate rewrites

def simplify(source, instruction, repeats=15):
    outputs = []
    for _ in range(repeats):
        # Fresh session each time, so earlier turns do not leak in  [§3.2]
        reply = chat_model(new_session=True, prompt=instruction + source)
        outputs.append(reply)
    return outputs
Pseudocode
# Build one rating sheet per rater  [arxiv:2212.14882 §3.3]
# Input:  source texts, candidates per source, rater
# Output: one (source, random candidate) pair per source

def build_sheet(sources, candidates, rater):
    # Each source gets one randomly chosen rewrite  [§3.3]
    return [(s, random_choice(candidates[s])) for s in sources]
Pseudocode
# Restate a stored answer on request  [arxiv:2412.05022 §6]
# Input:  customer question, mode ("easy" or "translate")
# Output: text to speak and show; no algorithm stated in the paper

def restate(question, mode):
    # Find the stored answer for the matched topic  [§6]
    answer = knowledge_base(match_topic(question))
    if mode == "easy":
        # Service is German only in the paper  [§6]
        return easy_language_service(answer, lang="de")
    # Translation service; the demo used Danish  [§6]
    return translation_service(answer, target="da")

Each pair in the radiology study is rated on three statements on a 5-point scale, with a free-text follow-up for wrong text, missing information and harmful conclusions 2212.14882 §3.3. Summaries use medians and quartiles, because the ratings are ordinal 2212.14882 §3.4.

Contradictions

  • Harm maximum. The text says raters never chose "strongly disagree" for harm 2212.14882 §4.1 §5.1.3. The summary table gives a maximum of 5 for harm, and the per-report table has 5 in every report [arxiv:2212.14882 tab. 1] tab. 2. The tables are used here.
  • Share finding harm. The text says about half of the raters found critical statements 2212.14882 §5.1.3. The follow-up table says 36% listed harmful conclusions 2212.14882 tab. 3. The paper does not explain the gap.
  • Between the two papers. None known. They measure different things, and the robot paper reports no numbers that could be set against the radiology ratings 2412.05022 §7.

Papers

Related

ConceptPrompt engineeringSteering a frozen language model through its input: instructions, demonstrations, formats, reasoning cues, personas and layouts, and with weight access, learned soft prompts. Format changes alone moved one model's accuracy by up to 76 points; a zero-shot prompt tied 10-shot on one translation model. Evidence is mostly from 2021-2025 models, often one prompt per task.ConceptHuman evaluationJudging generated outputs, dialogues or robot behaviour with people and turning the judgments into a verdict. Averaged Likert ratings can reverse the true preference when systems are close; a per-annotator probability question recovered more expected orderings in one study. Pairwise trials feed a ranking model. Results are author-reported, from twenty-one papers.ConceptLLM evaluationHow to measure what a large language model can do, and what changes the reported number: prompt, format, decoding, model snapshot, metric, judge and ground truth. Variants cover fixed-prompt benchmarks, setting and prompt sweeps, complexity curves, minimal-edit probes, LLM and human judges, simulated participants, reading-time fit, bias probes, safety rates, surveys and hallucination checks.ConceptText summarizationShorten a text, such as news, spoken dialogue or clinical notes, into a summary that keeps the key content. Extractive methods copy sentences, abstractive methods write new ones, and hybrid methods do both. Fine-tuning needs labelled pairs; a prompted LLM needs none. Long inputs need a long-context model. Automatic scores can disagree with human ratings.ConceptHuman-in-the-loopA person stays inside an AI system's decision or learning loop: choosing outputs, teaching through labels or corrections, sharing control, or limiting autonomy. Twelve cards cover these. Evidence ranges from position papers and a survey to user studies, trials and robot experiments.ConceptLLM ethicsEthical and social risks of language models: risk taxonomies and a 1,725-risk database, auditing blueprints, bias, toxicity and stereotype measures, cultural and religious bias, memorised training text, web corpus filtering, carbon cost, authorship and copyright, education, mental-health and clinical use, and agency. Most results are author-reported; cards note how each was measured.

Source and license

Compiled from the papers listed on this page, all published under CC BY, CC BY-SA or CC0, with citations. See each paper's page for its authors and license.