Retrieval augmentation of large language models for lay language generation.

The complex linguistic structures and specialized terminology of expert-authored content limit the accessibility of biomedical literature to the general public. Automated methods have the potential to render this literature more interpretable to readers with different educational backgrounds. Prior work has framed such lay language generation as a summarization or simplification task. However, adapting biomedical text for the lay public includes the additional and distinct task of background explanation: adding external content in the form of definitions, motivation, or examples to enhance comprehensibility. This task is especially challenging because the source document may not include the required background knowledge. Furthermore, background explanation capabilities have yet to be formally evaluated, and little is known about how best to enhance them. To address this problem, we introduce Retrieval-Augmented Lay Language (RALL) generation, which intuitively fits the need for external knowledge beyond that in expert-authored source documents. In addition, we introduce CELLS, the largest (63k pairs) and broadest-ranging (12 journals) parallel corpus for lay language generation. To evaluate RALL, we augmented state-of-the-art text generation models with information retrieval of either term definitions from the UMLS and Wikipedia, or embeddings of explanations from Wikipedia documents. Of these, embedding-based RALL models improved summary quality and simplicity while maintaining factual correctness, suggesting that Wikipedia is a helpful source for background explanation in this context. We also evaluated the ability of both an open-source Large Language Model (Llama 2) and a closed-source Large Language Model (GPT-4) in background explanation, with and without retrieval augmentation. Results indicate that these LLMs can generate simplified content, but that the summary quality is not ideal. Taken together, this work presents the first comprehensive study of background explanation for lay language generation, paving the path for disseminating scientific knowledge to a broader audience. Our code and data are publicly available at: https://github.com/LinguisticAnomalies/pls_retrieval.

Psycholinguistic Markers of COVID-19 Conspiracy Tweets and Predictors of Tweet Dissemination.

The adoption of conspiracy theories about COVID-19 has been fairly widespread among the general public and associated with the rejection of self-protective behaviors. Despite their significance, however, a gap remains in our understanding of the underlying characteristics of messages used to disseminate COVID-19 conspiracies. We used the construct of resonance as a framework to examine a sample of more than 1.8 million posts to Twitter about COVID-19 made between April and June 2020. Our analyses focused on the psycholinguistic properties that distinguish conspiracy theory tweets from other COVID-19 topics and predict their spread. COVID-19 conspiracy tweets were distinct and most likely to resonate when they provided explanations and expressed negative emotions. The results highlight the sensemaking functions served by conspiracy tweets in response to the profound upheaval caused by the pandemic.

Clinician Practice Patterns That Result in the Diagnosis of Coccidioidomycosis Before or During Hospitalization.

BACKGROUND: Coccidioidomycosis (CM) is common and important within endemic regions, requiring specific testing for diagnosis. Long delays in diagnosis have been ascribed to ambulatory clinicians. However, how their testing practices have impacted patient care has not been systematically unexplored. METHODS: We analyzed practice patterns for CM diagnoses over 3 years within a large Arizona healthcare system, including diagnosis location, patient characteristics, and care-seeking patterns associated with missed diagnosis. RESULTS: For 2043 CM diagnoses, 72.9% were made during hospital admission, 21.7% in ambulatory clinics, 3.2% in emergency units, and only 0.5% in urgent care units. A 40.6% subgroup of hospitalized patients required neither intensive care unit or hospital-requiring procedures, had a median length of stay of only 3 days, but still incurred both substantial costs ($27.0 million) and unnecessary antibiotic administrations. Prior to hospital diagnosis (median of 32 days), 45.1% of patients had 1 or more visits with symptoms consistent with CM. During those visits, 71.3% were not tested for CM. Diagnoses were delayed a median of 27 days. CONCLUSIONS: Lack of testing for CM in ambulatory care settings within a region endemic for CM resulted in a large number of hospital admissions, attendant costs, and unneeded antibacterial drug use, much of which would otherwise be unnecessary. Improving this practice is challenging since many clinicians did not train where CM is common, resulting in significant inertia to change. Determining the best way to retrain clinicians to diagnose CM earlier is an opportunity to explore which strategies might be the most effective.

Factors Influencing Willingness to Share Health Misinformation Videos on the Internet: Web-Based Survey.

BACKGROUND: The rapidly evolving digital environment of the social media era has increased the reach of both quality health information and misinformation. Platforms such as YouTube enable easy sharing of attractive, if not always evidence-based, videos with large personal networks and the public. Although much research has focused on characterizing health misinformation on the internet, it has not sufficiently focused on describing and measuring individuals' information competencies that build resilience. OBJECTIVE: This study aims to assess individuals' willingness to share a non-evidence-based YouTube video about strengthening the immune system; to describe types of evidence that individuals view as supportive of the claim by the video; and to relate information-sharing behavior to several information competencies, namely, information literacy, science literacy, knowledge of the immune system, interpersonal trust, and trust in health authority. METHODS: A web-based survey methodology with 150 individuals across the United States was used. Participants were asked to watch a YouTube excerpt from a morning TV show featuring a wellness pharmacy representative promoting an immunity-boosting dietary supplement produced by his company; answer questions about the video and report whether they would share it with a cousin who was frequently sick; and complete instruments pertaining to the information competencies outlined in the objectives. RESULTS: Most participants (105/150, 70%) said that they would share the video with their cousins. Their confidence in the supplement would be further boosted by a friend's recommendations, positive reviews on a crowdsourcing website, and statements of uncited effectiveness studies on the producer's website. Although all information literacy competencies analyzed in this study had a statistically significant relationship with the outcome, each competency was also highly correlated with the others. Information literacy and interpersonal trust independently predicted the largest amount of variance in the intention to share the video (17% and 16%, respectively). Interpersonal trust was negatively related to the willingness to share the video. Science literacy explained 7% of the variance. CONCLUSIONS: People are vulnerable to web-based misinformation and are likely to propagate it on the internet. Information literacy and science literacy are associated with less vulnerability to misinformation and a lower propensity to spread it. Of the two, information literacy holds a greater promise as an intervention target. Understanding the role of different kinds of trust in information sharing merits further research.

Automated Extraction of Diagnostic Criteria From Electronic Health Records for Autism Spectrum Disorders: Development, Evaluation, and Application.

BACKGROUND: Electronic health records (EHRs) bring many opportunities for information utilization. One such use is the surveillance conducted by the Centers for Disease Control and Prevention to track cases of autism spectrum disorder (ASD). This process currently comprises manual collection and review of EHRs of 4- and 8-year old children in 11 US states for the presence of ASD criteria. The work is time-consuming and expensive. OBJECTIVE: Our objective was to automatically extract from EHRs the description of behaviors noted by the clinicians in evidence of the diagnostic criteria in the Diagnostic and Statistical Manual of Mental Disorders (DSM). Previously, we reported on the classification of entire EHRs as ASD or not. In this work, we focus on the extraction of individual expressions of the different ASD criteria in the text. We intend to facilitate large-scale surveillance efforts for ASD and support analysis of changes over time as well as enable integration with other relevant data. METHODS: We developed a natural language processing (NLP) parser to extract expressions of 12 DSM criteria using 104 patterns and 92 lexicons (1787 terms). The parser is rule-based to enable precise extraction of the entities from the text. The entities themselves are encompassed in the EHRs as very diverse expressions of the diagnostic criteria written by different people at different times (clinicians, speech pathologists, among others). Due to the sparsity of the data, a rule-based approach is best suited until larger datasets can be generated for machine learning algorithms. RESULTS: We evaluated our rule-based parser and compared it with a machine learning baseline (decision tree). Using a test set of 6636 sentences (50 EHRs), we found that our parser achieved 76% precision, 43% recall (ie, sensitivity), and >99% specificity for criterion extraction. The performance was better for the rule-based approach than for the machine learning baseline (60% precision and 30% recall). For some individual criteria, precision was as high as 97% and recall 57%. Since precision was very high, we were assured that criteria were rarely assigned incorrectly, and our numbers presented a lower bound of their presence in EHRs. We then conducted a case study and parsed 4480 new EHRs covering 10 years of surveillance records from the Arizona Developmental Disabilities Surveillance Program. The social criteria (A1 criteria) showed the biggest change over the years. The communication criteria (A2 criteria) did not distinguish the ASD from the non-ASD records. Among behaviors and interests criteria (A3 criteria), 1 (A3b) was present with much greater frequency in the ASD than in the non-ASD EHRs. CONCLUSIONS: Our results demonstrate that NLP can support large-scale analysis useful for ASD surveillance and research. In the future, we intend to facilitate detailed analysis and integration of national datasets.

Improving Consumer Understanding of Medical Text: Development and Validation of a New SubSimplify Algorithm to Automatically Generate Term Explanations in English and Spanish.

BACKGROUND: While health literacy is important for people to maintain good health and manage diseases, medical educational texts are often written beyond the reading level of the average individual. To mitigate this disconnect, text simplification research provides methods to increase readability and, therefore, comprehension. One method of text simplification is to isolate particularly difficult terms within a document and replace them with easier synonyms (lexical simplification) or an explanation in plain language (semantic simplification). Unfortunately, existing dictionaries are seldom complete, and consequently, resources for many difficult terms are unavailable. This is the case for English and Spanish resources. OBJECTIVE: Our objective was to automatically generate explanations for difficult terms in both English and Spanish when they are not covered by existing resources. The system we present combines existing resources for explanation generation using a novel algorithm (SubSimplify) to create additional explanations. METHODS: SubSimplify uses word-level parsing techniques and specialized medical affix dictionaries to identify the morphological units of a term and then source their definitions. While the underlying resources are different, SubSimplify applies the same principles in both languages. To evaluate our approach, we used term familiarity to identify difficult terms in English and Spanish and then generated explanations for them. For each language, we extracted 400 difficult terms from two different article types (General and Medical topics) balanced for frequency. For English terms, we compared SubSimplify's explanation with the explanations from the Consumer Health Vocabulary, WordNet Synonyms and Summaries, as well as Word Embedding Vector (WEV) synonyms. For Spanish terms, we compared the explanation to WordNet Summaries and WEV Embedding synonyms. We evaluated quality, coverage, and usefulness for the simplification provided for each term. Quality is the average score from two subject experts on a 1-4 Likert scale (two per language) for the synonyms or explanations provided by the source. Coverage is the number of terms for which a source could provide an explanation. Usefulness is the same expert score, however, with a 0 assigned when no explanations or synonyms were available for a term. RESULTS: SubSimplify resulted in quality scores of 1.64 for English (P<.001) and 1.49 for Spanish (P<.001), which were lower than those of existing resources (Consumer Health Vocabulary [CHV]=2.81). However, in coverage, SubSimplify outperforms all existing written resources, increasing the coverage from 53.0% to 80.5% in English and from 20.8% to 90.8% in Spanish (P<.001). This result means that the usefulness score of SubSimplify (1.32; P<.001) is greater than that of most existing resources (eg, CHV=0.169). CONCLUSIONS: Our approach is intended as an additional resource to existing, manually created resources. It greatly increases the number of difficult terms for which an easier alternative can be made available, resulting in greater actual usefulness.

NegAIT: A new parser for medical text simplification using morphological, sentential and double negation.

Many different text features influence text readability and content comprehension. Negation is commonly suggested as one such feature, but few general-purpose tools exist to discover negation and studies of the impact of negation on text readability are rare. In this paper, we introduce a new negation parser (NegAIT) for detecting morphological, sentential, and double negation. We evaluated the parser using a human annotated gold standard containing 500 Wikipedia sentences and achieved 95%, 89% and 67% precision with 100%, 80%, and 67% recall, respectively. We also investigate two applications of this new negation parser. First, we performed a corpus statistics study to demonstrate different negation usage in easy and difficult text. Negation usage was compared in six corpora: patient blogs (4K sentences), Cochrane reviews (91K sentences), PubMed abstracts (20K sentences), clinical trial texts (48K sentences), and English and Simple English Wikipedia articles for different medical topics (60K and 6K sentences). The most difficult text contained the least negation. However, when comparing negation types, difficult texts (i.e., Cochrane, PubMed, English Wikipedia and clinical trials) contained significantly (p<0.01) more morphological negations. Second, we conducted a predictive analytics study to show the importance of negation in distinguishing between easy and difficulty text. Five binary classifiers (Naïve Bayes, SVM, decision tree, logistic regression and linear regression) were trained using only negation information. All classifiers achieved better performance than the majority baseline. The Naïve Bayes' classifier achieved the highest accuracy at 77% (9% higher than the majority baseline).

Effects on Text Simplification: Evaluation of Splitting Up Noun Phrases.

To help increase health literacy, we are developing a text simplification tool that creates more accessible patient education materials. Tool development is guided by a data-driven feature analysis comparing simple and difficult text. In the present study, we focus on the common advice to split long noun phrases. Our previous corpus analysis showed that easier texts contained shorter noun phrases. Subsequently, we conducted a user study to measure the difficulty of sentences containing noun phrases of different lengths (2-gram, 3-gram, and 4-gram); noun phrases of different conditions (split or not); and, to simulate unknown terms, pseudowords (present or not). We gathered 35 evaluations for 30 sentences in each condition (3 × 2 × 2 conditions) on Amazon's Mechanical Turk (N = 12,600). We conducted a 3-way analysis of variance for perceived and actual difficulty. Splitting noun phrases had a positive effect on perceived difficulty but a negative effect on actual difficulty. The presence of pseudowords increased perceived and actual difficulty. Without pseudowords, longer noun phrases led to increased perceived and actual difficulty. A follow-up study using the phrases (N = 1,350) showed that measuring awkwardness may indicate when to split noun phrases. We conclude that splitting noun phrases benefits perceived difficulty but hurts actual difficulty when the phrasing becomes less natural.

Moving Beyond Readability Metrics for Health-Related Text Simplification.

Limited health literacy is a barrier to understanding health information. Simplifying text can reduce this barrier and possibly other known disparities in health. Unfortunately, few tools exist to simplify text with demonstrated impact on comprehension. By leveraging modern data sources integrated with natural language processing algorithms, we are developing the first semi-automated text simplification tool. We present two main contributions. First, we introduce our evidence-based development strategy for designing effective text simplification software and summarize initial, promising results. Second, we present a new study examining existing readability formulas, which are the most commonly used tools for text simplification in healthcare. We compare syllable count, the proxy for word difficulty used by most readability formulas, with our new metric 'term familiarity' and find that syllable count measures how difficult words 'appear' to be, but not their actual difficulty. In contrast, term familiarity can be used to measure actual difficulty.

Utilizing Large Language Models to Generate Synthetic Data to Increase the Performance of BERT-Based Neural Networks.

An important problem impacting healthcare is the lack of available experts. Machine learning (ML) models may help resolve this by aiding in screening and diagnosing patients. However, creating large, representative datasets to train models is expensive. We evaluated large language models (LLMs) for data creation. Using Autism Spectrum Disorders (ASD), we prompted GPT-3.5 and GPT-4 to generate 4,200 synthetic examples of behaviors to augment existing medical observations. Our goal is to label behaviors corresponding to autism criteria and improve model accuracy with synthetic training data. We used a BERT classifier pretrained on biomedical literature to assess differences in performance between models. A random sample (N=140) from the LLM-generated data was also evaluated by a clinician and found to contain 83% correct behavioral example-label pairs. Augmenting the dataset increased recall by 13% but decreased precision by 16%. Future work will investigate how different synthetic data characteristics affect ML outcomes.

A Comparison of Google and ChatGPT for Automatic Generation of Health-related Multiple-choice Questions.

Critical to producing accessible content is an understanding of what characteristics affect understanding and comprehension. To answer this question, we are producing a large corpus of health-related texts with associated questions that can be read or listened to by study participants to measure the difficulty of the underlying content, which can later be used to better understand text difficulty and user comprehension. In this paper, we examine methods for automatically generating multiple-choice questions using Google's related questions and ChatGPT. Overall, we find both algorithms generate reasonable questions that are complementary; ChatGPT questions are more similar to the snippet while Google related-search questions have more lexical variation.

Text and Audio Simplification: Human vs. ChatGPT.

Text and audio simplification to increase information comprehension are important in healthcare. With the introduction of ChatGPT, evaluation of its simplification performance is needed. We provide a systematic comparison of human and ChatGPT simplified texts using fourteen metrics indicative of text difficulty. We briefly introduce our online editor where these simplification tools, including ChatGPT, are available. We scored twelve corpora using our metrics: six text, one audio, and five ChatGPT simplified corpora (using five different prompts). We then compare these corpora with texts simplified and verified in a prior user study. Finally, a medical domain expert evaluated the user study texts and five, new ChatGPT simplified versions. We found that simple corpora show higher similarity with the human simplified texts. ChatGPT simplification moves metrics in the right direction. The medical domain expert's evaluation showed a preference for the ChatGPT style, but the text itself was rated lower for content retention.

Effects of Added Emphasis and Pause in Audio Delivery of Health Information.

Health literacy is crucial to supporting good health and is a major national goal. Audio delivery of information is becoming more popular for informing oneself. In this study, we evaluate the effect of audio enhancements in the form of information emphasis and pauses with health texts of varying difficulty and we measure health information comprehension and retention. We produced audio snippets from difficult and easy text and conducted the study on Amazon Mechanical Turk (AMT). Our findings suggest that emphasis matters for both information comprehension and retention. When there is no added pause, emphasizing significant information can lower the perceived difficulty for difficult and easy texts. Comprehension is higher (54%) with correctly placed emphasis for the difficult texts compared to not adding emphasis (50%). Adding a pause lowers perceived difficulty and can improve retention but adversely affects information comprehension.

Crowdsourcing with Enhanced Data Quality Assurance: An Efficient Approach to Mitigate Resource Scarcity Challenges in Training Large Language Models for Healthcare.

Large Language Models (LLMs) have demonstrated immense potential in artificial intelligence across various domains, including healthcare. However, their efficacy is hindered by the need for high-quality labeled data, which is often expensive and time-consuming to create, particularly in low-resource domains like healthcare. To address these challenges, we propose a crowdsourcing (CS) framework enriched with quality control measures at the pre-, real-time-, and post-data gathering stages. Our study evaluated the effectiveness of enhancing data quality through its impact on LLMs (Bio-BERT) for predicting autism-related symptoms. The results show that real-time quality control improves data quality by 19% compared to pre-quality control. Fine-tuning Bio-BERT using crowdsourced data generally increased recall compared to the Bio-BERT baseline but lowered precision. Our findings highlighted the potential of crowdsourcing and quality control in resource-constrained environments and offered insights into optimizing healthcare LLMs for informed decision-making and improved patient care.

Transparent deep learning to identify autism spectrum disorders (ASD) in EHR using clinical notes.

OBJECTIVE: Machine learning (ML) is increasingly employed to diagnose medical conditions, with algorithms trained to assign a single label using a black-box approach. We created an ML approach using deep learning that generates outcomes that are transparent and in line with clinical, diagnostic rules. We demonstrate our approach for autism spectrum disorders (ASD), a neurodevelopmental condition with increasing prevalence. METHODS: We use unstructured data from the Centers for Disease Control and Prevention (CDC) surveillance records labeled by a CDC-trained clinician with ASD A1-3 and B1-4 criterion labels per sentence and with ASD cases labels per record using Diagnostic and Statistical Manual of Mental Disorders (DSM5) rules. One rule-based and three deep ML algorithms and six ensembles were compared and evaluated using a test set with 6773 sentences (N = 35 cases) set aside in advance. Criterion and case labeling were evaluated for each ML algorithm and ensemble. Case labeling outcomes were compared also with seven traditional tests. RESULTS: Performance for criterion labeling was highest for the hybrid BiLSTM ML model. The best case labeling was achieved by an ensemble of two BiLSTM ML models using a majority vote. It achieved 100% precision (or PPV), 83% recall (or sensitivity), 100% specificity, 91% accuracy, and 0.91 F-measure. A comparison with existing diagnostic tests shows that our best ensemble was more accurate overall. CONCLUSIONS: Transparent ML is achievable even with small datasets. By focusing on intermediate steps, deep ML can provide transparent decisions. By leveraging data redundancies, ML errors at the intermediate level have a low impact on final outcomes.

Development and evaluation of a biomedical search engine using a predicate-based vector space model.

Although biomedical information available in articles and patents is increasing exponentially, we continue to rely on the same information retrieval methods and use very few keywords to search millions of documents. We are developing a fundamentally different approach for finding much more precise and complete information with a single query using predicates instead of keywords for both query and document representation. Predicates are triples that are more complex datastructures than keywords and contain more structured information. To make optimal use of them, we developed a new predicate-based vector space model and query-document similarity function with adjusted tf-idf and boost function. Using a test bed of 107,367 PubMed abstracts, we evaluated the first essential function: retrieving information. Cancer researchers provided 20 realistic queries, for which the top 15 abstracts were retrieved using a predicate-based (new) and keyword-based (baseline) approach. Each abstract was evaluated, double-blind, by cancer researchers on a 0-5 point scale to calculate precision (0 versus higher) and relevance (0-5 score). Precision was significantly higher (p<.001) for the predicate-based (80%) than for the keyword-based (71%) approach. Relevance was almost doubled with the predicate-based approach-2.1 versus 1.6 without rank order adjustment (p<.001) and 1.34 versus 0.98 with rank order adjustment (p<.001) for predicate--versus keyword-based approach respectively. Predicates can support more precise searching than keywords, laying the foundation for rich and sophisticated information search.

User evaluation of the effects of a text simplification algorithm using term familiarity on perception, understanding, learning, and information retention.

BACKGROUND: Adequate health literacy is important for people to maintain good health and manage diseases and injuries. Educational text, either retrieved from the Internet or provided by a doctor's office, is a popular method to communicate health-related information. Unfortunately, it is difficult to write text that is easy to understand, and existing approaches, mostly the application of readability formulas, have not convincingly been shown to reduce the difficulty of text. OBJECTIVE: To develop an evidence-based writer support tool to improve perceived and actual text difficulty. To this end, we are developing and testing algorithms that automatically identify difficult sections in text and provide appropriate, easier alternatives; algorithms that effectively reduce text difficulty will be included in the support tool. This work describes the user evaluation with an independent writer of an automated simplification algorithm using term familiarity. METHODS: Term familiarity indicates how easy words are for readers and is estimated using term frequencies in the Google Web Corpus. Unfamiliar words are algorithmically identified and tagged for potential replacement. Easier alternatives consisting of synonyms, hypernyms, definitions, and semantic types are extracted from WordNet, the Unified Medical Language System (UMLS), and Wiktionary and ranked for a writer to choose from to simplify the text. We conducted a controlled user study with a representative writer who used our simplification algorithm to simplify texts. We tested the impact with representative consumers. The key independent variable of our study is lexical simplification, and we measured its effect on both perceived and actual text difficulty. Participants were recruited from Amazon's Mechanical Turk website. Perceived difficulty was measured with 1 metric, a 5-point Likert scale. Actual difficulty was measured with 3 metrics: 5 multiple-choice questions alongside each text to measure understanding, 7 multiple-choice questions without the text for learning, and 2 free recall questions for information retention. RESULTS: Ninety-nine participants completed the study. We found strong beneficial effects on both perceived and actual difficulty. After simplification, the text was perceived as simpler (P<.001) with simplified text scoring 2.3 and original text 3.2 on the 5-point Likert scale (score 1: easiest). It also led to better understanding of the text (P<.001) with 11% more correct answers with simplified text (63% correct) compared to the original (52% correct). There was more learning with 18% more correct answers after reading simplified text compared to 9% more correct answers after reading the original text (P=.003). There was no significant effect on free recall. CONCLUSIONS: Term familiarity is a valuable feature in simplifying text. Although the topic of the text influences the effect size, the results were convincing and consistent.

Audio delivery of health information: An NLP study of information difficulty and bias in listeners.

Health literacy is the ability to understand, process, and obtain health information and make suitable decisions about health care [3]. Traditionally, text has been the main medium for delivering health information. However, virtual assistants are gaining popularity in this digital era; and people increasingly rely on audio and smart speakers for health information. We aim to identify audio/text features that contribute to the difficulty of the information delivered over audio. We are creating a health-related audio corpus. We selected text snippets and calculated seven text features. Then, we converted the text snippets to audio snippets. In a pilot study with Amazon Mechanical Turk (AMT) workers, we measured the perceived and actual difficulty of the audio using the response of multiple choice and free recall questions. We collected demographic information as well as bias about doctors' gender, task preference, and health information preference. Thirteen workers completed thirty audio snippets and related questions. We found a strong correlation between text features lexical chain, and the dependent variables, and multiple choice response, percentage of matching word, percentage of similar word, cosine similarity, and time taken (in seconds). In addition, doctors were generally perceived to be more competent than warm. How warm workers perceive male doctors correlated significantly with perceived difficulty.

Evaluation of an online text simplification editor using manual and automated metrics for perceived and actual text difficulty.

Objective: Simplifying healthcare text to improve understanding is difficult but critical to improve health literacy. Unfortunately, few tools exist that have been shown objectively to improve text and understanding. We developed an online editor that integrates simplification algorithms that suggest concrete simplifications, all of which have been shown individually to affect text difficulty. Materials and Methods: The editor was used by a health educator at a local community health center to simplify 4 texts. A controlled experiment was conducted with community center members to measure perceived and actual difficulty of the original and simplified texts. Perceived difficulty was measured using a Likert scale; actual difficulty with multiple-choice questions and with free recall of information evaluated by the educator and 2 sets of automated metrics. Results: The results show that perceived difficulty improved with simplification. Several multiple-choice questions, measuring actual difficulty, were answered more correctly with the simplified text. Free recall of information showed no improvement based on the educator evaluation but was better for simplified texts when measured with automated metrics. Two follow-up analyses showed that self-reported education level and the amount of English spoken at home positively correlated with question accuracy for original texts and the effect disappears with simplified text. Discussion: Simplifying text is difficult and the results are subtle. However, using a variety of different metrics helps quantify the effects of changes. Conclusion: Text simplification can be supported by algorithmic tools. Without requiring tool training or linguistic knowledge, our simplification editor helped simplify healthcare related texts.

Improving the Quality of Suggestions for Medical Text Simplification Tools.

Text continues to be an important medium for communicating health-related information. We have built a text simplification tool that gives concrete suggestions on how to simplify health and medical texts. An important component of the tool identifies difficult words and suggests simpler synonyms based on pre-existing resources (WordNet and UMLS). These candidate substitutions are not always appropriate in all contexts. In this paper, we introduce a filtering algorithm that utilizes semantic similarity based on word embeddings to determine if the candidate substitution is appropriate in the context of the text. We provide an analysis of our approach on a new dataset of 788 labeled substitution examples. The filtering algorithm is particularly helpful at removing obvious examples and can improve the precision by 3% at a recall level of 95%.