Scoring clinical patient notes (PNs) written by medical students is a necessary but resource-intensive task in medical education. This paper describes the organization and key lessons from a Kaggle competition on automated scoring of such notes. 1,471 teams took part in the competition and developed an extensive, publicly available code repository of varying solutions evaluated over the first public dataset for this task. The most successful approaches from this community effort are described and utilized in the development of a PN scoring system. We discuss the choice of models and system architecture with a view to operational use and scalability, and evaluate its performance on both the public Kaggle data (10 clinical cases, 43,985 PNs) and an extended internal dataset (178 clinical cases, 6,940 PNs). The results show that the system significantly outperforms a state-of-the-art existing tool for PN scoring and that task-adaptive pretraining using masked language modeling can be an effective approach even for small training samples.

This paper reports findings from the First Shared Task on Automated Prediction of Difficulty and Response Time for Multiple-Choice Questions. The task was organized as part of the 19th Workshop on Innovative Use of NLP for Building Educational Applications (BEA’24), held in conjunction with NAACL 2024, and called upon the research community to contribute solutions to the problem of modeling difficulty and response time for clinical multiple-choice questions (MCQs). A set of 667 previously used and now retired MCQs from the United States Medical Licensing Examination (USMLE®) and their corresponding difficulties and mean response times were made available for experimentation. A total of 17 teams submitted solutions and 12 teams submitted system report papers describing their approaches. This paper summarizes the findings from the shared task and analyzes the main approaches proposed by the participants.

This paper presents the ACTA system, which performs automated short-answer grading in the domain of high-stakes medical exams. The system builds upon previous work on neural similarity-based grading approaches by applying these to the medical domain and utilizing contrastive learning as a means to optimize the similarity metric. ACTA is evaluated against three strong baselines and is developed in alignment with operational needs, where low-confidence responses are flagged for human review. Learning curves are explored to understand the effects of training data on performance. The results demonstrate that ACTA leads to substantially lower number of responses being flagged for human review, while maintaining high classification accuracy.

This paper presents a corpus of 43,985 clinical patient notes (PNs) written by 35,156 examinees during the high-stakes USMLE® Step 2 Clinical Skills examination. In this exam, examinees interact with standardized patients - people trained to portray simulated scenarios called clinical cases. For each encounter, an examinee writes a PN, which is then scored by physician raters using a rubric of clinical concepts, expressions of which should be present in the PN. The corpus features PNs from 10 clinical cases, as well as the clinical concepts from the case rubrics. A subset of 2,840 PNs were annotated by 10 physician experts such that all 143 concepts from the case rubrics (e.g., shortness of breath) were mapped to 34,660 PN phrases (e.g., dyspnea, difficulty breathing). The corpus is available via a data sharing agreement with NBME and can be requested at https://www.nbme.org/services/data-sharing.

This study examines the relationship between the linguistic characteristics of a test item and the complexity of the response process required to answer it correctly. Using data from a large-scale medical licensing exam, clustering methods identified items that were similar with respect to their relative difficulty and relative response-time intensiveness to create low response process complexity and high response process complexity item classes. Interpretable models were used to investigate the linguistic features that best differentiated between these classes from a descriptive and predictive framework. Results suggest that nuanced features such as the number of ambiguous medical terms help explain response process complexity beyond superficial item characteristics such as word count. Yet, although linguistic features carry signal relevant to response process complexity, the classification of individual items remains challenging.

This paper brings together approaches from the fields of NLP and psychometric measurement to address the problem of predicting examinee proficiency from responses to short-answer questions (SAQs). While previous approaches train on manually labeled data to predict the human-ratings assigned to SAQ responses, the approach presented here models examinee proficiency directly and does not require manually labeled data to train on. We use data from a large medical exam where experimental SAQ items are embedded alongside 106 scored multiple-choice questions (MCQs). First, the latent trait of examinee proficiency is measured using the scored MCQs and then a model is trained on the experimental SAQ responses as input, aiming to predict proficiency as its target variable. The predicted value is then used as a “score” for the SAQ response and evaluated in terms of its contribution to the precision of proficiency estimation.

This paper investigates whether transfer learning can improve the prediction of the difficulty and response time parameters for 18,000 multiple-choice questions from a high-stakes medical exam. The type the signal that best predicts difficulty and response time is also explored, both in terms of representation abstraction and item component used as input (e.g., whole item, answer options only, etc.). The results indicate that, for our sample, transfer learning can improve the prediction of item difficulty when response time is used as an auxiliary task but not the other way around. In addition, difficulty was best predicted using signal from the item stem (the description of the clinical case), while all parts of the item were important for predicting the response time.

One of the most resource-intensive problems in the educational testing industry relates to ensuring that newly-developed exam questions can adequately distinguish between students of high and low ability. The current practice for obtaining this information is the costly procedure of pretesting: new items are administered to test-takers and then the items that are too easy or too difficult are discarded. This paper presents the first study towards automatic prediction of an item’s probability to “survive” pretesting (item survival), focusing on human-produced MCQs for a medical exam. Survival is modeled through a number of linguistic features and embedding types, as well as features inspired by information retrieval. The approach shows promising first results for this challenging new application and for modeling the difficulty of expert-knowledge questions.

Predicting the construct-relevant difficulty of Multiple-Choice Questions (MCQs) has the potential to reduce cost while maintaining the quality of high-stakes exams. In this paper, we propose a method for estimating the difficulty of MCQs from a high-stakes medical exam, where all questions were deliberately written to a common reading level. To accomplish this, we extract a large number of linguistic features and embedding types, as well as features quantifying the difficulty of the items for an automatic question-answering system. The results show that the proposed approach outperforms various baselines with a statistically significant difference. Best results were achieved when using the full feature set, where embeddings had the highest predictive power, followed by linguistic features. An ablation study of the various types of linguistic features suggested that information from all levels of linguistic processing contributes to predicting item difficulty, with features related to semantic ambiguity and the psycholinguistic properties of words having a slightly higher importance. Owing to its generic nature, the presented approach has the potential to generalize over other exams containing MCQs.

We present a novel approach to automatic question answering that does not depend on the performance of an information retrieval (IR) system and does not require that the training data come from the same source as the questions. We evaluate the system performance on a challenging set of university-level medical science multiple-choice questions. Best performance is achieved when combining a neural approach with an IR approach, both of which work independently. Unlike previous approaches, the system achieves statistically significant improvement over the random guess baseline even for questions that are labeled as challenging based on the performance of baseline solvers.

NLP approaches to automatic text adaptation often rely on user-need guidelines which are generic and do not account for the differences between various types of target groups. One such group are adults with high-functioning autism, who are usually able to read long sentences and comprehend difficult words but whose comprehension may be impeded by other linguistic constructions. This is especially challenging for real-world user-generated texts such as product reviews, which cannot be controlled editorially and are thus a particularly good applcation for automatic text adaptation systems. In this paper we present a mixed-methods survey conducted with 24 adult web-users diagnosed with autism and an age-matched control group of 33 neurotypical participants. The aim of the survey was to identify whether the group with autism experienced any barriers when reading online reviews, what these potential barriers were, and what NLP methods would be best suited to improve the accessibility of online reviews for people with autism. The group with autism consistently reported significantly greater difficulties with understanding online product reviews compared to the control group and identified issues related to text length, poor topic organisation, and the use of irony and sarcasm.

Developing plausible distractors (wrong answer options) when writing multiple-choice questions has been described as one of the most challenging and time-consuming parts of the item-writing process. In this paper we propose a fully automatic method for generating distractor suggestions for multiple-choice questions used in high-stakes medical exams. The system uses a question stem and the correct answer as an input and produces a list of suggested distractors ranked based on their similarity to the stem and the correct answer. To do this we use a novel approach of combining concept embeddings with information retrieval methods. We frame the evaluation as a prediction task where we aim to “predict” the human-produced distractors used in large sets of medical questions, i.e. if a distractor generated by our system is good enough it is likely to feature among the list of distractors produced by the human item-writers. The results reveal that combining concept embeddings with information retrieval approaches significantly improves the generation of plausible distractors and enables us to match around 1 in 5 of the human-produced distractors. The approach proposed in this paper is generalisable to all scenarios where the distractors refer to concepts.

When processing a text, humans and machines must disambiguate between different uses of the pronoun it, including non-referential, nominal anaphoric or clause anaphoric ones. In this paper we use eye-tracking data to learn how humans perform this disambiguation and use this knowledge to improve the automatic classification of it. We show that by using gaze data and a POS-tagger we are able to significantly outperform a common baseline and classify between three categories of it with an accuracy comparable to that of linguistic-based approaches. In addition, the discriminatory power of specific gaze features informs the way humans process the pronoun, which, to the best of our knowledge, has not been explored using data from a natural reading task.

In recent years gaze data has been increasingly used to improve and evaluate NLP models due to the fact that it carries information about the cognitive processing of linguistic phenomena. In this paper we conduct a preliminary study towards the automatic identification of multiword expressions based on gaze features from native and non-native speakers of English. We report comparisons between a part-of-speech (POS) and frequency baseline to: i) a prediction model based solely on gaze data and ii) a combined model of gaze data, POS and frequency. In spite of the challenging nature of the task, best performance was achieved by the latter. Furthermore, we explore how the type of gaze data (from native versus non-native speakers) affects the prediction, showing that data from the two groups is discriminative to an equal degree for the task. Finally, we show that late processing measures are more predictive than early ones, which is in line with previous research on idioms and other formulaic structures.

Given the lack of large user-evaluated corpora in disability-related NLP research (e.g. text simplification or readability assessment for people with cognitive disabilities), the question of choosing suitable training data for NLP models is not straightforward. The use of large generic corpora may be problematic because such data may not reflect the needs of the target population. The use of the available user-evaluated corpora may be problematic because these datasets are not large enough to be used as training data. In this paper we explore a third approach, in which a large generic corpus is combined with a smaller population-specific corpus to train a classifier which is evaluated using two sets of unseen user-evaluated data. One of these sets, the ASD Comprehension corpus, is developed for the purposes of this study and made freely available. We explore the effects of the size and type of the training data used on the performance of the classifiers, and the effects of the type of the unseen test datasets on the classification performance.

Eye tracking studies from the past few decades have shaped the way we think of word complexity and cognitive load: words that are long, rare and ambiguous are more difficult to read. However, online processing techniques have been scarcely applied to investigating the reading difficulties of people with autism and what vocabulary is challenging for them. We present parallel gaze data obtained from adult readers with autism and a control group of neurotypical readers and show that the former required higher cognitive effort to comprehend the texts as evidenced by three gaze-based measures. We divide all words into four classes based on their viewing times for both groups and investigate the relationship between longer viewing times and word length, word frequency, and four cognitively-based measures (word concreteness, familiarity, age of acquisition and imagability).

This paper presents an approach for automatic evaluation of the readability of text simplification output for readers with cognitive disabilities. First, we present our work towards the development of the EasyRead corpus, which contains easy-to-read documents created especially for people with cognitive disabilities. We then compare the EasyRead corpus to the simplified output contained in the LocalNews corpus (Feng, 2009), the accessibility of which has been evaluated through reading comprehension experiments including 20 adults with mild intellectual disability. This comparison is made on the basis of 13 disability-specific linguistic features. The comparison reveals that there are no major differences between the two corpora, which shows that the EasyRead corpus is to a similar reading level as the user-evaluated texts. We also discuss the role of Simple Wikipedia (Zhu et al., 2010) as a widely-used accessibility benchmark, in light of our finding that it is significantly more complex than both the EasyRead and the LocalNews corpora.

The paper presents a corpus of text data and its corresponding gaze fixations obtained from autistic and non-autistic readers. The data was elicited through reading comprehension testing combined with eye-tracking recording. The corpus consists of 1034 content words tagged with their POS, syntactic role and three gaze-based measures corresponding to the autistic and control participants. The reading skills of the participants were measured through multiple-choice questions and, based on the answers given, they were divided into groups of skillful and less-skillful readers. This division of the groups informs researchers on whether particular fixations were elicited from skillful or less-skillful readers and allows a fair between-group comparison for two levels of reading ability. In addition to describing the process of data collection and corpus development, we present a study on the effect that word length has on reading in autism. The corpus is intended as a resource for investigating the particular linguistic constructions which pose reading difficulties for people with autism and hopefully, as a way to inform future text simplification research intended for this population.