Improving broad-coverage medical entity linking with semantic type prediction and large-scale datasets.

OBJECTIVES: Biomedical natural language processing tools are increasingly being applied for broad-coverage information extraction-extracting medical information of all types in a scientific document or a clinical note. In such broad-coverage settings, linking mentions of medical concepts to standardized vocabularies requires choosing the best candidate concepts from large inventories covering dozens of types. This study presents a novel semantic type prediction module for biomedical NLP pipelines and two automatically-constructed, large-scale datasets with broad coverage of semantic types. METHODS: We experiment with five off-the-shelf biomedical NLP toolkits on four benchmark datasets for medical information extraction from scientific literature and clinical notes. All toolkits adopt a staged approach of mention detection followed by two stages of medical entity linking: (1) generating a list of candidate concepts, and (2) picking the best concept among them. We introduce a semantic type prediction module to alleviate the problem of overgeneration of candidate concepts by filtering out irrelevant candidate concepts based on the predicted semantic type of a mention. We present MedType, a fully modular semantic type prediction model which we integrate into the existing NLP toolkits. To address the dearth of broad-coverage training data for medical information extraction, we further present WikiMed and PubMedDS, two large-scale datasets for medical entity linking. RESULTS: Semantic type filtering improves medical entity linking performance across all toolkits and datasets, often by several percentage points of F-1. Further, pretraining MedType on our novel datasets achieves state-of-the-art performance for semantic type prediction in biomedical text. CONCLUSIONS: Semantic type prediction is a key part of building accurate NLP pipelines for broad-coverage information extraction from biomedical text. We make our source code and novel datasets publicly available to foster reproducible research.

Ambiguity in medical concept normalization: An analysis of types and coverage in electronic health record datasets.

OBJECTIVES: Normalizing mentions of medical concepts to standardized vocabularies is a fundamental component of clinical text analysis. Ambiguity-words or phrases that may refer to different concepts-has been extensively researched as part of information extraction from biomedical literature, but less is known about the types and frequency of ambiguity in clinical text. This study characterizes the distribution and distinct types of ambiguity exhibited by benchmark clinical concept normalization datasets, in order to identify directions for advancing medical concept normalization research. MATERIALS AND METHODS: We identified ambiguous strings in datasets derived from the 2 available clinical corpora for concept normalization and categorized the distinct types of ambiguity they exhibited. We then compared observed string ambiguity in the datasets with potential ambiguity in the Unified Medical Language System (UMLS) to assess how representative available datasets are of ambiguity in clinical language. RESULTS: We found that <15% of strings were ambiguous within the datasets, while over 50% were ambiguous in the UMLS, indicating only partial coverage of clinical ambiguity. The percentage of strings in common between any pair of datasets ranged from 2% to only 36%; of these, 40% were annotated with different sets of concepts, severely limiting generalization. Finally, we observed 12 distinct types of ambiguity, distributed unequally across the available datasets, reflecting diverse linguistic and medical phenomena. DISCUSSION: Existing datasets are not sufficient to cover the diversity of clinical concept ambiguity, limiting both training and evaluation of normalization methods for clinical text. Additionally, the UMLS offers important semantic information for building and evaluating normalization methods. CONCLUSIONS: Our findings identify 3 opportunities for concept normalization research, including a need for ambiguity-specific clinical datasets and leveraging the rich semantics of the UMLS in new methods and evaluation measures for normalization.

Modeling Dialogue Acts with Content Word Filtering and Speaker Preferences.

We present an unsupervised model of dialogue act sequences in conversation. By modeling topical themes as transitioning more slowly than dialogue acts in conversation, our model de-emphasizes content-related words in order to focus on conversational function words that signal dialogue acts. We also incorporate speaker tendencies to use some acts more than others as an additional predictor of dialogue act prevalence beyond temporal dependencies. According to the evaluation presented on two dissimilar corpora, the CNET forum and NPS Chat corpus, the effectiveness of each modeling assumption is found to vary depending on characteristics of the data. De-emphasizing content-related words yields improvement on the CNET corpus, while utilizing speaker tendencies is advantageous on the NPS corpus. The components of our model complement one another to achieve robust performance on both corpora and outperform state-of-the-art baseline models.

Data mining and education.

An emerging field of educational data mining (EDM) is building on and contributing to a wide variety of disciplines through analysis of data coming from various educational technologies. EDM researchers are addressing questions of cognition, metacognition, motivation, affect, language, social discourse, etc. using data from intelligent tutoring systems, massive open online courses, educational games and simulations, and discussion forums. The data include detailed action and timing logs of student interactions in user interfaces such as graded responses to questions or essays, steps in rich problem solving environments, games or simulations, discussion forum posts, or chat dialogs. They might also include external sensors such as eye tracking, facial expression, body movement, etc. We review how EDM has addressed the research questions that surround the psychology of learning with an emphasis on assessment, transfer of learning and model discovery, the role of affect, motivation and metacognition on learning, and analysis of language data and collaborative learning. For example, we discuss (1) how different statistical assessment methods were used in a data mining competition to improve prediction of student responses to intelligent tutor tasks, (2) how better cognitive models can be discovered from data and used to improve instruction, (3) how data-driven models of student affect can be used to focus discussion in a dialog-based tutoring system, and (4) how machine learning techniques applied to discussion data can be used to produce automated agents that support student learning as they collaborate in a chat room or a discussion board.

When are tutorial dialogues more effective than reading?

It is often assumed that engaging in a one-on-one dialogue with a tutor is more effective than listening to a lecture or reading a text. Although earlier experiments have not always supported this hypothesis, this may be due in part to allowing the tutors to cover different content than the noninteractive instruction. In 7 experiments, we tested the interaction hypothesis under the constraint that (a) all students covered the same content during instruction, (b) the task domain was qualitative physics, (c) the instruction was in natural language as opposed to mathematical or other formal languages, and (d) the instruction conformed with a widely observed pattern in human tutoring: Graesser, Person, and Magliano's 5-step frame. In the experiments, we compared 2 kinds of human tutoring (spoken and computer mediated) with 2 kinds of natural-language-based computer tutoring (Why2-Atlas and Why2-AutoTutor) and 3 control conditions that involved studying texts. The results depended on whether the students' preparation matched the content of the instruction. When novices (students who had not taken college physics) studied content that was written for intermediates (students who had taken college physics), then tutorial dialogue was reliably more beneficial than less interactive instruction, with large effect sizes. When novices studied material written for novices or intermediates studied material written for intermediates, then tutorial dialogue was not reliably more effective than the text-based control conditions.