A pinboard by
Maury Shenk

My family are Sparrho investors. I am particularly interested in AI aspects of the platform.


AI methods to analyze unstructured text, summarize it, recommend related text, and related use cases

This pinboard covers AI-based text analysis including natural language processing (NLP), through methods such as categorization (e.g. clustering, word/phrase embedding), keyword generation, summarization, recommendation and translation. It excludes related research on (a) general sentiment analysis of text, (b) NLP for conversational applications and (c) natural language generation (e.g. text from images) and text-to-image generation.


An innovative solution for breast cancer textual big data analysis

Abstract: The digitalization of stored information in hospitals now allows for the exploitation of medical data in text format, as electronic health records (EHRs), initially gathered for other purposes than epidemiology. Manual search and analysis operations on such data become tedious. In recent years, the use of natural language processing (NLP) tools was highlighted to automatize the extraction of information contained in EHRs, structure it and perform statistical analysis on this structured information. The main difficulties with the existing approaches is the requirement of synonyms or ontology dictionaries, that are mostly available in English only and do not include local or custom notations. In this work, a team composed of oncologists as domain experts and data scientists develop a custom NLP-based system to process and structure textual clinical reports of patients suffering from breast cancer. The tool relies on the combination of standard text mining techniques and an advanced synonym detection method. It allows for a global analysis by retrieval of indicators such as medical history, tumor characteristics, therapeutic responses, recurrences and prognosis. The versatility of the method allows to obtain easily new indicators, thus opening up the way for retrospective studies with a substantial reduction of the amount of manual work. With no need for biomedical annotators or pre-defined ontologies, this language-agnostic method reached an good extraction accuracy for several concepts of interest, according to a comparison with a manually structured file, without requiring any existing corpus with local or new notations.

Pub.: 06 Dec '17, Pinned: 08 Dec '17

A Deep Relevance Matching Model for Ad-hoc Retrieval

Abstract: In recent years, deep neural networks have led to exciting breakthroughs in speech recognition, computer vision, and natural language processing (NLP) tasks. However, there have been few positive results of deep models on ad-hoc retrieval tasks. This is partially due to the fact that many important characteristics of the ad-hoc retrieval task have not been well addressed in deep models yet. Typically, the ad-hoc retrieval task is formalized as a matching problem between two pieces of text in existing work using deep models, and treated equivalent to many NLP tasks such as paraphrase identification, question answering and automatic conversation. However, we argue that the ad-hoc retrieval task is mainly about relevance matching while most NLP matching tasks concern semantic matching, and there are some fundamental differences between these two matching tasks. Successful relevance matching requires proper handling of the exact matching signals, query term importance, and diverse matching requirements. In this paper, we propose a novel deep relevance matching model (DRMM) for ad-hoc retrieval. Specifically, our model employs a joint deep architecture at the query term level for relevance matching. By using matching histogram mapping, a feed forward matching network, and a term gating network, we can effectively deal with the three relevance matching factors mentioned above. Experimental results on two representative benchmark collections show that our model can significantly outperform some well-known retrieval models as well as state-of-the-art deep matching models.

Pub.: 23 Nov '17, Pinned: 01 Dec '17

Deep Learning to Classify Radiology Free-Text Reports.

Abstract: Purpose To evaluate the performance of a deep learning convolutional neural network (CNN) model compared with a traditional natural language processing (NLP) model in extracting pulmonary embolism (PE) findings from thoracic computed tomography (CT) reports from two institutions. Materials and Methods Contrast material-enhanced CT examinations of the chest performed between January 1, 1998, and January 1, 2016, were selected. Annotations by two human radiologists were made for three categories: the presence, chronicity, and location of PE. Classification of performance of a CNN model with an unsupervised learning algorithm for obtaining vector representations of words was compared with the open-source application PeFinder. Sensitivity, specificity, accuracy, and F1 scores for both the CNN model and PeFinder in the internal and external validation sets were determined. Results The CNN model demonstrated an accuracy of 99% and an area under the curve value of 0.97. For internal validation report data, the CNN model had a statistically significant larger F1 score (0.938) than did PeFinder (0.867) when classifying findings as either PE positive or PE negative, but no significant difference in sensitivity, specificity, or accuracy was found. For external validation report data, no statistical difference between the performance of the CNN model and PeFinder was found. Conclusion A deep learning CNN model can classify radiology free-text reports with accuracy equivalent to or beyond that of an existing traditional NLP model. (©) RSNA, 2017 Online supplemental material is available for this article.

Pub.: 15 Nov '17, Pinned: 01 Dec '17

SkipFlow: Incorporating Neural Coherence Features for End-to-End Automatic Text Scoring

Abstract: Deep learning has demonstrated tremendous potential for Automatic Text Scoring (ATS) tasks. In this paper, we describe a new neural architecture that enhances vanilla neural network models with auxiliary neural coherence features. Our new method proposes a new \textsc{SkipFlow} mechanism that models relationships between snapshots of the hidden representations of a long short-term memory (LSTM) network as it reads. Subsequently, the semantic relationships between multiple snapshots are used as auxiliary features for prediction. This has two main benefits. Firstly, essays are typically long sequences and therefore the memorization capability of the LSTM network may be insufficient. Implicit access to multiple snapshots can alleviate this problem by acting as a protection against vanishing gradients. The parameters of the \textsc{SkipFlow} mechanism also acts as an auxiliary memory. Secondly, modeling relationships between multiple positions allows our model to learn features that represent and approximate textual coherence. In our model, we call this \textit{neural coherence} features. Overall, we present a unified deep learning architecture that generates neural coherence features as it reads in an end-to-end fashion. Our approach demonstrates state-of-the-art performance on the benchmark ASAP dataset, outperforming not only feature engineering baselines but also other deep learning models.

Pub.: 14 Nov '17, Pinned: 08 Dec '17

Quantitative text feature analysis of autobiographical interview data: prediction of episodic details, semantic details and temporal discounting.

Abstract: Autobiographical memory and episodic future thinking (i.e. the capacity to project oneself into an imaginary future) are typically assessed using the Autobiographical Interview (AI). In the AI, subjects are provided with verbal cues (e.g. "your wedding day") and are asked to freely recall (or imagine) the cued past (or future) event. Narratives are recorded, transcribed and analyzed using an established manual scoring procedure (Levine et al., 2002). Here we applied automatic text feature extraction methods to a relatively large (n = 86) set of AI data. In a first proof-of-concept approach, we used regression models to predict internal (episodic) and semantic detail sum scores from low-level linguistic features. Across a range of different regression methods, prediction accuracy averaged at about 0.5 standard deviations. Given the known association of episodic future thinking with temporal discounting behavior, i.e. the preference for smaller-sooner over larger-later rewards, we also ran models predicting temporal discounting directly from linguistic features of AI narratives. Here, prediction accuracy was much lower, but involved the same text feature components as prediction of internal (episodic) details. Our findings highlight the potential feasibility of using tools from quantitative text analysis to analyze AI datasets, and we discuss potential future applications of this approach.

Pub.: 10 Nov '17, Pinned: 14 Nov '17

ADEPt, a semantically-enriched pipeline for extracting adverse drug events from free-text electronic health records.

Abstract: Adverse drug events (ADEs) are unintended responses to medical treatment. They can greatly affect a patient's quality of life and present a substantial burden on healthcare. Although Electronic health records (EHRs) document a wealth of information relating to ADEs, they are frequently stored in the unstructured or semi-structured free-text narrative requiring Natural Language Processing (NLP) techniques to mine the relevant information. Here we present a rule-based ADE detection and classification pipeline built and tested on a large Psychiatric corpus comprising 264k patients using the de-identified EHRs of four UK-based psychiatric hospitals. The pipeline uses characteristics specific to Psychiatric EHRs to guide the annotation process, and distinguishes: a) the temporal value associated with the ADE mention (whether it is historical or present), b) the categorical value of the ADE (whether it is assertive, hypothetical, retrospective or a general discussion) and c) the implicit contextual value where the status of the ADE is deduced from surrounding indicators, rather than explicitly stated. We manually created the rulebase in collaboration with clinicians and pharmacists by studying ADE mentions in various types of clinical notes. We evaluated the open-source Adverse Drug Event annotation Pipeline (ADEPt) using 19 ADEs specific to antipsychotics and antidepressants medication. The ADEs chosen vary in severity, regularity and persistence. The average F-measure and accuracy achieved by our tool across all tested ADEs were 0.83 and 0.83 respectively. In addition to annotation power, the ADEPT pipeline presents an improvement to the state of the art context-discerning algorithm, ConText.

Pub.: 10 Nov '17, Pinned: 14 Nov '17

Artificial Intelligence Learning Semantics via External Resources for Classifying Diagnosis Codes in Discharge Notes.

Abstract: Automated disease code classification using free-text medical information is important for public health surveillance. However, traditional natural language processing (NLP) pipelines are limited, so we propose a method combining word embedding with a convolutional neural network (CNN).Our objective was to compare the performance of traditional pipelines (NLP plus supervised machine learning models) with that of word embedding combined with a CNN in conducting a classification task identifying International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM) diagnosis codes in discharge notes.We used 2 classification methods: (1) extracting from discharge notes some features (terms, n-gram phrases, and SNOMED CT categories) that we used to train a set of supervised machine learning models (support vector machine, random forests, and gradient boosting machine), and (2) building a feature matrix, by a pretrained word embedding model, that we used to train a CNN. We used these methods to identify the chapter-level ICD-10-CM diagnosis codes in a set of discharge notes. We conducted the evaluation using 103,390 discharge notes covering patients hospitalized from June 1, 2015 to January 31, 2017 in the Tri-Service General Hospital in Taipei, Taiwan. We used the receiver operating characteristic curve as an evaluation measure, and calculated the area under the curve (AUC) and F-measure as the global measure of effectiveness.In 5-fold cross-validation tests, our method had a higher testing accuracy (mean AUC 0.9696; mean F-measure 0.9086) than traditional NLP-based approaches (mean AUC range 0.8183-0.9571; mean F-measure range 0.5050-0.8739). A real-world simulation that split the training sample and the testing sample by date verified this result (mean AUC 0.9645; mean F-measure 0.9003 using the proposed method). Further analysis showed that the convolutional layers of the CNN effectively identified a large number of keywords and automatically extracted enough concepts to predict the diagnosis codes.Word embedding combined with a CNN showed outstanding performance compared with traditional methods, needing very little data preprocessing. This shows that future studies will not be limited by incomplete dictionaries. A large amount of unstructured information from free-text medical writing will be extracted by automated approaches in the future, and we believe that the health care field is about to enter the age of big data.

Pub.: 08 Nov '17, Pinned: 14 Nov '17

Using Clinical Notes and Natural Language Processing for Automated HIV Risk Assessment.

Abstract: Universal HIV screening programs are costly, labor-intensive, and often fail to identify high-risk individuals. Automated risk assessment methods that leverage longitudinal electronic health records (EHRs) could catalyze targeted screening programs. While social and behavioral determinants of health are typically captured in narrative documentation, previous analyses have considered only structured EHR fields. We examined whether natural language processing (NLP) would improve predictive models of HIV diagnosis.181 HIV+ individuals received care at New York Presbyterian Hospital prior to a confirmatory HIV diagnosis and 543 HIV negative controls were selected using propensity score matching and included in the study cohort. EHR data including demographics, laboratory tests, diagnosis codes and unstructured notes prior to HIV diagnosis were extracted for modeling. Three predictive algorithms were developed using machine-learning algorithms: (1) a baseline model with only structured EHR data, (2) baseline plus NLP topics, and (3) baseline plus NLP clinical keywords.Predictive models demonstrated a range of performance with F-measures of 0.59 for the baseline model, 0.63 for the baseline + NLP topic model, and 0.74 for the baseline + NLP keyword model. The baseline + NLP keyword model yielded the highest precision by including keywords including 'msm', 'unprotected', 'hiv', and 'methamphetamine' and structured EHR data indicative of additional HIV risk factors.NLP improved the predictive performance of automated HIV risk assessment by extracting terms in clinical text indicative of high-risk behavior. Future studies should explore more advanced techniques for extracting social and behavioral determinants from clinical text.

Pub.: 31 Oct '17, Pinned: 14 Nov '17

Personalized word representations Carrying Personalized Semantics Learned from Social Network Posts

Abstract: Distributed word representations have been shown to be very useful in various natural language processing (NLP) application tasks. These word vectors learned from huge corpora very often carry both semantic and syntactic information of words. However, it is well known that each individual user has his own language patterns because of different factors such as interested topics, friend groups, social activities, wording habits, etc., which may imply some kind of personalized semantics. With such personalized semantics, the same word may imply slightly differently for different users. For example, the word "Cappuccino" may imply "Leisure", "Joy", "Excellent" for a user enjoying coffee, by only a kind of drink for someone else. Such personalized semantics of course cannot be carried by the standard universal word vectors trained with huge corpora produced by many people. In this paper, we propose a framework to train different personalized word vectors for different users based on the very successful continuous skip-gram model using the social network data posted by many individual users. In this framework, universal background word vectors are first learned from the background corpora, and then adapted by the personalized corpus for each individual user to learn the personalized word vectors. We use two application tasks to evaluate the quality of the personalized word vectors obtained in this way, the user prediction task and the sentence completion task. These personalized word vectors were shown to carry some personalized semantics and offer improved performance on these two evaluation tasks.

Pub.: 29 Oct '17, Pinned: 14 Nov '17

Integrating Natural Language Processing and Machine Learning Algorithms to Categorize Oncologic Response in Radiology Reports.

Abstract: A significant volume of medical data remains unstructured. Natural language processing (NLP) and machine learning (ML) techniques have shown to successfully extract insights from radiology reports. However, the codependent effects of NLP and ML in this context have not been well-studied. Between April 1, 2015 and November 1, 2016, 9418 cross-sectional abdomen/pelvis CT and MR examinations containing our internal structured reporting element for cancer were separated into four categories: Progression, Stable Disease, Improvement, or No Cancer. We combined each of three NLP techniques with five ML algorithms to predict the assigned label using the unstructured report text and compared the performance of each combination. The three NLP algorithms included term frequency-inverse document frequency (TF-IDF), term frequency weighting (TF), and 16-bit feature hashing. The ML algorithms included logistic regression (LR), random decision forest (RDF), one-vs-all support vector machine (SVM), one-vs-all Bayes point machine (BPM), and fully connected neural network (NN). The best-performing NLP model consisted of tokenized unigrams and bigrams with TF-IDF. Increasing N-gram length yielded little to no added benefit for most ML algorithms. With all parameters optimized, SVM had the best performance on the test dataset, with 90.6 average accuracy and F score of 0.813. The interplay between ML and NLP algorithms and their effect on interpretation accuracy is complex. The best accuracy is achieved when both algorithms are optimized concurrently.

Pub.: 29 Oct '17, Pinned: 14 Nov '17

Capturing the Patient's Perspective: a Review of Advances in Natural Language Processing of Health-Related Text.

Abstract: Background: Natural Language Processing (NLP) methods are increasingly being utilized to mine knowledge from unstructured health-related texts. Recent advances in noisy text processing techniques are enabling researchers and medical domain experts to go beyond the information encapsulated in published texts (e.g., clinical trials and systematic reviews) and structured questionnaires, and obtain perspectives from other unstructured sources such as Electronic Health Records (EHRs) and social media posts. Objectives: To review the recently published literature discussing the application of NLP techniques for mining health-related information from EHRs and social media posts. Methods: Literature review included the research published over the last five years based on searches of PubMed, conference proceedings, and the ACM Digital Library, as well as on relevant publications referenced in papers. We particularly focused on the techniques employed on EHRs and social media data. Results: A set of 62 studies involving EHRs and 87 studies involving social media matched our criteria and were included in this paper. We present the purposes of these studies, outline the key NLP contributions, and discuss the general trends observed in the field, the current state of research, and important outstanding problems. Conclusions: Over the recent years, there has been a continuing transition from lexical and rule-based systems to learning-based approaches, because of the growth of annotated data sets and advances in data science. For EHRs, publicly available annotated data is still scarce and this acts as an obstacle to research progress. On the contrary, research on social media mining has seen a rapid growth, particularly because the large amount of unlabeled data available via this resource compensates for the uncertainty inherent to the data. Effective mechanisms to filter out noise and for mapping social media expressions to standard medical concepts are crucial and latent research problems. Shared tasks and other competitive challenges have been driving factors behind the implementation of open systems, and they are likely to play an imperative role in the development of future systems.

Pub.: 25 Oct '17, Pinned: 14 Nov '17

A hybrid approach for measuring semantic similarity based on IC-weighted path distance in WordNet

Abstract: As a valuable tool for text understanding, semantic similarity measurement enables discriminative semantic-based applications in the fields of natural language processing, information retrieval, computational linguistics and artificial intelligence. Most of the existing studies have used structured taxonomies such as WordNet to explore the lexical semantic relationship, however, the improvement of computation accuracy is still a challenge for them. To address this problem, in this paper, we propose a hybrid WordNet-based approach CSSM-ICSP to measuring concept semantic similarity, which leverage the information content(IC) of concepts to weight the shortest path distance between concepts. To improve the performance of IC computation, we also develop a novel model of the intrinsic IC of concepts, where a variety of semantic properties involved in the structure of WordNet are taken into consideration. In addition, we summarize and classify the technical characteristics of previous WordNet-based approaches, as well as evaluate our approach against these approaches on various benchmarks. The experimental results of the proposed approaches are more correlated with human judgment of similarity in term of the correlation coefficient, which indicates that our IC model and similarity detection approach are comparable or even better for semantic similarity measurement as compared to others.

Pub.: 05 Sep '17, Pinned: 08 Dec '17

Unsupervised Sentence Representations as Word Information Series: Revisiting TF--IDF

Abstract: Sentence representation at the semantic level is a challenging task for Natural Language Processing and Artificial Intelligence. Despite the advances in word embeddings (i.e. word vector representations), capturing sentence meaning is an open question due to complexities of semantic interactions among words. In this paper, we present an embedding method, which is aimed at learning unsupervised sentence representations from unlabeled text. We propose an unsupervised method that models a sentence as a weighted series of word embeddings. The weights of the word embeddings are fitted by using Shannon's word entropies provided by the Term Frequency--Inverse Document Frequency (TF--IDF) transform. The hyperparameters of the model can be selected according to the properties of data (e.g. sentence length and textual gender). Hyperparameter selection involves word embedding methods and dimensionalities, as well as weighting schemata. Our method offers advantages over existing methods: identifiable modules, short-term training, online inference of (unseen) sentence representations, as well as independence from domain, external knowledge and language resources. Results showed that our model outperformed the state of the art in well-known Semantic Textual Similarity (STS) benchmarks. Moreover, our model reached state-of-the-art performance when compared to supervised and knowledge-based STS systems.

Pub.: 17 Oct '17, Pinned: 14 Nov '17