Jessica Karpinski

Advancing Multi-label Classification Using Deep Neural Networks ---------- is developing advanced machine learning methods with its multi-label classification model, ----------. Multi-label (ML) text classification has applications spanning industries such as marketing, product management, academia, governance, and biomedicine. It is used to assign textual documents one or more labels for organization and analysis purposes. -------- has been developing technology that improves current models and systems to advance accuracy and efficiency. ---------- is ----------’s novel end-to-end deep learning framework that has been validated for use in biomedical text classification. ---------- Architecture Researchers have been advancing machine learning methods for the past several decades but those methods often require human effort for feature engineering and have certain limitations for what they can achieve. ---------- has utilized the knowledge, skills, and expertise of Dr. ----------, the Director of Natural Language Processing (NLP) Research at the company, to develop its multi-label classification model known as ----------. The architecture of this model is comprised of three major modules: 1. a document encoding network that utilizes input in the form of raw text and converts it into output of high-dimensional vectors representing the whole document. This is done through a two-step process of: a. the embeddings from Language Models (ELMo) network that uses the raw text to generate contextualized embeddings for each word; and b. an RRN network that uses those contextualized word embeddings as inputs and generates the appropriate document representation 2. a label prediction network that takes document vectors as input and formulates a prediction confidence score for each label; and 3. a label count prediction network with several layers with an output layer for which the number of nodes equals the number of maximal permitted labels. The layer also inputs the same document vectors to create an output of the estimation of label counts for each document. Prior to evaluation, ---------- first had its prediction network training. As part of the first step of this training, the label prediction and the document encoding networks are updated through back propagation. Next, the label count prediction network was trained by updating the multi-layer perceptron (MLP) as the gradient descent stops at the layer of the document vector. All of the individual labels are ranked by their corresponding confidence scores that have been generated by the label prediction network. Then the top labels are used as the final output. Evaluation ---------- was evaluated with classification tasks for three different topics with publicly available data sets from biomedical literature and clinical notes. The topics were hallmarks of cancer classification, chemical exposure assessments, and diagnosis codes assignment. It was found that ---------- outperformed all of the other the multi-level classification baseline models and the team at ---------- intends to keep improving and advancing the ---------- network. ----------’s ---------- has proven to be more advanced and accurate than binary relevance methods in various biomedical contexts. Since the network does not require human effort for tasks such as feature engineering, the process of classification is expedited. With ----------’s efficiency and scalability, it can be used for large sets of labels, making it invaluable for biomedical researchers and companies to organize vast amounts of information. To learn more about how ----------’s technology can expedite your research processes, request a demo by clicking below.