About Me
I am currently a Research Associate in NLP4Health Group supervised by Prof.V.G.Vinod Vydiswaran at University of Michigan. I am interested in NLP, IR and CV models and applications in health-care domain. I am applying Ph.D programs related with NLP.
ResumeUniversity of Michigan
NLP, IR, CV
jiazhaol AT umich DOT edu
Shenyang, China
Video
This is one interesting HCI game I designed using computer vision. Thank my teammates, we got hightest score in EECS 442
Electronic Transcription of Prescriptions (ETP) in health care has been suggested to have positive impact for prescribing process in safety, quality, efficiency for both prescribers and pharmacists. We are trying to use Deep learning method to automatically transcripe prescription to pharmacy instructions. More details in following paper.
We found movie revenue followed two Gaussion Distrubution: High revenue and Low revenue. Based on this obeservation, we trained High-low binary Random Forest Classifier (RFC performed best) based on labels from Gaussion Mixture Model (GMM) clustering. After classification, we trained movie Gradient Boosting Regression (GBR performed best) separately and using Back-off strategy to solving cold start problem. We also tested generalization of our classification + regression model on Europe Soccer Players.
Identifying medication relations between drugs and associated attributes automatically from clinical narratives can help develop advanced tools for decision support. We investigate the strengths of neural network models to identify eight medication relations. We find that relation extraction is sensitive to complexity of data patterns as well as model capacity. Our results show that Bi-LSTM models achieved an overall F1 score of 0.892 on eight tasks, outperforming SVM and CNN models.
Traditional methods simply relay on manual hash tags, title and description of video compared with free text query, which suffers from high bias and low accuracy. We proposed a novel video segments retrieval system based on architecture Attentive Convolutional Nerual Network (ACNN), achieving state-of-the-art performance on TACoS dataset.
In the study, we systematically analyzed questions posted on pregnancy forums by young mothers and contrasted it to a unique dataset of Facebook posts by expecting adolescent women. Distribution of themes across these platforms showed significant differences on roles played by the two platforms. We concluded that Facebook is chosen as a self-expression place to seek emotional support while health forum served as professional information provider.
Prediction of soccer competition results is always an atractive topic. However, traditional ranking method, Network ofNetworks (NoN), can only capture the similarity between playersor teams rather than the win-lose relationship. Broadly used method Deep Neural Network (DNN) cannot converge within limited dataset. In this work, we proposed a graph-based method PageRank++ to predict the rank of teams in a league. Our algorithm converges faster without using large dataset compared with DNN and higher accuracy compared with baselines, Naive Rank and Naive PageRank.
I am responsible for research project in Machine Translation / Simplification in Health Care domain and techinical part User study Web application developing. More details and writings in Research above. My work is supervised by Prof Vinod.
I did research projects about Topic Model on social medias, Deep Learning Models on clinical notes. More details and writings in Research above. My work was supervised by Prof. Vinod.
I was grader of lecture EECS 498 Information Retrieval and Web Search taught by Prof. Rada Mihalcea. Main resposiblity is to grade codes and reports.
GPA: 3.856/4.00
GPA: 88/100