Jiaxing Xu's Page
Somewhere beyond the barricade is there a world you long to see.
Hi! I am a research intern in Automatic Drive Group, SenseTime Group Limited. I just graduate from College of Software, Beihang University. Before joining SenseTime, I worked as a research intern in Knowledge Computing Group, Microsoft Research Asia and a research assistant in Knowledge Engineering Group of Tsinghua University.
My research interests lie in natural language processing and data mining, especially on dialog system and graph neural network. I do believe these fields make invaluable contributions to the real world.
New!! I'm applying for PhD programs and searching for a Research Assistant position currently. Don't hesitate to contact me if you are interested.Curriculum Vitae
Researchers always have the demand to know what areas a conference focus on and how this conference developed. Thus, we purpose a method to automatically extract keywords from every paper of a conference, construct it to a concept graph, and combine those small graphs to a concept graph of a particular conference. This framework could not only help researchers quickly understand a conference, but also help us discover new concepts.
For languages with no annotated resources, transferring knowledge from rich-resource languages is an effective solution for named entity recognition (NER). While all existing methods directly transfer from source-learned model to a target language, in this paper, we propose to fine-tune the learned model with a few similar examples given a test case, which could benefit the prediction by leveraging the structural and semantic information conveyed in such similar examples. We also present an active learning algorithm to quickly find informative samples from unlabeled data that could fast improve the performance of our model.
The baseline method of non-task-oriented dialog system is directly imputing the goal, knowledge and conversation into the encoder without using the structure features. To enhance the performance, we consider constructing a graph based on the required knowledge and using GNN to learn the structure features of the graph. Then the GNN output is combined with the origin encoder to improve the dialog agent.
In order to handle unique properties of player trajectory prediction, we deliver a two-folded research procedure in which we firstly proposed a novel LSTM model called TAS-LSTM to appropriately modeling the group feature under team aware situations. We further consider the influence of the ball’s movement and proposed a new approach.
• Teaching machines to converse naturally with humans is challenging and really interesting.
• We propose a fantastic system to help people construct their own chatbots:
• An interactive syntax tree help people to define question rules.
• Use visualization methods to understand how to make mock sentences.
• Use BLSTM-CRF-NER model and LSTM classifier to construct chatbot.
This development selects the application scenario of the trusted deployment of the lease contract, optimizes the existing NLP model, and creates an algorithm for contract processing to automatically generate the java smart contract code and deploy it to the blockchain to supervise the transaction process in real time. The protocol and user interface automatically execute the contract content and record the behavior track in real time, thus ensuring the security and convenience of the lease transaction.
Manual disease coding is time-consuming and expensive. We develop a model based on 3000 real cases and large-scale drug database. Thus, we can recommend medication to patients through their symptoms.
Aside from my curiosity about computer science and my study preoccupation, I am also a big fan of music and go game. I would like to listen to the concert, play guitar or watch go game competition in my free time. If this involves sports, I favor swimming since it enhances my self-control and it is a benefit in my health insurance and spirit.