2016年5月8日 (日) 11:21的版本

Text Processing Team Schedule

Members

Former Members

Rong Liu (刘荣) : 优酷
Xiaoxi Wang (王晓曦) : 图灵机器人
Xi Ma (马习) : 清华大学研究生
DongXu Zhang (张东旭) : --

Current Members

Tianyi Luo (骆天一)
Chao Xing (邢超)
Qixin Wang (王琪鑫)
Yiqiao Pan (潘一桥)
Aodong Li (李傲冬)
Ziwei Bai (白子薇)
Aiting Liu (刘艾婷)

Work Process

Question answering system

Chao Xing

2016-05-05 :

            Coding for RNN-DSSM in tensorflow. Face an error when running rnn-dssm model in cpu : memory keep increasing. 
            Tensorflow's version in huilan is 0.7.0 and install by pip, this cause using error in creating gpu graph,
            one possible solution is build tensorflow from source code.

Aiting Liu

2016-05-08：Fetch data from 'http://news.ifeng.com/' and 'http://www.xinhuanet.com/'(13.4M)

2016-05-07：Fetch data from 'http://fangtan.china.com.cn/' and interview books (10M)

2016-05-04：Establish the overall framework of our chat robot,and continue to build database

Ziwei Bai

2016-05-04：Finding and dealing with the data for QA system

Reproduce DSSM Baseline (Chao Xing)

2016-04-28 : Given a talk to text team for some recently paper.

              Knowledge Base Completion via Search-Based Question Answering : pdf
              Open Domain Question Answering via Semantic Enrichment  : pdf
              A Neural Conversational Model : pdf
              And given a tiny results for CNN-DSSM in huilan's weekly report.

2016-04-27 : Code Multi-layer CNN, suffered from memory error in GPU in tensorflow.

              So I run such test on CPU, should slow.

2016-04-26 : Code done tricky & analysis such tricky.

2016-04-25 : Find a tricky to improve accuracy given by Tianyi.

            : Code for this tricky.

2016-04-23 : Set a series of experiment set.

              1. Try deep CNN-DSSM, current model just follow proposed model contain one convolution layer, need to be a tuneable parameter.
              2. Test whether mixture data effective to current model and deep CDSSM.
              3. Code Recurrent CNN-DSSM (new approach.)

2016-04-22 : Find a problem : Use labs' gpu machine 970 iteration per time is 1537 second but huilan's server is just 7 second.

              Achieve reasonable results when apply max-margin method to CNN-DSSM model.

2016-04-21 : True DSSM model doesn't work well, analysis as below:

               1. Not exactly reproduce DSSM model, because the original one is English version, I just adapt it to Chinese but after word segmentation. 
                  So the input is tri-gram words not tri-gram letter.
               2. Our dataset far from rich, because of we do not use pre-trained word vectors as initial vectors, we can hardly achieve good performance.
            : Request
               1. As we have rich pre-trained word vectors, maybe CDSSM or RDSSM corrected to our task.
               2. Different length of sequences seek to be fixed dimension vectors, just CNN and RNN can do such things, DNN can not do it by using 
                 fix length of word vectors
            : Coding done CDSSM. Test for it's performance.
               One problem : When you install tensorflow by pip 0.8.0 and you want to use conv2d function by gpu, you need make sure you had already 
                            install your cudnn's version as 4.0 not lastest 5.0.

2016-04-20 : Find reproduced DSSM model's bug, fix it.

2016-04-19 : Code mixture data model by less memory dependency done. Test it's performance.

2016-04-18 : Code mixture data model.

2016-04-16 : Code mixture data model, but face to memory error. Dr. Wang help me fix it.

2016-04-15 : Share Papers. Investigation a series of DSSM papers for future work. And show our intern students how to do research.

            : Original DSSM model : Learning Deep Structured Semantic Models for Web Search using Clickthrough Data pdf
            : CNN based DSSM model : A Latent Semantic Model with Convolutional-Pooling Structure for Information Retrieval pdf
            : Use DSSM model for a new area : Modeling Interestingness with Deep Neural Networks pdf
            : Latest approach for LSTM + RNN DSSM model : SEMANTIC MODELLING WITH LONG-SHORT-TERM MEMORY FOR INFORMATION RETRIEVAL pdf

2016-04-14 : Test dssm-dnn model, code dssm-cnn model.

              Continue investigate deep neural question answering system.

2016-04-13 : test dssm model, investigate deep neural question answering system.

            : Share theano ppt theano
            : Share tensorflow ppt tensorflow

2016-04-12 : Write done dssm tensor flow version.

2016-04-11 : Write tensorflow toolkit ppt for intern student.

2016-04-10 : Learn tensorflow toolkit.

2016-04-09 : Learn tensorflow toolkit.

2016-04-08 : Finish theano version.

Deep Poem Processing With Image (Ziwei Bai)

2016-04-20 :combine my program with Qixin Wang's

2016-04-10 : web spider to catch a thousand pices of images.

2016-04-13 :1、download theano for python2.7。 2.debug cnn.py

2016-04-15 :web spider to catch 30 thousands pices of images and store them into a matrix

2016-04-16 :modify the code of CNN and spider

2016-04-17 :train convouloutional neural network

RNN Piano Processing (Jiyuan Zhang)

2016-4-12：select appropriate midis and run rnnrbm model

2016-4-13：view rnnrbm model‘s code

2016-4-14~15:coding to select 4/4 beat of midis

2016-4-17~22:run data, failed several times ，then modify code and view rnnrbm model's code

2016-4-25~29:replace rnnrbm with lstmrbm, then run lstmrbm's model

Question & Answering (Aiting Liu)

2016-04-24 : make my biweekly report

2016-04-23 : read Fader's paper (2011)

2016-04-20 : read Fader's paper (2013)

2016-04-15 : learn dssm and sent2vec

2016-04-16 : try to figure out how the PARALAX dataset is constructed

2016-04-17 : download the PARALAX dataset and try to turn it into what we want it to be

Generation Model (Aodong li)

2016-05-07 :

   study attention-based model
   learn some details about the poem generation model
   change my focus onto lyrics generation model

2016-05-06 : read the paper about poem generation and learn about LSTM

2016-05-05 : check in and have an overview of generation model

“Schedule”版本间的差异