0lidaxiang / ir

This is the IR course of NTUST in 2017. IR means that Information Retrieval and Its Applications, including Vector Model, word2Vec technology and so on.

IR

This is the homeworks of NTUST CSIE IR course in 2017. Homework3, homework4 and homework5 will test result in Kaggle website.

Kaggle Three Competitions's link is below:
Homework3
Homework4
Homework5

sourceData

This is the data files that should be put in data or source floder of homework1,homework3,homework4,homework5.

Homework1

This program is right.
Vector Model.

Homework2

This program is right.
Compute MAP.

Homework3

This program is wrong.EM Algorithm maybe right.
EM Algorithm.

Homework4

This program result effect is bad.Algorithm maybe right. Rocchio Algorithm applying in IR system.

Homework5

This program is right. But the keras train maybe not right.The CBOW result may be better only need 20~50 times train. My 2500 times train's good result is just a coincidence.And maybe we should given the word in query but not in document the socre of the three lowest TF_IDF socre, teacher's advice.

Our goal is to implement a word embedding method by using Keras, and then you should leverage the learned embeddings to do retrieval. Use CBOW, word embeddings, One-hot, Vector model. And we also can use Rocchio Algorithm to imporve the result's MAP.