Microsoft AI Challenge 2018 - Solution

My solution to Microsoft AI Challenge 2018

This solution achieved a spot in Top 20 teams out of 2000 teams that were participating[*].

Microsoft organized Microsoft AI Challenge 2018 for India in November 2018. The problem statement was to figure out the right response to queries(web) given 10 possible responses.

How to use?

• Create a python(3.6) enviroment with libraries from requirements.txt
• Place the dataset in data folder i.e. it should contain Data.tsv and eval1_unlabelled.tsv as provided by the organisers
• Run preprocessing python preprocess.py(It standardizes the column names, creates a local validation dataset and undersamples the training dataset for fast and effective training. All the processed files will be saved in processed folder)
• Run the training python training.py, it saves the model parameters in model_weights folder as well
• The above steps also generates the submission in the current directory

Files

• dynamic_clip_attention.py - Model
• utils.py - Utility Functions
• preprocess.py - Preprocessing
• training.py - Training and prediction

References

This solution is a modified version of https://github.com/wjbianjason/Dynamic-Clip-Attention

[*] The actual solution was an ensemble of models.