Words contain meanings. This task is to learn the vector representations of words for meaning encoding.
Phase 1: implementation of word embedding models for English data
Phase 2: implementation of same models for Chinese PTT data
[Phase 1 Word Correlation Competition]
[Phase 2 Word Analogy Competition]
Sentiment can be analyzed based on the observed language. This task is to predict the sentiment for English sentences using recursive neural networks.
Bonus 1: prediction of the sentiment for Chinese data
Bonus 2: implementation of convolutional neural networks for sentiment analysis
Language understanding is the core component of intelligent systems, where there are mainly two types of information to be extracted: intents and slots. Intent classification and slot filling should be applied in order to understand natural language.
Task 1: filling the slots given the input sentence
Task 2: prediction of the user intention from the input sentence
Language generation is to emit natural language texts given 1) natural language in another language or 2) user intent associated with specified slots.
Task 1: Machine translation generates target language given the source language
Task 2: Natural language generation generates natural language given the structured semantic frames
Game playing is an interaction between a user and an environment. Deep reinforcement learning can learn the agent's action for plying the game.
Task: implement deep Q-network for the agent to play a Atari game.
Assist A1
Assist A1 and A4
Assist A2
Assist A2
Assist A3
Assist A3
Assist A4 and A5
Assist A5