Learning to Generate Reviews and Discovering Sentiment

[flrngel]

https://arxiv.org/abs/1704.01444
paper from OpenAI

Summary

This paper shows that "Your training dataset should cover your target tasks" by using mLSTM and Sentiment Analysis

Abstract

• When given sufficient amounts of capacity, training data, and compute time, byte-level recurrent language model can achieve good performance

4. Experimental Setup and Results

4.3. Capacity Ceiling

• There is a notable drop in paper's approach transitioning from sentence to document datasets