The book chapter on sentiment analysis results in a validation accuracy of 100% right in the first epoch. (Thanks to @Sharan for pointing this).
It seems that the error has something to do with the buf_size
in the paddle reader: https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_understand_sentiment.py#L177
On increasing the buf_size
to higher values (like 30000), we get reasonable results.