Performance on the paper's dataset about hierarchical-attention-networks

[superzhangxing]

@gabrer Hi, do you have re-implemented on dataset of original paper?
In my implementation, I can't get the same performance according to the paper.

[MH23333]

@superzhangxing Hi, in my implementation, I can only get 63.7% accuracy rate on the dev dataset of yelp2013 under the same configuration according to the paper. The accuracy rate on the test dataset is supposed to be a bit lower, which is different from the performance in the paper. Is there some idea? Can those tricks reduce the gap？

[superzhangxing]

@MH23333 Hi, do you use pre-trained word embeddings or random initialized ones? I train them with word2vec method on train and dev dataset. I believe it will improve the accurate. I have the same configuration according to the paper and get around 67% accurate on dev and test dataset. Such as optimizer with SGD+Momentum, and momentum parameter with 0.9. The only trick I think is aligning the sentence length in each batch to accelerate the training speed. and it also has been mentioned in the paper.

[MH23333]

@superzhangxing Appreciate for your quickly reply！I use word embeddings in the way same as you, and set all hyper parameters mentioned in the paper. Maybe other unmentioned hyper parameters have an important influence on the results. How to set the following two parameters may be important: sentence length(how many words in a sentnece) and document length(how many sentences in a document).
I will do more experiments. Many thanks!

[superzhangxing]

@MH23333 The max sentence length and max document length are both set 40. Please note that I use dynamic rnn , so I don't use fixed sentence length or fixed document length. I'm not sure whether it makes the influence.

[MH23333]

@superzhangxing Thanks a lot! I also use dynamic rnn and masked attention. More contrast tests will be performed. Hope I can get better results.