How about results

[xyznlp]

Good job. How about rouge scores?

[astariul]

I tried to reproduce results reported in the paper Get to the Point, using same parameters.

But I was unsuccessful, results are much lower than expected, the ROUGE scores are half lower than the ones reported in the paper.

Any idea / pretrained model available ?

[xcfcode]

@Colanim Have you got good results?

[astariul]

Nope, I gave up and tried this repo, which is really easy to reproduce and gave better results anyway.

[xcfcode]

@Colanim Thanks a lot!

[xcfcode]

@Colanim @ymfa I have changed something which can help get better results. @mollynatsu Maybe someone can try to reproduce the results, This is an excellent repo, do not give up. After which I can get ROUGE-L F1 score about 30

I hope this can help the one who will hold this repo in the future.

• The input of decoder should be (word embedding + t-1 time context vector )

https://github.com/ymfa/seq2seq-summarizer/blob/f6b3de3f179d56537d5badff6d3f950a41eff423/model.py#L127

• The ptr calculation should consider the word embedding

https://github.com/ymfa/seq2seq-summarizer/blob/f6b3de3f179d56537d5badff6d3f950a41eff423/model.py#L171

• Init GRU params

    elif isinstance(m, nn.GRU):
        print("Init GRU params...... ")
        for param in m.parameters():
            if len(param.shape) >= 2:
                init.orthogonal_(param.data)
            else:
                init.normal_(param.data)

• I use adam with lr equal to 0.001
• Change attention to MLP attention and create mask
• norm the loss by traget_len