search

SteveKommrusch / OpenNMT-py-ggnn-example

Training, validation, and test files for GGNN encoder option in OpenNMT-py.
10 stars 4 forks source link

Using graph as an input to the model #6

Open Ali-Hatami opened 7 months ago

Ali-Hatami commented 7 months ago

Hi,

I want to use a graph as an input instead of a text in my translation model. In fact, it is a translation model from graph to text. Can I use this lib? It is a bit difficult to follow how can use this. Could you please help me? Thanks

SteveKommrusch commented 7 months ago

Hello,

I think you can use this lib to support a graph-to-sequence translation model. But the output sequencer (the decoder) that it currently supports is the LSTM style, not a transformer. The GGNN I added to OpenNMT-py is based on the paper https://arxiv.org/abs/1511.05493 and the OpenNMT-py connections I added support the graph neural network sequence generator shown in the center of Figure 2 from our paper at https://arxiv.org/abs/2002.06799 . I tried to make the usage rather generic so others could use it for graph-to-sequence generation. You can see some usage Q&A by reading the issues for this repo: https://github.com/SteveKommrusch/OpenNMT-py-ggnn-example/issues .

In my own research I have found that transformer models outperform graph-to-sequence models (and my work was on computer code, not human languages). I think the transformer's key/query attention system is usually better at finding relevant connections than the hand-coded edges for a graph network.

But if you are interested in trying the ggnn, I can help on your attempt. Do you have a write-up (draft paper?) of what you are trying to do?

Regards, Steve