The latest releases of TensorFlow 2 have been accompanied by excellent tutorials which made it straightforward to manipulate and adapt complex models such as RNNs, CNNs, and GANs to a variety of tasks. However, I did not find any tutorials showing any of these models in structured data, such as the constituent tree.
For instance, in Natural Language Processing with Deep Learning course from Stanford was shown a tree RNN for sentiment analysis and there are even some implementations using older versions of TensorFlow. However, these implementations are very complex to understand and adapted to other tasks.
Then, could a tutorial be published showing how to apply tree RNN to the sentiment analysis task using Tensorflow 2? I think the main points to address would be how to adapt the computation graph of the model to the different trees existing in the training set (i.e. should the model be created for each instance of the training set?)
I think this tutorial would help the community understand and apply dynamic models to a myriad of tasks.
@yashk2810 for FYI
Thank you for posting this.
Agree with you. This should be an interesting topic.
Will get back to the documentation & community team
The latest releases of TensorFlow 2 have been accompanied by excellent tutorials which made it straightforward to manipulate and adapt complex models such as RNNs, CNNs, and GANs to a variety of tasks. However, I did not find any tutorials showing any of these models in structured data, such as the constituent tree.
For instance, in Natural Language Processing with Deep Learning course from Stanford was shown a tree RNN for sentiment analysis and there are even some implementations using older versions of TensorFlow. However, these implementations are very complex to understand and adapted to other tasks.
Then, could a tutorial be published showing how to apply tree RNN to the sentiment analysis task using Tensorflow 2? I think the main points to address would be how to adapt the computation graph of the model to the different trees existing in the training set (i.e. should the model be created for each instance of the training set?)
I think this tutorial would help the community understand and apply dynamic models to a myriad of tasks.