Capability to get probabilities for each tag in crf_decode

[agarwalishan]

Describe the feature and the current behavior/state. I am talking about this specific repository. https://www.tensorflow.org/addons/api_docs/python/tfa/text/crf_decode For now crf_decode function returns the tag indices with highest probability and a score for the whole sequence. But the crf_decode do not returns the probabilities for every tag for each token in sequence like we get in softmax. So I request to please add this capability because probabilities of tags help us know the confidence score and many more things.

Relevant information

• Are you willing to contribute it (yes/no): yes
• Are you willing to maintain it going forward? (yes/no): no
• Is there a relevant academic paper? (if so, where): no
• Is there already an implementation in another framework? (if so, where): I guess in pytorch
• Was it part of tf.contrib? (if so, where): yes, tf.contrib.crf.crf_decode

Which API type would this fall under (layer, metric, optimizer, etc.) It will fall under 'text'

Who will benefit with this feature? I believe many developers around the globe are looking for this feature.

Any other info. Confidence scores for any model are very important for further decisions and modifications.

[Ghostvv]

Hi, we implemented the probability for each tag here: https://github.com/RasaHQ/rasa/blob/master/rasa/utils/tensorflow/crf.py

[howl-anderson]

Regarding to returns the probabilities for every tag for each token in sequence. Do you means you want the emission probability of each words?

if so, I think you can use tfa.layers.CRF to get the potentials (as logits).

In following sample code:

layer = tfa.layers.CRF(4, use_boundary=False)
decoded_sequence, potentials, sequence_length, chain_kernel = layer(inputs)

potentials is the thing you want.

[Ghostvv]

yeah, but potentials are not probabilities. CRF also takes into account transition probabilities

[Ghostvv]

there is an open PR for that: https://github.com/tensorflow/addons/pull/1935

[howl-anderson]

@Ghostvv Thank you for the information you provided. I will do some investigation on the PR. Then, hopefully, do some work based on that.

[seanpmorgan]

TensorFlow Addons is transitioning to a minimal maintenance and release mode. New features will not be added to this repository. For more information, please see our public messaging on this decision: TensorFlow Addons Wind Down

Please consider sending feature requests / contributions to other repositories in the TF community with a similar charters to TFA: Keras Keras-CV Keras-NLP