Intent recognition tool

[katerynaCh]

This PR implements an intent recognition tool that can classify text data into the following intents: [ 'Complain', 'Praise', 'Apologise', 'Thank', 'Criticize', 'Agree', 'Taunt', 'Flaunt', 'Joke', 'Oppose', 'Comfort', 'Care', 'Inform', 'Advise', 'Arrange', 'Introduce', 'Leave', 'Prevent', 'Greet', 'Ask for help' ]. This tool is meant to be used together with the speech transcription tool for speech-to-text conversion and hence the ROS nodes are expected to subscribe to output of speech transcription ROS node. Similarly, python demo is provided with integration of the speech transcription tool.

The model implements an architecture that is trained in a multimodal manner (audio, visual, text) with the goal of improving inference on one of the modalities, in our case, text. The reference to the paper describing the approach will be added in the places where it is missing soon, once the paper/preprint is published. We provide several pretrained models of different sizes, from bert-tiny to bert-base.

This tool can be fully tested once the speech transcription (whisper) tool (#433) is integrated (or by just reusing its code) but I would rather start the initial review sooner if possible.

I am not sure why some tools' tests are failing, but seems unrelated to this PR?

[tsampazk]

Hello again @katerynaCh. Thank you for the fixes so far.

I ran some more tests and debugging and added more comments. I also went ahead and applied some changes required for ROS1/2 in files that were previously not changed (it would make it awkward for the review process otherwise).

[tsampazk]

I went ahead and added the fix from #465 to the new ROS1 node.

Thank you very much @katerynaCh, everything looks good to me!

[passalis]

Thank you!