DeepLearningNLU

Coding Task NUL for the course "Deep learning for speech & language processing" WS 20/21

Notes on Coding Tasks

• data will be published on Dec 1st, 2020 at 12am on the excercise webpage
• you will get 2 files: one for training (contains labels), one for evaluation (does not contain labels)
• for evaluation, upload your predictions in the same format as the download
  • each data sample in a file has a unique id
    • upload only the data sample id and your predicted label
    • don't upload the input data (there is a size limit for each upload)
• format: JSON
  • python has built-in functionallity to read / write JSON
  • if you never used it (with python), there is a short overview: https://realpython.com/python-json/

NLU train data

{
    "0": {
        "intent": "AddToPlaylist",
        "text": "Add a tune to my elrow Guest List",
        "slots": {
            "music_item": "tune",
            "playlist_owner": "my",
            "playlist": "elrow Guest List"
        },
        "positions": {
            "music_item": [
                6,
                9
            ],
            "playlist_owner": [
                14,
                15
            ],
            "playlist": [
                17,
                32
            ]
        }
    },
    "1": {
        "intent": "AddToPlaylist",
        "text": "Add a guy is a guy to the infinite indie folk playlist.",
        "slots": {
            "entity_name": "a guy is a guy",
            "playlist": "infinite indie folk"
        },
        "positions": {
            "entity_name": [
                4,
                17
            ],
            "playlist": [
                26,
                44
            ]
        }
    },
...

---

Upload your predictions without "text" and "position" fields!

###Emotion train data features-entry contains a list of a list with 26 items.

length of inner list: 26 (float numbers - represent one preprocessed speech frame (logMel)) length of outer list: number of frames per data-point, e.g. 10 or 15, ... for easy parsing into an array, have a look at e.g. https://docs.scipy.org/doc/numpy-1.15.1/reference/generated/numpy.fromstring.html

{"0": {"valence": 0, 
       "activation": 1,
       "features": [[5.502810676891276, 5.389630715979907, 5.89079939835461, 5.462222408074205, 5.482823967430299, 6.776999420417389, 7.500191549370183, 8.304382413190043, 8.01058604345238, 6.033706406358724, 7.067212416470815, 6.4463667723745965, 6.923879604046541, 8.091544891207468, 7.972158444902915, 7.641215050710871, 7.1507373690379525, 7.507756141316357, 7.074741477204279, 7.992954055443716, 7.576059013864094, 7.7207549746760975, 8.649263126832178, 8.220149828892483, 6.811228837696475, 6.457873509412178], [...], ...]
       },
 "1: "valence": 1, 
       "activation": 1,
       "features": [[3.502810676891276, 5.389630715979907, 5.89079939835461, 5.462222408074205, 5.482823967430299, 6.776999420417389, 7.500191549370183, 8.304382413190043, 8.01058604345238, 6.033706406358724, 7.067212416470815, 6.4463667723745965, 2.923879604046541, 8.091544891207468, 7.972158444902915, 7.641215050710871, 7.1507373690379525, 7.507756141316357, 7.074741477204279, 7.992954055443716, 7.576059013864094, 7.7207549746760975, 8.649263126832178, 8.220149828892483, 6.811228837696475, 6.457873509412178], [...], ...]
       },
       ...

Upload your predictions without "features" fields!