classification.md

classification task-type

back to main README

For the classification task, the system reads one instance per line. Similarly as for word-level tasks, you have to define a sent_idxs for each dataset, and a column_idx for each task. If for example you have a file with labels <TAB> sentence, sent_idxs should be 1 and column_idx=0. It should be noted that you can use a list for the sent_idxs, if you want to use multiple input utterances (they will be concatenated internally, with a SEP token inbetween).

{
    "RTE": {
        "train_data_path": "data/glue/RTE.train",
        "dev_data_path": "data/glue/RTE.dev",
        "sent_idxs": [0,1],
        "tasks": {
            "rte": {
                "column_idx": 2,
                "task_type": "classification"
            }
        }
    }
}

Joint sequence labeling and classification within dataset

We also support classification of labels inside a conll-like format. In this situation, the name of the task should match the label used in the comment, and the column_idx should be set to -1.

If your data looks like this:

# text: tell me the weather report for half moon bay
# intent: weather/find
1	tell	_	NoLabel
2	me	_	NoLabel
3	the	_	NoLabel
4	weather	_	B-weather/noun
5	report	_	I-weather/noun
6	for	_	NoLabel
7	half	_	B-location
8	moon	_	I-location
9	bay	_	I-location

The dataset configuration should be (note that the task name `intent' should match with line 2 of the data file):

{
    "NLU": {
        "train_data_path": "data/nlu/en/train-en.conllu",
        "dev_data_path": "data/nlu/en/eval-en.conllu",
        "word_idx": 1,
        "tasks": {
            "slots": {
                "task_type": "seq",
                "column_idx": 3
            },
            "intent": {
                "task_type": "classification",
                "column_idx": -1
            }
        }
    }
}