Skip to content

Latest commit

 

History

History
183 lines (158 loc) · 5.58 KB

add_custom_features.md

File metadata and controls

183 lines (158 loc) · 5.58 KB
Raw

Add custom features for custom analysis

If you want to perform custom analysis with your custom features that are not supported in the original task processors, and these features are only related to one particular system/dataset instead of the task itself, you can define the custom features and analysis in the metadata section in the output JSON file.

Example of bucket analysis

Here is the output json format for bucket analysis with custom features. Discrete features should have "dtype": "string", such as the subject of the sentence, etc. Continuous features should have "dtype": "float", such as the count of particular words, the output logits/probability, etc.

{
  "metadata": {
    "custom_features": {
      "feature-level": {
        "discrete-custom-feature-name": {
          "cls_name": "Value",
          "dtype": "string",
          "description": "(optional) description of the feature"
        },
        "continuous-custom-feature-name": {
          "cls_name": "Value",
          "dtype": "float",
          "description": "(optional) description of the feature"
        }
      }
    },
    "custom_analyses": [
      {
        "cls_name": "BucketAnalysis",
        "level": "feature-level",
        "feature": "discrete-custom-feature-name",
        "num_buckets": 15,
        "method": "discrete",
        "sample_limit": 50
      },
      {
        "cls_name": "BucketAnalysis",
        "level": "feature-level",
        "feature": "continuous-custom-feature-name",
        "num_buckets": 15,
        "method": "continuous",
        "sample_limit": 50
      }
    ]
  },
  "examples": [
    {
      "predicted_label": "true",
      "discrete-custom-feature-name": "discrete-feature-value",
      "continuous-custom-feature-name": "continuous-feature-value"
    }
  ]
}

where

  • feature-level: the fine-grained level of the analysis
    • example for sentence-level analysis
    • span for span-level analysis (e.g. in named entity recognition analysis)
    • token for token-level analysis (e.g. in named entity conditional text generation)
  • num_bucket: the number of buckets to be used
  • sample_limit: if the number of examples are larger than sample_limit, randomly select sample_limit examples for analysis

Example of ComboCountAnalysis

ComboCountAnalysis is used to count feature combinations (e.g. for confusion matrices). It will return counts of each combination of values for the features named in features.

{
  "metadata": {
    "custom_analyses": [
      {
        "cls_name": "ComboCountAnalysis",
        "level": "feature-level",
        "description": "(optional) description of the analysis",
        "features": ["feature-name-1", "feature-name-2"]
      }
    ]
  },
  "examples": [
    {
      "predicted_label": "true",
      "custom-feature-name": "feature-value"
    }
  ]
}

where

  • feature-level: the fine-grained level of the analysis
    • example for sentence-level analysis
    • span for span-level analysis (e.g. in named entity recognition analysis)
    • token for token-level analysis (e.g. in named entity conditional text generation)
  • features: a list of feature names where each feature are predefined. These features can be true_label, predicted_label, or other custom features.

Example of CalibrationAnalysis

CalibrationAnalysis measures calibration, the association between a model-predicted probability and the likelihood of the answer being correct. It must be used on tasks that use the Accuracy metric. For tasks that use the Accuracy metric, by default, calibration analysis is automatically performed if your model inputs have the confidence feature. You can directly add the confidence values to the examples in a json-formatted system output file. You may also customize calibration analysis with custom feature name and settings.

Default calibration analysis format example:

{
  "examples": [
    {
      "predicted_label": "positive",
      "confidence": 0.22101897026820283
    }
  ]
}

Custom calibration analysis format example:

{
  "metadata": {
    "custom_features": {
      "example": {
        "probability": {
          "cls_name": "Value",
          "dtype": "float",
          "description": "model-predicted probability"
        }
      }
    },
    "custom_analyses": [
      {
        "cls_name": "CalibrationAnalysis",
        "level": "example",
        "feature": "probability",
        "num_buckets": 5,
        "sample_limit": 50
      }
    ]
  },
  "examples": [
    {
      "predicted_label": "positive",
      "probability": 0.22101897026820283
    }
  ]
}

where

  • feature: the customized name of the confidence feature
  • num_bucket: the number of buckets of the same size between interval [0, 1]

Example output json files

Note

When running analysis with SDK command, --custom-dataset-file-type and --output-file-type json are required.

Example command:

explainaboard --task text-classification --custom-dataset-paths path_to_my_custom_data.tsv --system-outputs path_to_my_output_with_custom_features.json --report-json path_to_my_report.json --custom-dataset-file-type tsv --output-file-type json