Skip to content


[ allennlp.interpret.saliency_interpreters.saliency_interpreter ]


class SaliencyInterpreter(Registrable):
 | def __init__(self, predictor: Predictor) -> None

A SaliencyInterpreter interprets an AllenNLP Predictor's outputs by assigning a saliency score to each input token.


class SaliencyInterpreter(Registrable):
 | ...
 | def saliency_interpret_from_json(self, inputs: JsonDict) -> JsonDict

This function finds saliency values for each input token.


  • inputs : JsonDict
    The input you want to interpret (the same as the argument to a Predictor, e.g., predict_json()).


  • interpretation : JsonDict
    Contains the normalized saliency values for each input token. The dict has entries for each instance in the inputs JsonDict, e.g., {instance_1: ..., instance_2:, ... }. Each one of those entries has entries for the saliency of the inputs, e.g., {grad_input_1: ..., grad_input_2: ... }.