Producing explainable rules via deep learning

A computer program product for producing explainable rules via deep learning, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to perform a method comprising:

• receiving, utilizing the processor at a deep neural network (DNN) having a plurality of intermediate layers, a plurality of sentences each having an associated label;
• training the DNN, utilizing the processor and the plurality of sentences and associated labels;
• producing a linguistic expression (LE) utilizing the trained DNN and the processor, the LE having a word-based rule model that includes a plurality of human-readable conditions that are understandable by a human user,
• wherein each condition within the LE is derived from one of a plurality of intermediate layers within the DNN;
• outputting the LE in human-readable form for modification thereof by a human;
• receiving input from the human user specifying how to adjust at least one of the conditions of the word-based rule model;
• modifying the at least one condition according to the user input;
• selecting one of the sentences;
• feeding the one of the sentences as input to the DNN for generating a predicted label for the one of the sentences;
• computing an error corresponding to the predicted label compared with its correct label;
• backpropagating the computed error to update conditions included in the linguistic expressions;
• receiving an unlabeled sentence;
• and
• applying the linguistic expression with the updated conditions to the unlabeled sentence for determining a label for the unlabeled sentence in response to determining that one or more conditions of the linguistic expression with the updated conditions is not met for the unlabeled sentence, assigning a label indicating that the condition is not met is assigned to the unlabeled sentence.

A system, comprising:

• a processor;
• and
• logic integrated with the processor, executable by the processor, or integrated with and executable by the processor, the logic being configured to:
• receive, at a deep neural network (DNN), a plurality of sentences each having an associated label;
• train the DNN, utilizing the plurality of sentences and associated labels;
• produce a linguistic expression (LE) utilizing the trained DNN, wherein the linguistic expression includes a human-readable model, such that the linguistic expression includes a word-based rule that includes one or more human-readable conditions and that is understandable by a human user,
• wherein the DNN includes a plurality of intermediate layers, where each of the plurality of intermediate layers corresponds to a condition within the linguistic expression used to perform a determination of a label;
• outputting the word-based rule of the human-readable model for output in readable form to the human user;
• receiving an instruction from the human user specifying how to adjust at least one of the conditions of the word-based rule;
• modifying the at least one condition according to the instruction for updating the at least one condition;
• computing an error corresponding to a predicted label compared with its correct label;
• backpropagating the computed error to further update conditions included in the linguistic expressions;
• receiving unlabeled sentences;
• applying the linguistic expression with the updated conditions to the unlabeled sentences for determining labels for the unlabeled sentences;
• and
• in response to determining that one or more conditions of the linguistic expression is not met for one of the sentences, assigning a label to the one of the sentences indicating that the condition is not met.

A computer-implemented method, comprising:

• identifying a collection of sentences having labels;
• inputting the sentences into a deep neural network (DNN) having several layers, the DNN analyzing the sentences and producing linguistic expressions (LEs) corresponding to the sentences, each of the linguistic expression including a human-readable model having a word-based rule that includes one or more human-readable conditions;
• outputting the human-readable models of the linguistic expressions to a human expert for reading of the human-readable models by the human expert;
• receiving input from the human expert in response to outputting the human-readable models, the input being for adjusting one or more conditions of the human-readable models;
• in response to receiving the input from the human expert, modifying the linguistic expressions, wherein the modified linguistic expressions form a predictive model that, given an unlabeled sentence, predicts one or more labels for the unlabeled sentence;
• (a) choosing a subset of sentences and their corresponding labels from the collection of sentences;
• (b) for each sentence in the subset, feeding the sentence as input to the DNN and propagating it through all of the DNN's layers until its predicted label is obtained as output from the DNN, at which point the predicted label is compared with its correct label, and a corresponding error is computed;
• (c) backpropagating the computed error to update conditions included in the linguistic expressions;
• and
• repeating steps (b) and (c) for one or more additional subsets in the collection of sentences;
• receiving an unlabeled sentence;
• applying the linguistic expression with the updated conditions to the unlabeled sentence for determining a label for the unlabeled sentence;
• and
• in response to determining that one or more conditions of the linguistic expression with the updated conditions is not met for the unlabeled sentence, assigning a label to the unlabeled sentence indicating that the condition is not met.