Learning rules and dictionaries with neuro-symbolic artificial intelligence
A system, computer program product, and method are provided for jointly learning dictionary based rules and dictionary candidates. Natural language text is received and parsed into subsets, with the subset being subjected to natural language processing to identify one or more verbs within the subset. The identified verbs are evaluated with respect to a dictionary and one or more rules. The evaluation is directed at each predicate in the rules with respect to the identified verbs. A neural network is leveraged to jointly induce modification of the rules and one or more dictionaries responsive to the evaluation.
1. A computer system comprising:
a processor operatively coupled to memory;
an artificial intelligence (AI) platform, in communication with the processor, the AI platform comprising:
a text module configured to receive natural language text, and parse the received text into a subset;
subject the subset to natural language processing, and identify one or more verbs within the subset;
an evaluator, operatively coupled to the NLP module, configured to evaluate the identified one or more verbs with respect to at least one dictionary and one or more rules, including evaluating each predicate in the one or more rules with respect to the identified one or more verbs;
an artificial neural network (ANN) configured to:
process a loss function generated in response to the evaluating, wherein the loss function is utilized by a machine learning module to interface with one or more dictionary and one or more rules; and
compute one or more gradients from the loss function, wherein the loss function is utilized as feedback by the ANN, and wherein a gradient descent of the one or more gradients is an optimization algorithm used to train the ANN;
a machine learning (ML) module, operatively coupled to the evaluator, configured to leverage the ANN to jointly induce modification of the one or more rules and the at least one dictionary responsive to the computation of the one or more gradient descents; and
utilizing, by a machine learning (ML) module, the loss function to selectively and jointly amend the at least one dictionary and the one or more rules based on the computed one or more gradients.
2. The computer system of claim 1 , wherein evaluation of the identified one or more verbs with respect to the at least one dictionary and the one or more rules includes the evaluator to assess a score for each of the identified one or more verbs across the one or more rules, and select a maximum score as a subset score for the subset, wherein the score is parameters that are updated through backpropagation.
3. The computer system of claim 2 , wherein the subset has one or more sentences, and further comprising the evaluator to identify a ground truth label for each sentence in the subset, compare the ground truth label with the subset score, the comparison to generate a loss function.
4. The computer system of claim 3 , further comprising the ML module configured to selectively amend the at least one dictionary and the one or more rules responsive to the computed one or more gradients.
5. The computer system of claim 4 , further comprising the ML module configured to back-propagate the computed one or more gradients through the neural network responsive to the selective dictionary and rule amendment.
6. The computer system of claim 1 , wherein the leverage of the neural network further compromises the neural network to process a loss function in response to the evaluation, and to compute one or more gradients from the loss function.
7. The computer system of claim 6 , wherein the jointly induced modification of the one or more rules and the at least one dictionary is in response to the computed one or more gradients.
8. A computer program product comprising a computer readable storage medium having program code embodied therewith that, when executed by at least one processor, cause a computing device to:
receive natural language text, and parse the received text into a subset;
subject the subset to natural language processing, and utilize a machine-learning model to identify one or more verbs within the subset;
evaluating the identified one or more verbs with respect to at least one and one or more rules, including evaluating each predicate in the one or more rules with respect to the identified one or more verbs; and
leveraging an artificial neural network (ANN) to:
process a loss function generated in response to the evaluating, wherein the loss function is utilized by a machine learning module to interface with one or more dictionary and one or more rules; and
compute one or more gradient descents from the loss function, wherein the loss function is utilized as feedback to train the ANN, and wherein the gradient descent is an optimization algorithm used to train the ANN;
jointly inducing modification of the one or more rules and the at least one dictionary responsive to the computation of the one or more gradient descents; and
utilizing, by a machine learning (ML) module, the loss function to selectively and jointly amend the at least one dictionary and the one or more rules based on the computed one or more gradients.
9. The computer program product of claim 8 , wherein the program code configured to evaluate the identified one or more verbs with respect to the at least one dictionary and the one or more rules includes assessing a score for each of the identified one or more verbs across the one or more rules, and selecting a maximum score as a subset score for the subset.
10. The computer program product of claim 9 , wherein the subset has one or more sentences, and further comprising program code configured to identify a ground truth label for each sentence in the subset, compare the ground truth label with the subset score, the comparison to generate the loss function.
11. The computer program product of claim 10 , further comprising program code configured to selectively amend the at least one dictionary and the one or more rules responsive to the computed one or more gradients.
12. The computer program product of claim 11 , further comprising program code configured to back-propagate the computed one or more gradients through the neural network responsive to the selective dictionary and rule amendment.
13. A computer implemented method, for learning both rules and dictionaries, the computer-implemented method comprising:
receive natural language text, and parse the received text into a subset;
subject the subset to natural language processing, and utilize a machine-learning model to identify one or more verbs within the subset;
evaluating the identified one or more verbs with respect to at least one dictionary and one or more rules, including evaluating each predicate in the one or more rules with respect to the identified one or more verbs; and
leveraging an artificial neural network (ANN) to:
process a loss function generated in response to the evaluating, wherein the loss function is utilized by a machine learning module to interface with one or more dictionary and one or more rules; and
compute one or more gradient descents from the loss function, wherein the loss function is utilized as feedback to train the ANN, and wherein the gradient descent of the one or more gradients is an optimization algorithm used to train the ANN;
jointly inducing modification of the one or more rules and the at least one dictionary responsive to the computation of the one or more gradients; and
utilizing, by a machine learning (ML) module, the loss function to selectively and jointly amend the at least one dictionary and the one or more rules based on the computed one or more gradients.
14. The method of claim 13 , wherein evaluating the identified one or more verbs with respect to the at least one dictionary and the one or more rules includes assessing a score for each of the identified one or more verbs across the one or more rules, and selecting a maximum score as a subset score for the subset.
15. The method of claim 14 , wherein the subset has one or more sentences, and further comprising identifying a ground truth label for each sentence in the subset, comparing the ground truth label with the subset score, the comparison generating the loss function.
16. The method of claim 15 , further comprising selectively amending the at least one dictionary and the one or more rules responsive to the computed one or more gradients.
17. The method of claim 16 , further comprising back-propagating the computed one or more gradients through the neural network responsive to the selective dictionary and rule amendment.
18. The method of claim 17 , wherein the selective amendment further comprises updating one or more scores, the scores including a rule score, a predicate score, and a word score.