Hierarchical multi-armed bandit for lexicon expansion
View Patent ↗Disclosed herein are methods, systems, and computer program products for selecting an artificial intelligence (AI) model. Aspects include receiving, by a multi-armed bandit agent, candidate entities from multiple set expansion models and selecting a subset of the candidate entities for a first set expansion task, wherein a first candidate entity of the selected subset of candidate entities is selected from a first model by a subject matter expert. Aspects also include selecting, by the multi-armed bandit agent, based on the first candidate entity selected by the subject matter expert, the first model from the set expansion models for generating further candidate entities.
1 . A computer-implemented method for selecting an artificial intelligence (AI) model, the computer-implemented method comprising:
receiving, by a multi-armed bandit agent, candidate entities from multiple set expansion models, wherein the multiple set expansion models comprise a first AI model and a second AI model previously trained on different datasets;
selecting a subset of the candidate entities for a first set expansion task, wherein a first candidate entity of the selected subset of candidate entities is selected from the first AI model by a subject matter expert;
selecting, by the multi-armed bandit agent, based on the first candidate entity selected by the subject matter expert, the first AI model from the set expansion models for generating further candidate entities;
in response to the first candidate entity from the first AI model being selected, selecting, without having knowledge of the different datasets and while avoiding knowledge of the training distribution of the different datasets, the second AI model configured to output the first candidate entity, wherein both the first AI model and the second AI model are configured to output the first candidate entity based on receiving input of a set of positive inputs and a set of negative inputs such that a reward is given to the first AI model and the second AI model for generating the first candidate entity, wherein the set of positive inputs and the set of negative inputs are input to the first AI model and the second AI model after the first AI model and the second AI model have been previously trained on the different datasets; and
causing the first AI model and the second AI model selected to be executed for generating an expansion to a lexicon for natural language processing (NLP).
2 . The computer-implemented method of claim 1 , wherein the NLP is executed for at least one of a query understanding and a question answering.
3 . The computer-implemented method of claim 1 , wherein the first AI model outputs a first list comprising the first candidate entity and the second AI model outputs a second list comprising the first candidate entity.
4 . The computer-implemented method of claim 1 , wherein to alleviate a cold- start problem with the multi-armed bandit agent, frequency-based boosting (Marble-B) is used to generate initial candidate entities from the expansion models.
5 . The computer-implemented method of claim 4 , wherein for the frequency-based boosting:
all AI models of the expansion models are executed in order to generate the initial candidate entities; and
in response to executing the all AI models of the expansion to generate the initial candidate entities, the initial candidate entities from the all AI models are captured and one or more most frequent candidate entities from the initial candidate entities are presented to the subject matter expert.
6 . The computer-implemented method of claim 1 , wherein to alleviate a cold- start problem with the multi-armed bandit agent, history-based boosting (Marble-H) is used to generate initial candidate entities from the expansion models.
7 . The computer-implemented method of claim 6 , wherein for the history-based boosting, in response to the subject matter expert selecting, at time t, the first candidate entity from the first AI model from the expansion models, the first AI model is rewarded and further, a second model that also generates the first candidate entity at time t+1, is also rewarded.
8 . A system comprising:
a memory; and
one or more processing units coupled with the memory, the one or more processing units configured to perform a method for selecting an artificial intelligence (AI) model, the method comprising:
receiving, by a multi-armed bandit agent, candidate entities from multiple set expansion models, wherein the multiple set expansion models comprise a first AI model and a second AI model previously trained on different datasets;
selecting a subset of the candidate entities for a first set expansion task, wherein a first candidate entity of the selected subset of candidate entities is selected from the first AI model by a subject matter expert;
selecting, by the multi-armed bandit agent, based on the first candidate entity selected by the subject matter expert, the first AI model from the set expansion models for generating further candidate entities;
in response to the first candidate entity from the first AI model being selected, selecting, without having knowledge of the different datasets and while avoiding knowledge of the training distribution of the different datasets, the second AI model; configured to output the first candidate entity, wherein both the first AI model and the second AI model are configured to output the first candidate entity based on receiving input of a set of positive inputs and a set of negative inputs such that a reward is given to the first AI model and the second AI model for generating the first candidate entity, wherein the set of positive inputs and the set of negative inputs are input to the first AI model and the second AI model after the first AI model and the second AI model have been previously trained on the different datasets; and
causing the first AI model and the second AI model selected to be executed for generating an expansion to a lexicon for natural language processing (NLP).
9 . The system of claim 8 , wherein the NLP is executed for at least one of a query understanding and a question answering.
10 . The system of claim 8 , wherein the first AI model outputs a first list comprising the first candidate entity and the second AI model outputs a second list comprising the first candidate entity.
11 . The system of claim 8 , wherein to alleviate a cold-start problem with the multi-armed bandit agent, frequency-based boosting (Marble-B) is used to generate initial candidate entities from the expansion models.
12 . The system of claim 11 , wherein for the frequency-based boosting:
all AI models of the expansion models are executed in order to generate the initial candidate entities; and
in response to executing the all AI models of the expansion to generate the initial candidate entities, the initial candidate entities from the all AI models are captured and one or more most frequent candidate entities from the initial candidate entities are presented to the subject matter expert.
13 . The system of claim 8 , wherein to alleviate a cold-start problem with the multi-armed bandit agent, history-based boosting (Marble-H) is used to generate initial candidate entities from the expansion models.
14 . The system of claim 13 , wherein for the history-based boosting, in response to the subject matter expert selecting, at time t, the first candidate entity from the first AI model from the expansion models, the first AI model is rewarded and further, a second model that also generates the first candidate entity at time t+1, is also rewarded.
15 . A computer program product comprising a memory with one or more computer executable instructions stored thereon, which when executed by one or more processing units cause the one or more processing units to perform a method for selecting an artificial intelligence (AI) model, the method comprising:
receiving, by a multi-armed bandit agent, candidate entities from multiple set expansion models, wherein the multiple set expansion models comprise a first AI model and a second AI model previously trained on different datasets;
selecting a subset of the candidate entities for a first set expansion task, wherein a first candidate entity of the selected subset of candidate entities is selected from the first AI model by a subject matter expert;
selecting, by the multi-armed bandit agent, based on the first candidate entity selected by the subject matter expert, the first AI model from the set expansion models for generating further candidate entities;
in response to the first candidate entity from the first AI model being selected, selecting, without having knowledge of the different datasets and while avoiding knowledge of the training distribution of the different datasets, the second AI model, configured to output the first candidate entity, wherein both the first AI model and the second AI model are configured to output the first candidate entity based on receiving input of a set of positive inputs and a set of negative inputs such that a reward is given to the first AI model and the second AI model for generating the first candidate entity, wherein the set of positive inputs and the set of negative inputs are input to the first AI model and the second AI model after the first AI model and the second AI model have been previously trained on the different datasets; and
causing the first AI model and the second AI model selected to be executed for generating an expansion to a lexicon for natural language processing (NLP).
16 . The computer program product of claim 15 , wherein the expansion models comprise AI models of different types; and
wherein the first AI model outputs a first list comprising the first candidate entity and the second AI model outputs a second list comprising the first candidate entity.
17 . The computer program product of claim 15 , wherein to alleviate a cold-start problem with the multi-armed bandit agent, frequency-based boosting (Marble-B) is used to generate initial candidate entities from the expansion models.
18 . The computer program product of claim 17 , wherein for the frequency-based boosting:
all AI models of the expansion models are executed in order to generate the initial candidate entities; and
in response to executing the all AI models of the expansion to generate the initial candidate entities, the initial candidate entities from the all AI models are captured and one or more most frequent candidate entities from the initial candidate entities are presented to the subject matter expert.
19 . The computer program product of claim 15 , wherein to alleviate a cold-start problem with the multi-armed bandit agent, history-based boosting (Marble-H) is used to generate initial candidate entities from the expansion models.
20 . The computer program product of claim 19 , wherein for the history-based boosting, in response to the subject matter expert selecting, at time t, the first candidate entity from the first AI model from the expansion models, the first AI model is rewarded and further, a second model that also generates the first candidate entity at time t+1, is also rewarded.