Classifying business summaries against a hierarchical industry classification structure using supervised machine learning
A classification system is provided for classifying text-based business summaries, referred to herein as “summaries,” against a hierarchical industry classification structure. The classification system includes a word-based sub classifier that uses a neural network to generate a vector space for each summary in a training set, where each summary in the training set is known to correspond to a particular industry classification in the hierarchical industry classification structure. Weight values in the hidden layer of a neural network used by the word-based sub classifier are changed to improve the predictive capabilities of the neural network in the business summary classification context. Embodiments include increasing representation in the training set for underrepresented parent industry classifications and attributes of the hierarchical industry classification structure, such as distances between industry classifications and whether industry classifications are in the same subgraph. The completion of training of the word-based sub classifier is based upon whether a performance metric, such as an hF 1 score, satisfies one or more early stopping criteria. The classification system also includes a category-based sub classifier and a meta classifier.
1. A computing apparatus comprising:
one or more processors;
one or more memories; and
a word-based sub classifier configured to generate a first probability distribution for a text-based business summary across a plurality of industry classifications in a hierarchical industry classification structure, wherein:
the word-based sub classifier is trained on a training set comprising a plurality of text-based business summaries, wherein each text-based business summary, from the plurality of text-based business summaries, has a known correspondence to an industry classification, from the plurality of industry classifications, in the hierarchical industry classification structure, and
training of the word-based sub classifier is completed when an evaluation metric satisfies one or more early stopping criteria,
a category-based sub classifier configured to generate a second probability distribution for the text-based business summary across the plurality of industry classifications in the hierarchical industry classification structure,
a meta classifier configured to determine a predicted probability distribution for the text-based business summary across the plurality of industry classifications in the hierarchical industry classification structure based upon the first probability distribution generated by the word-based sub classifier and second probability distribution generated by the category-based sub classifier.
2. The computing apparatus as recited in claim 1 , wherein the word-based sub classifier is further configured to use a neural network to determine a vector representation for a particular text-based business summary from the plurality of text-based business summaries by:
determining a vector representation for each word in the particular text-based business summary, and
determining an average vector representation based upon the vector representations for each word in the particular text-based business summary.
3. The computing apparatus as recited in claim 1 , wherein training the word-based sub classifier includes updating weight values in a hidden layer of a neural network used by the word-based sub classifier to minimize a loss function, where a loss in the loss function represents a difference between an estimated probability and a known probability that a particular text-based business summary, from the plurality of text-based business summaries in the training set, corresponds to a particular industry classification in the hierarchical industry classification structure.
4. The computing apparatus as recited in claim 3 , wherein updating weight values in a hidden layer of a neural network used by the word-based sub classifier to minimize a loss function includes, for a particular industry classification in the hierarchical industry classification, increasing representation of a parent industry classification to the particular industry classification in the hierarchical industry classification.
5. The computing apparatus as recited in claim 4 , wherein the word-based sub classifier is further configured to revise a value that corresponds to the parent industry classification in a one-hot vector representation for the particular text-based validation summary.
6. The computing apparatus as recited in claim 1 , wherein:
the evaluation metric is an hF 1 score calculated based upon results of processing a set of validation summaries using the word-based sub classifier.
7. The computing apparatus as recited in claim 1 , wherein the predicted probability distribution for the text-based business summary across the plurality of industry classifications in the hierarchical industry classification structure is a geometric mean of the first probability distribution generated by the word-based sub classifier and the second probability distribution generated by the category-based sub classifier.
8. One or more non-transitory computer-readable media storing instructions which, when processed by one or more processors, cause:
a word-based sub classifier to generate a first probability distribution for a text-based business summary across a plurality of industry classifications in a hierarchical industry classification structure, wherein:
the word-based sub classifier is trained on a training set comprising a plurality of text-based business summaries, wherein each text-based business summary, from the plurality of text-based business summaries, has a known correspondence to an industry classification, from the plurality of industry classifications, in the hierarchical industry classification structure, and
training of the word-based sub classifier is completed when an evaluation metric satisfies one or more early stopping criteria,
a category-based sub classifier to generate a second probability distribution for the text-based business summary across the plurality of industry classifications in the hierarchical industry classification structure,
a meta classifier to determine a predicted probability distribution for the text-based business summary across the plurality of industry classifications in the hierarchical industry classification structure based upon the first probability distribution generated by the word-based sub classifier and second probability distribution generated by the category-based sub classifier.
9. The one or more non-transitory computer-readable media as recited in claim 8 , further comprising additional instructions which, when processed by the one or more processors, cause the word-based sub classifier to use a neural network to determine a vector representation for a particular text-based business summary from the plurality of text-based business summaries by:
determining a vector representation for each word in the particular text-based business summary, and
determining an average vector representation based upon the vector representations for each word in the particular text-based business summary.
10. The one or more non-transitory computer-readable media as recited in claim 8 , wherein training the word-based sub classifier includes updating weight values in a hidden layer of a neural network used by the word-based sub classifier to minimize a loss function, where a loss in the loss function represents a difference between an estimated probability and a known probability that a particular text-based business summary, from the plurality of text-based business summaries in the training set, corresponds to a particular industry classification in the hierarchical industry classification structure.
11. The one or more non-transitory computer-readable media as recited in claim 10 , wherein updating weight values in a hidden layer of a neural network used by the word-based sub classifier to minimize a loss function includes, for a particular industry classification in the hierarchical industry classification, increasing representation of a parent industry classification to the particular industry classification in the hierarchical industry classification.
12. The one or more non-transitory computer-readable media as recited in claim 11 , further comprising additional instructions which, when processed by the one or more processors, cause the word-based sub classifier to revise a value that corresponds to the parent industry classification in a one-hot vector representation for the particular text-based validation summary.
13. The one or more non-transitory computer-readable media as recited in claim 8 , wherein:
the evaluation metric is an hF 1 score calculated based upon results of processing a set of validation summaries using the word-based sub classifier.
14. The one or more non-transitory computer-readable media as recited in claim 8 , wherein the predicted probability distribution for the text-based business summary across the plurality of industry classifications in the hierarchical industry classification structure is a geometric mean of the first probability distribution generated by the word-based sub classifier and the second probability distribution generated by the category-based sub classifier.
15. A computer-implemented method comprising:
a word-based sub classifier generating a first probability distribution for a text-based business summary across a plurality of industry classifications in a hierarchical industry classification structure, wherein:
the word-based sub classifier is trained on a training set comprising a plurality of text-based business summaries, wherein each text-based business summary, from the plurality of text-based business summaries, has a known correspondence to an industry classification, from the plurality of industry classifications, in the hierarchical industry classification structure, and
training of the word-based sub classifier is completed when an evaluation metric satisfies one or more early stopping criteria,
a category-based sub classifier generating a second probability distribution for the text-based business summary across the plurality of industry classifications in the hierarchical industry classification structure,
a meta classifier determining a predicted probability distribution for the text-based business summary across the plurality of industry classifications in the hierarchical industry classification structure based upon the first probability distribution generated by the word-based sub classifier and second probability distribution generated by the category-based sub classifier.
16. The computer-implemented method as recited in claim 15 , further comprising the word-based sub classifier using a neural network to determine a vector representation for a particular text-based business summary from the plurality of text-based business summaries by:
determining a vector representation for each word in the particular text-based business summary, and
determining an average vector representation based upon the vector representations for each word in the particular text-based business summary.
17. The computer-implemented method as recited in claim 15 , wherein training the word-based sub classifier includes updating weight values in a hidden layer of a neural network used by the word-based sub classifier to minimize a loss function, where a loss in the loss function represents a difference between an estimated probability and a known probability that a particular text-based business summary, from the plurality of text-based business summaries in the training set, corresponds to a particular industry classification in the hierarchical industry classification structure.
18. The computer-implemented method as recited in claim 17 , wherein updating weight values in a hidden layer of a neural network used by the word-based sub classifier to minimize a loss function includes, for a particular industry classification in the hierarchical industry classification, increasing representation of a parent industry classification to the particular industry classification in the hierarchical industry classification.
19. The computer-implemented method as recited in claim 18 , further comprising the word-based sub classifier revising a value that corresponds to the parent industry classification in a one-hot vector representation for the particular text-based validation summary.
20. The computer-implemented method as recited in claim 15 , wherein:
the evaluation metric is an hF 1 score calculated based upon results of processing a set of validation summaries using the word-based sub classifier.
21. The computer-implemented method as recited in claim 15 , wherein the predicted probability distribution for the text-based business summary across the plurality of industry classifications in the hierarchical industry classification structure is a geometric mean of the first probability distribution generated by the word-based sub classifier and the second probability distribution generated by the category-based sub classifier.