Computing system for classifying tax effective date
A computing system is provided, including a processor configured to, at inference time, input an article to a first trained machine learning (ML) model. The processor is further configured to extract, via the first trained ML model, candidate dates from the article. The processor is further configured to input the extracted candidate dates to a second trained ML model, and classify, via the second trained ML model, each extracted candidate date into one of plurality of classifications including a tax effective date and a date other than a tax effective date. The processor is further configured to create a modified article that highlights the candidate date classified as the tax effective date and the candidate date classified as the date other than the tax effective date, in which the tax effective date and the date other than the tax effective date are highlighted in a visually distinguishable manner.
1 . A computing system, comprising:
a processor configured to:
at inference time,
input an article to a first trained machine learning (ML) model;
extract, via the first trained ML model, candidate dates from the article;
input the extracted candidate dates to a second trained ML model; and
classify, via the second trained ML model, each extracted candidate date into one of plurality of classifications including a tax effective date, a date other than a tax effective date, and a not-a-date, wherein
the first trained ML model has been trained on a first training data set including first training articles to identify the candidate dates,
the second trained ML model has been trained on a second training data set including second training articles that include a first set of dates that are labeled with a tax effective date classification and a second set of dates that are labeled with a date other than a tax effective date classification, and that also include text that is not-a-date, to classify the candidate dates into one of the plurality of classifications including the tax effective date classification, the date other than a tax effective date classification, and a not-a-date classification,
the first trained ML model is a named entity recognition model,
the second trained ML model is a neural network model that utilizes pretrained contextual embedding representations of words and entities,
the tax effective date, the date other than the tax effective date, and the not-a-date are highlighted in a different manner so as to be visually distinguishable from each other, and
the second training articles of the second training data set are divided into sentences by a sentence divider and processed by the second trained ML model one sentence at a time, each sentence being included together with start and end offsets of the candidate dates for classification.
2 . The computing system of claim 1 , wherein the processor is further configured to create a modified article that highlights the candidate date classified as the tax effective date.
3 . The computing system of claim 2 , wherein the processor is further configured to create a modified article that highlights the candidate date classified as the date other than the tax effective date.
4 . The computing system of claim 1 , wherein the processor is further configured to:
at training time,
receive the first training data set including the first training articles that include training dates;
input the first training data set to a first ML model; and
train the first ML model to extract the candidate dates from the first training articles, to thereby generate the first trained ML model.
5 . The computing system of claim 4 , wherein the first trained ML model is a spaCy NER model.
6 . The computing system of claim 4 , wherein the processor is further configured to:
at training time,
receive the second training data set including the second training articles that include the first set of dates that are labeled with the tax effective date classification and the second set of dates that are labeled with the date other than a tax effective date classification;
input the second training data set to a second ML model; and
train the second ML model to classify each extracted candidate date into one of plurality of classifications including the tax effective date and the date other than a tax effective date, to thereby generate the trained second ML model.
7 . The computing system of claim 6 , wherein the second ML model is a LUKE neural network model.
8 . The computing system of claim 1 , wherein the processor is further configured to:
input the article to a regular expression engine; and
extract, via the regular expression engine, candidate tax rates and/or tax amounts from the article.
9 . The computing system of claim 8 , wherein the processor is further configured to create a modified article that highlights the extracted candidate tax rates and the extracted candidate tax amounts.
10 . A computerized method, comprising:
at inference time,
inputting an article to a first trained machine learning (ML) model;
extracting, via the first trained ML model, candidate dates from the article;
inputting the extracted candidate dates to a second trained ML model; and
classifying, via the second trained ML model, each extracted candidate date into one of plurality of classifications including a tax effective date, a date other than a tax effective date, and a not-a-date, wherein
the first trained ML model has been trained on a first training data set including first training articles to identify the candidate dates,
the second trained ML model has been trained on a second training data set including second training articles that include a first set of dates that are labeled with a tax effective date classification and a second set of dates that are labeled with a date other than a tax effective date classification, and that also include text that is not-a-date, to classify the candidate dates into one of the plurality of classifications including the tax effective date classification, the date other than a tax effective date classification, and a not-a-date classification,
the first trained ML model is a named entity recognition model,
the second trained ML model is a neural network model that utilizes pretrained contextual embedding representations of words and entities,
the tax effective date, the date other than the tax effective date, and the not-a-date are highlighted in a different manner so as to be visually distinguishable from each other,
the second training articles of the second training data set are divided into sentences by a sentence divider and processed by the second trained ML model one sentence at a time, each sentence being included together with start and end offsets of the candidate dates for classification.
11 . The computerized method of claim 10 , further comprising:
creating a modified article that highlights the candidate date classified as the tax effective date.
12 . The computerized method of claim 11 , further comprising:
creating a modified article that highlights the candidate date classified as the date other than a tax effective date.
13 . The computerized method of claim 12 , further comprising:
at training time,
receiving the first training data set including the first training articles that include training dates;
inputting the first training data set to a first ML model;
training the first ML model to extract the candidate dates from the first training articles; and
generating the first trained ML model.
14 . The computerized method of claim 10 , further comprising:
at training time,
receiving the second training data set including the second training articles that include the first set of dates that are labeled with the tax effective date classification and the second set of dates that are labeled with the date other than a tax effective date classification;
inputting the second training data set to a second ML model;
training the second ML model to classify each extracted candidate date into one of plurality of classifications including the tax effective date and the date other than a tax effective date; and
generating the trained second ML model.
15 . The computerized method of claim 10 , further comprising:
inputting the article to a regular expression engine; and
extracting, via the regular expression engine, candidate tax rates and/or tax amounts from the article.
16 . The computerized method of claim 15 , further comprising:
creating a modified article that highlights the extracted candidate tax rates and the extracted candidate tax amounts.
17 . A computing system, comprising:
a processor configured to:
at inference time,
input an article to a first trained machine learning (ML) model;
extract, via the first trained ML model, candidate dates from the article;
input the extracted candidate dates to a second trained ML model;
classify, via the second trained ML model, each extracted candidate date into one of plurality of classifications including a tax effective date, a date other than a tax effective date, and a not-a-date;
at training time,
receive a first training data set including first training articles that include dates;
input the first training data set to a first ML model;
train the first ML model on the first training data set to identify the candidate dates from the first training articles, to thereby generate the first trained ML model;
receive a second training data set including second training articles that include a first set of dates that are labeled with a tax effective date classification and a second set of dates that are labeled with a date other than a tax effective date classification, and that also include text that is not-a-date;
input the second training data set to a second ML model; and
train the second ML model on the second training data set to classify each extracted candidate date into one of plurality of classifications including the tax effective date, the date other than a tax effective date, and a not-a-date classification, to thereby generate the trained second ML model, wherein
the first trained ML model is a named entity recognition model,
the second trained ML model is a neural network model that utilizes pretrained contextual embedding representations of words and entities,
the tax effective date, the date other than the tax effective date, and the not-a-date are highlighted in a different manner so as to be visually distinguishable from each other,
the second training articles of the second training data set are divided into sentences by a sentence divider and processed by the second trained ML model one sentence at a time, each sentence being included together with start and end offsets of the candidate dates for classification.