Systems and methods for machine learning key-value extraction on documents
A machine learning based key-value extraction model extracts fields/entities from documents. The input images are processed through OCR. A list of words (uni-grams) and their coordinates are extracted from the original images. Following word cleaning and manipulation, n-gram creation (multi-words), and feature engineering, the transformed data is fed into a classification algorithm to predict if a uni-gram or n-gram is one of the target entities or a non-entity. Following the first step that includes unique feature engineering, a second step improves extraction accuracy among the fields/entities.
1. A method to train a machine-learning model to extract key-values from documents, the method comprising:
receiving a collection of training document images;
creating a training data set from the collection; and
training a classification model using the training data set, wherein training the classification model further comprises applying at least a plurality of features to the training data set, the plurality of features comprising:
a first feature corresponding to a location of an n-gram,
a second feature corresponding to a plurality of letter cases,
a third feature corresponding to textual character type,
a fourth feature corresponding to a regular expression,
a fifth feature corresponding to a number of characters used to distinguish between different entities among target classes, and
a sixth feature corresponding to punctuation.
2. The method of claim 1 , wherein the first feature corresponding to the location of the n-gram uses a graph-like data structure.
3. The method of claim 2 , wherein the graph-like data structure indicates the nearest words neighboring the location of the n-gram.
4. The method as in one of claims 1-3 , wherein the second feature corresponding to letter case indicates numerical and non-numerical character types, wherein the plurality of letter cases includes upper case(s) and lower case(s).
5. The method of claim 1 , further comprising:
applying optical character recognition to a document image, wherein applying optical character recognition outputs a plurality of document objects;
applying threshold values to priors of the plurality of document objects;
determining, from the plurality of document objects, a subset of document objects based at least in part on some of the priors of the plurality of document objects not satisfying the threshold values; and
applying the trained classification model to the subset of document objects.
6. The method of claim 1 , further comprising:
applying optical character recognition to a document image, wherein applying optical character recognition outputs a plurality of document objects; and
applying the trained classification model to the plurality of document objects, wherein applying the trained classification model further comprises:
outputting entities and non-entities with corresponding probabilities indicating a likelihood of correctness for each entity and non-entity.
7. The method of claim 1 , wherein the sixth feature corresponding to punctuation is based on a unique punctuation characters that are observable within an entity.
8. A system comprising:
a data store configured to store computer-executable instructions; and
a hardware processor in communication with the data store, the hardware processor, when executing the computer-executable instructions, configured to:
receive a collection of training document images;
create a training data set from the collection; and
train a classification model using the training data set, wherein training the classification model further comprises applying at least a plurality of features to the training data set, the plurality of features comprising:
a first feature corresponding to a location of an n-gram,
a second feature corresponding to a plurality of letter cases,
a third feature corresponding to textual character type,
a fourth feature corresponding to a regular expression,
a fifth feature corresponding to a number of characters used to distinguish between different entities among target classes, and
a sixth feature corresponding to punctuation.
9. The system of claim 8 , wherein the first feature corresponding to the location of the n-gram uses a graph-like data structure.
10. The system of claim 9 , wherein the graph-like data structure indicates the nearest words neighboring the location of the n-gram.
11. The system as in one of claims 8-10 , wherein the second feature corresponding to letter case indicates numerical and non-numerical character types, wherein the plurality of letter cases includes upper case(s) and lower case(s).
12. The system of claim 7 , wherein the hardware processor is further configured to execute further computer-executable instructions to:
apply optical character recognition to a document image, wherein applying optical character recognition outputs a plurality of document objects; and
apply the trained classification model to the plurality of document objects, wherein applying the trained classification model further comprises:
outputting entities and non-entities with corresponding probabilities indicating a likelihood of correctness for each entity and non-entity.
13. A system comprising:
a data store configured to store computer-executable instructions; and
a hardware processor in communication with the data store, the computing device, when executing the computer-executable instructions, configured to:
receive a collection of training document images;
create a training data set from the collection; and
train a classification model using the training data set, wherein training the classification model further comprises applying at least a plurality of features to the training data set, the plurality of features comprising:
a first feature corresponding to a location of an n-gram,
a second feature corresponding to letter case,
a third feature corresponding to textual character type,
a fourth feature corresponding to a regular expression,
a fifth feature corresponding to a number of characters, and
a sixth feature corresponding to punctuation,
wherein the hardware processor is further configured to execute further computer-executable instructions to:
apply optical character recognition to a document image, wherein applying optical character recognition outputs a plurality of document objects;
apply threshold values to priors of the plurality of document objects;
determine, from the plurality of document objects, a subset of document objects based at least in part on some of the priors of the plurality of document objects not satisfying the threshold values; and
apply the classification model to the subset of document objects.