Mediums, methods, and systems for classifying columns of a data store based on character level labeling
Exemplary embodiments pertain to new techniques for classifying or labeling organized data. A major impediment to implementing high-quality machine learning is the lack of readily accessible labeled data. In some cases, data can be classified using a classifier, but these solutions can be inaccurate and slow. Exemplary embodiments address the problem of obtaining accurate labeled data in a timely manner by applying a classifier configured to operate on character-level embeddings. Among other advantages, this can help the classifier to recognize information contained within a data unit, such as a cell of a table. The classifier may operate within the organizational structure of the data, such as by operating across a particular row or column of a table. Because data within a particular row or column is often temporally organized (e.g., transactions that are logged in chronological order), row- or column-based approaches can yield more accurate results.
1. A computer-implemented method comprising:
receiving formatted input data, the formatted input data comprising a plurality of data units organized into a plurality of organizational units;
retrieving classifiable data from a first one of the organizational units;
sending the classifiable data to a classifier, the classifier configured to perform a character-level classification and output a label from a predetermined set of labels;
receiving a label for the classifiable data from the classifier; and
assigning the label to the first one of the organizational units,
wherein the classifier comprises a convolutional neural network (CNN) configured to operate on the organizational units, the CNN comprising a conditional random field (CRF).
2. The computer-implemented method of claim 1 , wherein the data units are cells in a table and the organizational units are rows or columns in the table.
3. The computer-implemented method of claim 1 , wherein the classifier is configured to extract information at a sub-data-unit level.
4. The computer-implemented method of claim 1 , wherein:
retrieving the classifiable data comprises breaking the first one of the organizational units into chunks of a predetermined size,
sending the classifiable data to the classifier comprises sending the chunks to the classifier;
receiving the label for the classifiable data comprises receiving a plurality of chunk labels, each chunk label corresponding to one of the chunks, and
assigning the label comprises selecting one of the chunk labels as the label for the first one of the organizational units.
5. The computer-implemented method of claim 1 , wherein:
retrieving the classifiable data comprises breaking the first one of the organizational units into chunks of a predetermined size,
sending the classifiable data to the classifier comprises grouping the chunks into one or more batches, and sending the one or more batches to the classifier;
receiving the label for the classifiable data comprises receiving a plurality of batch labels, each batch label corresponding to one of the batches, and
assigning the label comprises selecting one of the batch labels as the label for the first one of the organizational units.
6. The computer-implemented method of claim 1 , wherein the data units are cells in a table and the organizational units are rows or columns in the table, and the CNN is configured to apply a convolution kernel that encompasses a particular data unit in the first organizational unit and an adjacent data unit in a second organizational unit.
7. The computer-implemented method of claim 1 , wherein the data units are arranged in a temporal order and the classifier comprises a temporal neural network (TNN) configured to operate on the organizational units in a temporal direction, the TNN configured to mask a portion of the data units in the formatted input data.
8. The computer-implemented method of claim 1 , wherein sending the classifiable data to the classifier comprises selecting a subset of the data units within the first organizational unit and sending only the subset of the data units to the classifier.
9. The computer-implemented method of claim 1 , further comprising creating a character embedding from the classifiable data, wherein sending the classifiable data to the classifier comprises sending the character embedding to the classifier.
10. The computer-implemented method of claim 1 , further comprising flattening the input data by concatenating multiple data units of the input data.
11. A non-transitory computer-readable storage medium, the computer-readable storage medium including instructions that when executed by a computer, cause the computer to:
receive formatted input data, the formatted input data comprising a plurality of data units organized into a plurality of organizational units;
retrieve classifiable data from a first one of the organizational units;
send the classifiable data to a classifier, the classifier configured to perform a character-level classification and output a label from a predetermined set of labels;
receive a label for the classifiable data from the classifier; and
assign the label to the first one of the organizational units,
wherein the classifier is a convolutional neural network (CNN) configured to operate on the organizational units, the CNN comprising a conditional random field (CRF).
12. The computer-readable storage medium of claim 11 , wherein the data units are cells in a table and the organizational units are rows or columns in the table.
13. The computer-readable storage medium of claim 11 , wherein the classifier is configured to extract information at a sub-data-unit level.
14. The computer-readable storage medium of claim 11 , wherein:
retrieving the classifiable data comprises breaking the first one of the organizational units into chunks of a predetermined size,
sending the classifiable data to the classifier comprises sending the chunks to the classifier;
receiving the label for the classifiable data comprises receiving a plurality of chunk labels, each chunk label corresponding to one of the chunks, and
assigning the label comprises selecting one of the chunk labels as the label for the first one of the organizational units.
15. The computer-readable storage medium of claim 11 , wherein:
retrieving the classifiable data comprises breaking the first one of the organizational units into chunks of a predetermined size,
sending the classifiable data to the classifier comprises grouping the chunks into one or more batches, and sending the one or more batches to the classifier;
receiving the label for the classifiable data comprises receiving a plurality of batch labels, each batch label corresponding to one of the batches, and
assigning the label comprises selecting one of the batch labels as the label for the first one of the organizational units.
16. The computer-readable storage medium of claim 11 , wherein the data units are cells in a table and the organizational units are rows or columns in the table, and the CNN is configured to apply a convolution kernel that encompasses a particular data unit in the first organizational unit and an adjacent data unit in a second organizational unit.
17. The computer-readable storage medium of claim 11 , wherein the data units are arranged in a temporal order and the classifier comprises a temporal neural network (TNN) configured to operate on the organizational units in a temporal direction, the TNN configured to mask a portion of the data units in the formatted input data.
18. The computer-readable storage medium of claim 11 , wherein sending the classifiable data to the classifier comprises select a subset of the data units within the first organizational unit and sending only the subset of the data units to the classifier.
19. The computer-readable storage medium of claim 11 , wherein the instructions further configure the computer to create a character embedding from the classifiable data, wherein sending the classifiable data to the classifier comprises sending the character embedding to the classifier.
20. The computer-readable storage medium of claim 11 , the instructions, when executed by the computer, cause the computer to flatten the input data by concatenating multiple data units of the input data.