Numeric embeddings for entity-matching
Pairwise entity matching systems and methods are disclosed herein. A deep learning model may be used to match entities from separate data tables. Entities may be preprocessed to fuse textual and numeric data early in the neural network architecture. Numeric data may be represented as a vector of a geometrically progressing function. By fusing textual and numeric data, including dates, early in the neural network architecture the neural network may better learn the relationships between the numeric and textual data. Once preprocessed, the paired entities may be scored and matched using a neural network.
1. One or more non-transitory computer-readable media storing computer-executable instructions that, when executed by a processor, perform a method of matching data table entities, the method comprising:
ingesting a set of data comprising a plurality of entities, wherein each entity of the plurality of entities comprises textual data and numeric data;
pairing entities from the plurality of entities to form an entity pair;
for the textual data in each entity of the entity pair:
tokenizing the textual data into a set of textual tokens; and
contextualizing the set of textual tokens by passing the set of textual tokens through at least one neural network;
for the numeric data in each entity of the entity pair:
converting each numeric value in the numeric data to an integer; and
mapping each integer to a vector;
concatenating the textual data with the numeric data to form a sequence for each entity in the entity pair; and
analyzing the sequences using a deep learning model to classify the entity pair.
2. The media of claim 1 , wherein each entity comprises one of a row or a column in a data table.
3. The media of claim 1 , wherein the at least one neural network is a one-dimensional convolutional neural network.
4. The media of claim 1 , wherein the deep learning model is a decomposable attention neural network.
5. The media of claim 1 , wherein:
the numeric data further comprises date data; and
the date data is converted to the integer by determining an amount of time from a reference date.
6. The media of claim 1 , wherein the vector is a vector of a geometric progression of frequencies.
7. The media of claim 1 , wherein the entity pair is classified using cross entropy loss.
8. A method of matching data table entities, the method comprising:
ingesting a set of data comprising a plurality of entities, wherein each entity of the plurality of entities comprises textual data and numeric data;
pairing entities from the set of data to form an entity pair;
contextualizing, for each entity of the entity pair, the textual data using at least one neural network;
mapping, for each entity of the entity pair, each numeric value of the numeric data to a vector;
concatenating, for each entity of the entity pair, the textual data with the numeric data to form a sequence; and
analyzing the sequences using a deep learning model to classify the entity pair.
9. The method of claim 8 , wherein each entity of the plurality of entities comprises one of a row or a column in a data table.
10. The method of claim 8 , further comprising:
augmenting each textual token with an index representing a field of origin in the entity for the textual token; and
augmenting each numeric value with the index representing the field of origin in the entity for the numeric value.
11. The method of claim 8 , further comprising:
pooling the sequences to reduce a dimension of the neural network;
concatenating the sequences; and
classifying the sequences using a binary classification.
12. The method of claim 8 , wherein the at least one neural network is a long short term memory network.
13. The method of claim 8 , further comprising:
tokenizing, for each entity of the entity pair, the textual data, wherein the textual data is tokenized at a character level.
14. A system for matching data table entities, the system comprising:
a data store storing a plurality of entities; and
one or more non-transitory computer-readable media storing computer-executable instructions that, when executed by a processor, perform a method of matching data table entities, the method comprising:
pairing entities from the plurality of entities, wherein each entity comprises textual data and numeric data;
for the textual data in each entity of the paired entities:
tokenizing the textual data into a set of textual tokens; and
contextualizing the set of textual tokens using at least one neural network;
for the numeric data in each entity of the paired entities:
converting each numeric value in the numeric data to an integer; and
mapping each integer to a vector of a geometrically progressing frequency;
concatenating the textual data with the numeric data to form a sequence for each entity of the paired entities; and
analyzing the sequences using a deep learning model to classify the paired entities.
15. The system of claim 14 , wherein each entity comprises one of a row or a column in a data table.
16. The system of claim 14 , further comprising applying a linear mapping to the numeric data to match a dimension of the numeric data to a dimension of the textual data.
17. The system of claim 14 , wherein the computer-executable instructions are further executed to classify the sequences using cross-entropy loss.
18. The system of claim 14 , wherein the at least one neural network is a recurrent neural network.
19. The system of claim 14 , wherein the system further comprises:
a trained function represented by a neural network for scoring the paired entities, wherein the paired entities are classified into a classification based on a score from the trained function.
20. The system of claim 19 , wherein the computer-executable instructions are further executed to perform a step of determining, by the trained function, a confidence score for the classification.