Neural networks for information extraction from transaction data

A system comprising:

• one or more processors;
• and
• a non-transitory computer-readable medium storing instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
• receiving, by an information extraction system including one or more processors, a transaction record, the transaction record including a plurality of tokens, the transaction record describing a transaction served by a service provider;
• determining, by a convolutional neural network (CNN) processing module of the information extraction system, a classification of the transaction record based on a collection of parameters that the CNN processing module learned from first training data, the determining comprising
• determining whether the service provider is known to the information extraction system or not known to the information extraction system,
• if determining that the service provider is known to the information extraction system, generating a class label identifying a name of the service provider, and
• if determining that the service provider is not known to the information extraction system, generating a predetermined class label indicating that the service provider is not a known service provider, wherein the CNN processing module has been configured through training to generate the predetermined class label in response to processing a transaction record corresponding to a service provider that is not known to the information extraction system;
• in response to determining that the service provider is not known to the information extraction system, as indicated by the generation of the predetermined class label by the CNN processing module, triggering a recurrent neural network (RNN) processing module of the information extraction system, comprising
• locating, by the RNN processing module and based at least in part on a character embedding model and a word embedding model both of which the RNN processing module learned from second training data, a particular token of the plurality of tokens of the transaction record, wherein the particular token identifies the name of the service provider;
• and
• generating, by the information extraction system, a report specifying the name of the service provider in the transaction record.

A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:

• receiving, by an information extraction system including one or more processors, a transaction record, the transaction record including a plurality of tokens, the transaction record describing a transaction served by a service provider;
• determining, by a convolutional neural network (CNN) processing module of the information extraction system, a classification of the transaction record based on a collection of parameters that the CNN processing module learned from first training data, the determining comprising
• determining whether the service provider is known to the information extraction system or not known to the information extraction system,
• if determining that the service provider is known to the information extraction system, generating a class label identifying a name of the service provider, and
• if determining that the service provider is not known to the information extraction system, generating a predetermined class label indicating that the service provider is not a known service provider, wherein the CNN processing module has been configured through training to generate the predetermined class label in response to processing a transaction record corresponding to a service provider that is not known to the information extraction system;
• in response to determining that the service provider is not known to the information extraction system, as indicated by generating the generation of the predetermined class label by the CNN processing module, triggering a recurrent neural network (RNN) processing module of the information extraction system, comprising
• locating, by the RNN processing module and based at least in part on a character embedding model and a word embedding model both of which the RNN processing module learned from second training data, a particular token of the plurality of tokens of the transaction record, wherein the particular token identifies the name of the service provider;
• and
• generating, by the information extraction system, a report specifying the name of the service provider in the transaction record.