File matching with rationale alignment using neural networks and an optimal transport algorithm

A non-transitory computer-readable medium comprising a computer program, that, when executed by a computer system, enables the computer system to perform the following method for matching an input file to one of a plurality of files in a datastore and displaying the rationale for the match, the method comprising:

• constructing a cost matrix and an alignment matrix for each input file-datastore file pair by performing the following:
• creating a first set of objects from the input file;
• creating a second set of objects from a datastore file;
• applying a first neural network to the first set of objects to obtain a first set of vectors representative of the first set of objects;
• applying a second neural network to the second set of objects to obtain a second set of vectors representative of the second set of objects;
• adding one or more dummy points to at least one of the sets of vectors to constrain alignment results;
• constructing a cost matrix that indicates the cost of aligning each possible pairing of first and second set vectors, wherein constructing the cost matrix comprises computing, for each vector pair having a first vector from the first set of vectors and a second vector from the second set of vectors, a cost for aligning the first and second vectors in the vector pair, and wherein any vector pairing that includes a dummy point is assigned a fixed cost;
• using the cost matrix and an optimal transport algorithm to compute an alignment matrix that indicates the vector pairings between the first and second set of vectors that result in the lowest total alignment cost;
• identifying the input file-datastore file pair associated with the lowest total cost of alignment based on the cost and alignment matrices;
• and
• displaying in a user interface an indication of matching objects in the identified pair using the alignment matrix associated with the identified pair, wherein an object having a vector representation that is paired with a dummy point in the alignment matrix is treated as unmatched.

A computer system for matching an input file to one of a plurality of files in a datastore and displaying the rationale for the match, the system comprising:

• one or more processors;
• one or more memory units coupled to the one or more processors, wherein the one or more memory units store instructions that, when executed by the one or more processors, cause the system to perform the operations of:
• constructing a cost matrix and an alignment matrix for each input file-datastore file pair by performing the following:
• creating a first set of objects from the input file;
• creating a second set of objects from a datastore file;
• applying a first neural network to the first set of objects to obtain a first set of vectors representative of the first set of objects;
• applying a second neural network to the second set of objects to obtain a second set of vectors representative of the second set of objects;
• adding one or more dummy points to at least one of the sets of vectors to constrain alignment results;
• constructing a cost matrix that indicates the cost of aligning each possible pairing of first and second set vectors, wherein constructing the cost matrix comprises computing, for each vector pair having a first vector from the first set of vectors and a second vector from the second set of vectors, a cost for aligning the first and second vectors in the vector pair, and wherein any vector pairing that includes a dummy point is assigned a fixed cost;
• using the cost matrix and an optimal transport algorithm to compute an alignment matrix that indicates the vector pairings between the first and second set of vectors that result in the lowest total alignment cost;
• identifying the input file-datastore file pair associated with the lowest total cost of alignment based on the cost and alignment matrices;
• and
• displaying in a user interface an indication of matching objects in the identified pair using the alignment matrix associated with the identified pair, wherein an object having a vector representation that is paired with a dummy point in the alignment matrix is treated as unmatched.