Granted Patent B2

US 12,591,598 · App. 18/912,545 · Granted Mar 31, 2026

Inventors: Maksymilian Clark Polaczuk (Wellington, NZ); Christopher Darius Herrmann (Wellington, NZ); Timothy Matthew Leathart (Wellington, NZ); Divya Jitesh Patel (Wellington, NZ)

Assignee: Xero Limited

G06F16/288G06F16/285G06V30/19093G06V2201/09G06V2201/10

Abstract

Claims (78)

1 . A computer-implemented method comprising:

providing, to a first machine learning model, entity information for each of a plurality of entities, wherein the first machine learning model has been trained to generate an entity numerical representation for each of the plurality of entities based on the respective entity information of the entity;

determining, by the trained first machine learning model, a plurality of entity numerical representations based on received respective entity information for the plurality of entities, wherein the entity numerical representation comprises a numerical representation of the information related to the entity;

receiving an entity search text from a computing device;

providing the entity search text to the trained first machine learning model to generate a query numerical representation of the entity search text;

comparing the query numerical representation with each of the plurality of entity numerical representations to determine a similarity score indicating a similarity between the query numerical representation and a respective entity numerical representation; and

based on the determined similarity scores, identifying an entity numerical representation most similar to the entity search text,

wherein the comparing the query numerical representation with each of the plurality of entity numerical representations is performed by a matrix multiplication operation.

2 . The method of claim 1 , wherein the trained first machine learning model comprises a character embedding module, word embedding module, character-word composite embedding module and a composite numerical representation processing module, and

wherein determining, by the trained first machine learning model, a plurality of entity numerical representations based on the received respective entity information for the plurality of entities comprises:

determining, by the character embedding module a character numerical representation for each character in the respective entity information;

determining, by the word embedding module a word numerical representation for each word in the respective entity information;

providing, to the character-word composite embedding module the determined character and word numerical representations;

determining, by the character-word composite embedding module a composite character-word numerical representation for each word based on the determined word numerical representations and the character numerical representations for each respective character in each word;

providing, to the composite numerical representation processing module the composite character-word numerical representation for each word in the entity information;

determining, by the composite numerical representation processing module an entity numerical representation based on the composite character-word numerical representations.

3 . The method of claim 1 , wherein the entity information comprises one or more web pages hosted by an entity server associated with the respective entity.

4 . The method of claim 1 , wherein each of the plurality of entity representations is associated with an attribute of the respective entity, and wherein the attribute comprises one or more of: entity location, entity category, entity type and entity employee information.

5 . The method of claim 1 , further comprising validating the entity information, by an entity information validation module, before providing the entity information to the first machine learning model.

6 . The method of claim 1 , further comprising:

extracting one or more candidate logo images from the entity information of one or more entities of the plurality of entities;

providing each candidate logo image to an optical character recognition (OCR) module to determine candidate logo text associated with each candidate logo image;

determining, as an output of the OCR module, candidate logo text associated with each candidate logo image;

for each candidate logo text, determining a logo text similarity metric indicating a similarity between the candidate logo text and an entity identifier of the respective entity; and

based on the determined logo text similarity metric, determining a candidate entity logo as a designated entity logo of that entity.

7 . The method of claim 1 , further comprising:

extracting one or more candidate logo images and respective metadata from the entity information of one or more entities of the plurality of entities;

determining a candidate logo feature vector for each of the one or more candidate logo images based on the respective metadata;

providing each candidate logo image to an optical character recognition (OCR) module to determine candidate logo text associated with each candidate logo image;

determining, as an output of the OCR module, candidate logo text associated with each candidate logo image;

for each candidate logo text, determining a logo text similarity metric indicating a similarity between the candidate logo text and an entity identifier of the respective entity;

for each candidate logo image, providing the candidate logo feature vector and the logo text similarity metric to a first logo determination model configured to determine a logo probability score;

determining, as an output of the first logo determination model, a logo probability score for each candidate logo image; and

based on the determined logo probability scores, determining a candidate entity logo as a designated entity logo of that entity.

8 . The method of claim 7 , wherein the entity search text comprises one or more representative logo images, and wherein the method further comprises:

providing each representative logo image to the OCR module to determine representative logo text associated with each representative logo image;

determining, as an output of the OCR module, representative logo text associated with each representative logo image;

for each representative logo text, determining a logo text similarity metric indicating a similarity between the representative logo text and a designated entity logo of the plurality of entities; and

based on the determined logo text similarity metric, determining a representative entity logo as being most similar to a logo image of the one or more representative logo images of the entity search text.

9 . A system comprising:

one or more processors; and

memory comprising computer code, which when executed by the one or more processors, causes the system to:

provide, to a first machine learning model, entity information for each of a plurality of entities, wherein the first machine learning model has been trained to generate an entity numerical representation for each of the plurality of entities based on the respective entity information of the entity;

determine, by the trained first machine learning model, a plurality of entity numerical representations based on received respective entity information for the plurality of entities, wherein the entity numerical representation comprises a numerical representation of the information related to the entity;

receive an entity search text from a computing device;

provide the entity search text to the trained first machine learning model to generate query numerical representation of the entity search text;

compare the query numerical representation with each of the plurality of entity numerical representations to determine a similarity score indicating a similarity between the query numerical representation and a respective entity numerical representation; and

based on the determined similarity scores, identify an entity numerical representation most similar to the entity search text,

wherein the comparing the query numerical representation with each of the plurality of entity numerical representations is performed by a matrix multiplication operation.

10 . The system of claim 9 , wherein the trained first machine learning model comprises a character embedding module, word embedding module, character-word composite embedding module and a composite numerical representation processing module, and

wherein determining, by the trained first machine learning model, a plurality of entity numerical representations based on the received respective entity information for the plurality of entities comprises:

determining, by the character embedding module a character numerical representation for each character in the respective entity information;

determining, by the word embedding module a word numerical representation for each word in the respective entity information;

providing, to the character-word composite embedding module the determined character and word numerical representations;

determining, by the character-word composite embedding module a composite character-word numerical representation for each word based on the determined word numerical representations and the character numerical representations for each respective character in each word;

providing, to the composite numerical representation processing module the composite character-word numerical representation for each word in the entity information;

determining, by the composite numerical representation processing module an entity numerical representation based on the composite character-word numerical representations.

11 . The system of claim 9 , wherein the entity information comprises one or more web pages hosted by an entity server associated with the respective entity.

12 . The system of claim 9 , wherein each of the plurality of entity representations is associated with an attribute of the respective entity, and wherein the attribute comprises one or more of: entity location, entity category, entity type and entity employee information.

13 . The system of claim 9 , further comprising validating the entity information, by an entity information validation module, before providing the entity information to the first machine learning model.

14 . A machine-readable medium storing computer readable code, which when executed by one or more processors is configured to perform operations including:

providing, to a first machine learning model, entity information for each of a plurality of entities, wherein the first machine learning model has been trained to generate an entity numerical representation for each of the plurality of entities based on the respective entity information of the entity;

determining, by the trained first machine learning model, a plurality of entity numerical representations based on received respective entity information for the plurality of entities, wherein the entity numerical representation comprises a numerical representation of the information related to the entity;

receiving an entity search text from a computing device;

providing the entity search text to the trained first machine learning model to generate query numerical representation of the entity search text;

comparing the query numerical representation with each of the plurality of entity numerical representations to determine a similarity score indicating a similarity between the query numerical representation and a respective entity numerical representation; and

based on the determined similarity scores, identifying an entity numerical representation most similar to the entity search text,

wherein the comparing the query numerical representation with each of the plurality of entity numerical representations is performed by a matrix multiplication operation.

15 . The machine-readable medium of claim 14 , wherein the trained first machine learning model comprises a character embedding module, word embedding module, character-word composite embedding module and a composite numerical representation processing module, and

wherein determining, by the trained first machine learning model, a plurality of entity numerical representations based on the received respective entity information for the plurality of entities comprises:

determining, by the character embedding module a character numerical representation for each character in the respective entity information;

determining, by the word embedding module a word numerical representation for each word in the respective entity information;

providing, to the character-word composite embedding module the determined character and word numerical representations;

determining, by the character-word composite embedding module a composite character-word numerical representation for each word based on the determined word numerical representations and the character numerical representations for each respective character in each word;

providing, to the composite numerical representation processing module the composite character-word numerical representation for each word in the entity information;

determining, by the composite numerical representation processing module an entity numerical representation based on the composite character-word numerical representations.

16 . The machine-readable medium of claim 14 , wherein each of the plurality of entity representations is associated with an attribute of the respective entity, and wherein the attribute comprises one or more of: entity location, entity category, entity type and entity employee information.

17 . The machine-readable medium of claim 14 , wherein the machine-readable medium is a non-transient machine-readable medium.

Assignments (1)

Priority Claims (1)

Continuity (2)

References Cited (25)