Artificial-intelligence multimodal cross-lingual cross-database item search, classification, and set compare

1 . A system comprising:

one or more data storage media configured to store specific computer-executable instructions; and

one or more computer hardware processors configured to communicate with the one or more data storage media, wherein the specific computer-executable instructions are configured to cause the one or more computer hardware processors to at least:

receive data for a first target item comprising (i) a target identifying mark, (ii) a target item category, and (iii) a target item name in a target language;

determine a set of source identifying marks received from a source electronic catalog comprising data for a plurality of source items,

wherein the plurality of source items are filtered to determine a candidate source item to be compared to the first target item,

wherein a comparison between the candidate source item and the first target item is used to determine whether the first target item has or does not have a matching item in the source electronic catalog, and

wherein data for each source item from the plurality of source items comprises (i) a source identifying mark, (ii) a source item category, and (iii) a source item name in a source language;

select, from the set of source identifying marks, a first source identifying mark based at least in part on textual similarity between the target identifying mark and the first source identifying mark;

generate by a translation model, based at least in part on the target item name, a translated target item name in the source language;

determine, from the data for the plurality of source items, a data set for the source items associated with (i) an identifying mark textually similar to the first source identifying mark and (ii) an item category textually similar the target item category;

determine, from the data set for the source items, a source name embedding for each source item name;

determine a target name embedding for the translated target item name;

determine, from the data set for the source items, data for a candidate source item based at least in part on a respective source name embedding and the target name embedding;

generate by an image generator model, based at least in part on (i) the target item name and (ii) a prompt instruction to generate a catalog image, target image data;

determine a target image embedding for the target image data;

determine a source image embedding for source image data associated with the candidate source item;

determine an image similarity indicator based at least in part on the target image embedding and the source image embedding;

determine that the image similarity indicator fails to satisfy a second threshold; and

provide a first classification result indicating that the first target item does not have a match with any source item from the plurality of source items from the source electronic catalog.

2 . The system of claim 1 , wherein to select the first source identifying mark based at least in part on textual similarity between the target identifying mark and the first source identifying mark, further computer-executable instructions configure the one or more computer hardware processors to at least:

generate by a language model, based at least in part on the target identifying mark, a target identifying mark embedding;

generate by the language model, based at least in part on the source identifying mark, a source identifying mark embedding;

apply a cosine similarity function to the target identifying mark embedding and the source identifying mark embedding that outputs a similarity score; and

determine that the similarity score satisfies a third threshold.

3 . The system of claim 1 , wherein to determine the data set for the source items, further computer-executable instructions configure the one or more computer hardware processors to at least:

determine a first source item category associated with a first source item from the plurality of source items;

generate by a language model, based at least in part on the target item category, a target item category embedding;

generate by the language model, based at least in part on the first source item category, a source item category embedding;

apply a cosine similarity function to the target item category embedding and the source item category embedding that outputs a similarity score; and

determine that the similarity score satisfies a third threshold.

4 . The system of claim 1 , wherein further computer-executable instructions configure the one or more computer hardware processors to at least:

process, from a plurality of target items, a second target item that results in a second classification result indicating that the second target item does have a match with a source item from the plurality of source items from the source electronic catalog.

5 . The system of claim 1 , wherein to determine the data for the candidate source item from the data set for the source items, further computer-executable instructions configure the one or more computer hardware processors to at least:

determine a ranking for the data set for the source items; and

select, from the data set for the source items, the data for the candidate source item based at least in part on the ranking.

6 . The system of claim 5 , wherein to determine the ranking for the data set for the source items, additional computer-executable instructions configure the one or more computer hardware processors to at least:

determine a first text overlap score based at least in part on the translated target item name and an item name for the first source item; and

order the data set for the source items based at least in part on the first text overlap score and a second text overlap score.

7 . A computer-implemented method comprising:

receiving data for a target item comprising (i) a target identifying mark, (ii) a target item category, and (iii) a first target item name;

determining a set of source identifying marks received from a source electronic catalog comprising data for a plurality of source items,

wherein data for each source item from the plurality of source items comprises (i) a source identifying mark, (ii) a source item category, and (iii) a source item name;

selecting, from the set of source identifying marks, a first source identifying mark based at least in part on textual similarity between the target identifying mark and the first source identifying mark;

determining, from data for the plurality of source items, a data set for the source items with (i) an identifying mark textually similar to the first source identifying mark and (ii) an item category textually similar the target item category;

determining, from the data set for the source items, a source name embedding for each source item name;

determining a target name embedding based at least in part on the first target item name;

determining, from the data set for the source items, a candidate source item based at least in part on the source name embedding and the target name embedding;

generating by an image generator model, based at least in part on (i) the first target item name and (ii) a prompt instruction to generate a catalog image, target image data;

determining, by a computer vision system, an image similarity indicator based at least in part on the target image data and source image data associated with the candidate source item;

determining that the image similarity indicator fails to satisfy a first threshold; and

providing a first classification result indicating that the target item does not have a match with any source item from the plurality of source items from the source electronic catalog.

8 . The computer-implemented method of claim 7 , further comprising:

receiving a second target item name associated with a second target item;

determining a second candidate source item based at least in part on the second target item name; and

determining a second classification result for the second target item and the second candidate source item, wherein determining the second classification result further comprising:

determining a textual similarity indicator based at least in part on (i) the second target item name and (ii) an item name associated with the second candidate source item;

determining that the textual similarity indicator satisfies a second threshold; and

identifying the second classification result that the second target item matches the second candidate source item without performing an image comparison for the second target item and the second candidate source item.

9 . The computer-implemented method of claim 8 , wherein determining the textual similarity indicator further comprises:

determine a first set of word chunks from the second target item name;

determine a second set of word chunks from the item name associated with the second candidate source item; and

determine a textual similarity indicator from the first set of word chunks and the second set of word chunks.

10 . The computer-implemented method of claim 8 , wherein determining the textual similarity indicator further comprises:

generating by a language model, based at least in part on (i) the second target item name, (ii) the item name associated with the second candidate source item, and (iii) a prompt asking whether items identifiable from input data are the same or different, the textual similarity indicator.

11 . The computer-implemented method of claim 7 , wherein determining the image similarity indicator further comprises:

generating by a multimodal language model, based at least in part on (i) the target image data, (ii) the source image data, and (iii) a prompt asking whether items in images associated with the target image data and the source image data are the same or different, the image similarity indicator.

12 . The computer-implemented method of claim 7 , wherein the source identifying mark comprises a brand.

13 . The computer-implemented method of claim 7 , wherein determining the image similarity indicator further comprises:

determining a target image embedding for the target image data;

determining a source image embedding for the source image data; and

applying a cosine similarity function to the target image embedding and the source image embedding that outputs the image similarity indicator.

14 . A system comprising:

one or more data storage media configured to store specific computer-executable instructions; and

one or more computer hardware processors configured to communicate with the one or more data storage media, wherein the specific computer-executable instructions are configured to cause the one or more computer hardware processors to at least:

receive data for a first target item comprising (i) a target identifying mark, (ii) a target item category, and (iii) target text;

select, from a set of source identifying marks, a first source identifying mark based at least in part on textual similarity between the target identifying mark and the first source identifying mark;

determine, from data for a plurality of source items, a first candidate source item with (i) a source identifying mark textually similar to the first source identifying mark and (ii) a source item category textually similar to the target item category,

wherein the first candidate source item is associated with source text;

determine a textual similarity indicator based at least in part on the target text and the source text;

generate by an image generator model, based at least in part on the target text, target image data;

determine, by a computer vision system, an image similarity indicator based at least in part on the target image data and source image data associated with the first candidate source item;

determine that one or more thresholds are satisfied based at least in part on the textual similarity indicator and the image similarity indicator; and

provide a first classification result indicating that the first target item matches the first candidate source item.

15 . The system of claim 14 , wherein to select the first source identifying mark based at least in part on textual similarity between the target identifying mark and the first source identifying mark, further computer-executable instructions configure the one or more computer hardware processors to at least:

generate by a language model, based at least in part on the target identifying mark, a target identifying mark embedding;

generate by the language model, based at least in part on the source identifying mark, a source identifying mark embedding;

apply a cosine similarity function to the target identifying mark embedding and the source identifying mark embedding that outputs a similarity score; and

determine that the similarity score satisfies a threshold.

16 . The system of claim 14 , wherein to determine textual similarity between the target text and the source text, further computer-executable instructions configure the one or more computer hardware processors to at least:

generate by a language model, based at least in part on the target text, a target text embedding;

generate by the language model, based at least in part on the source text, a source text embedding;

apply a cosine similarity function to the target text embedding and the source text embedding that outputs a similarity score; and

determine that the similarity score satisfies a threshold.

17 . The system of claim 14 , wherein to determine the image similarity indicator, further computer-executable instructions configure the one or more computer hardware processors to at least:

determine a target image embedding for the target image data;

determine a source image embedding for the source image data; and

apply a cosine similarity function to the target image embedding and the source image embedding that outputs the image similarity indicator.

18 . The system of claim 17 , wherein to determine the target image embedding, additional computer-executable instructions configure the one or more computer hardware processors to at least:

generate, by a vision transformer, the target image embedding based at least in part on the target image data.

19 . The system of claim 14 , wherein further computer-executable instructions configure the one or more computer hardware processors to at least:

receive text data associated with a second target item;

determine a second candidate source item based at least in part on the text data associated with the second target item; and

determine a second classification result for the second target item and the second candidate source item, wherein to determine the second classification result, the further computer-executable instructions configure the one or more computer hardware processors to at least:

determine a chunked similarity score based at least in part on (i) the text data associated with the second target item and (ii) text data associated with the second candidate source item;

determine that the chunked similarity score satisfies a threshold; and

identify the second classification result that the second target item matches the second candidate source item without performing an image comparison for the second target item and the second candidate source item.

20 . The system of claim 14 , wherein the target text is in a target language, wherein the image generator model corresponds to a multilingual image generator model, and wherein to generate the target image data, further computer-executable instructions configure the one or more computer hardware processors to at least:

generate the target image data with the image generator model that receives the target text in the target language as a prompt.