Granted Patent B2

US 12,749,104 · App. 18/392,700 · Granted Sep 29, 2026

Inventors: Raghavendra Subramanya (Frisco, TX); Praneeth Gubbala (Milpitas, CA); Xuan Zhang (Georgetown, TX); Shilpa Gopinath (Kerala, IN); Kamalpreet Singh (Dallas, TX); Dong Nguyen (Wylie, TX); Harish Bondalapati (Bentonville, AR); Bahula Bosetti (Bentonville, AR); Sneha Oladhri (Irving, TX); Nishant Kumar Mishra (Pradesh, IN); Lakshya Khandelwal (Kota Rajasthan, IN); Navesh Garg (Bangalore, IN)

Assignee: Walmart Apollo, LLC

G06Q30/0627G06F40/284G06F40/40

Abstract

Claims (74)

1 . A system for conversational item search using multi-facet actions, the system comprising:

a processor; and

a computer-readable medium storing instructions that are operative upon execution by the processor to:

obtain a single utterance of a user comprising a conversational search query, the conversational search query comprising a plurality of words identifying an item and a plurality of search refinement terms associated with the item;

convert the single utterance into a bidirectional sequence of tokens using a natural-language model comprising a set of stacked transformer encoder layers and a conditional random field (CRF) decoder, the bidirectional sequence of tokens representing the plurality of words;

extract, from the bidirectional sequence of tokens, a plurality of facet-actions, each facet-action comprising an entity corresponding to an item attribute, and entity-role identifying a semantic refinement value associated with the entity, and a mapped filter action selected from a set of filter operations including exact, exclude, greater-than, less-than, and range;

identify, based on the plurality of facet-actions, a corresponding plurality of multi-facet filters;

apply the plurality of multi-facet filters to a plurality of items in a catalog, the plurality of multi-facet filters removing items that are irrelevant to the conversational search query;

identify a plurality of candidate items from the plurality of items remaining after application of the plurality of multi-facet filters;

score the plurality of candidate items using scoring criteria comprising a semantic-similarity measure between the conversational search query and item-attribute data associated with each candidate item;

select a candidate item from the plurality of candidate items having a highest score; and

generate a search result comprising the selected candidate item, wherein the search result is presented to the user via a user interface.

2 . The system of claim 1 , wherein the instructions are further operative to:

convert the single utterance into a bidirectional tokenized query string comprising a plurality of bidirectional tokens representing the plurality of words, wherein an extractor analyzes the plurality of bidirectional tokens to extract the plurality of facet-actions from the single utterance.

3 . The system of claim 1 , wherein the instructions are further operative to:

identify the entity associated with an attribute of the item;

recognize an entity role corresponding to the entity; and

map the entity and the entity role to a set of actions, wherein a facet-action is associated with the entity, the entity role, and the set of actions mapped to the entity role.

4 . The system of claim 1 , wherein scoring the plurality of candidate items further comprises:

generate two sub-scores associated with each candidate item, wherein the two sub-scores are weighted and summed to obtain a score for each candidate item.

5 . The system of claim 1 , wherein the instructions are further operative to:

identify a first role associated with the entity and a second role associated with the entity;

identify a first action associated with the entity and the first role; and

identify a second action associated with the entity and the second role, wherein a first facet-action corresponds to the first action and a second facet-action corresponds to the second action.

6 . The system of claim 1 , wherein the instructions are further operative to:

identify a price entity, wherein the price entity is associated with a first numeric value role and a second numeric value role, the first numeric value role being lower than the second numeric value role; and

identify a first facet-action of greater than for the first numeric value role and a second facet-action of less than for the second numeric value role, wherein a first filter associated with the first facet-action and a second filter associated with the second facet-action are applied to remove items having a price that falls outside a specified price range.

7 . The system of claim 1 , wherein identifying the entity, recognizing the entity-role, and mapping the entity and the entity-role to the plurality of facet-actions is performed by a named-entity recognition model comprising the set of stacked transformer encoder layers and the conditional random field (CRF) decoder that jointly predicts entity labels and entity-role labels using a BILOU tagging scheme.

8 . A method for conversational item search using multi-facet actions, the method comprising:

obtaining a single utterance of a user comprising a conversational search query, the conversational search query comprising a plurality of words identifying an item and a plurality of search refinement terms associated with the item;

converting the single utterance into a bidirectional sequence of tokens using a natural-language model comprising a set of stacked transformer encoder layers and a conditional random field (CRF) decoder, the bidirectional sequence of tokens representing the plurality of words;

extracting, from the bidirectional sequence of tokens, a plurality of facet-actions corresponding to the plurality of search refinement terms from the plurality of words, a facet comprising an item attribute, wherein a facet-action comprises a set of filter actions corresponding to the item attribute, each facet-action comprising an entity corresponding to the item attribute, an entity-role identifying a semantic refinement value associated with the entity, and a mapped filter action selected from a set of filter operations including exact, exclude, greater-than, less-than, and range;

identifying, based on the plurality of facet-actions, a corresponding plurality of multi-facet filters;

applying the plurality of multi-facet filters to a plurality of items in a catalog, the plurality of multi-facet filters corresponding to the plurality of facet-actions removing items that are irrelevant to the conversational search query;

identifying a plurality of candidate items from the plurality of items remaining after application of the plurality of multi-facet filters;

scoring the plurality of candidate items using a set of scoring criteria comprising a semantic-similarity measure between the conversational search query and item-attribute data associated with each candidate item;

selecting a candidate item from the plurality of candidate items having a highest score; and

generating a search result comprising the selected candidate item, wherein the search result is presented to the user via a user interface.

9 . The method of claim 8 , wherein identifying the entity and the entity-role from the bidirectional sequence of tokens is performed by a named-entity recognition model comprising the set of stacked transformer encoder layers and the conditional random field (CRF) decoder that jointly predicts entity labels and entity-role labels using a BILOU tagging scheme.

10 . The method of claim 8 , wherein training of the natural-language model comprises optimizing a multi-task loss function including an entity-classification loss, an entity-role loss, and a masked-token prediction loss, a total loss being a sum of the entity-classification loss, the entity-role loss, and a mask-modeling loss.

11 . The method of claim 10 , wherein optimizing the masked-token prediction loss comprises masking 15% of tokens in each input sequence such that, for each selected token, 70% are replaced with a mask token, 10% are replaced with a random token, and 20% are left unchanged.

12 . The method of claim 8 , further comprising:

generating two sub-scores associated with each candidate item;

weighting the two sub-scores; and

summing to obtain a score for each candidate item.

13 . The method of claim 8 , further comprising:

identifying a first role associated with the entity forming a first entity-role and a second role associated with the entity forming a second entity-role; and

identifying a first action associated with the first entity-role and a second action associated with the second entity-role, wherein a first facet-action corresponds to the first action, and wherein a second facet-action corresponds to the second action.

14 . The method of claim 8 , wherein a facet-action in the plurality of facet-actions comprises at least one of a brand, price, rating, discount, color, material, and size associated with an item, and wherein a facet-action in the plurality of facet-actions comprises at least one of a greater than action, a less than action, a range action, an exclude action, and an exact match action.

15 . One or more computer storage devices having computer-executable instructions stored thereon, which, upon execution by a computer, cause the computer to perform operations comprising:

obtaining a single utterance of a user comprising a conversational search query, the conversational search query comprising a plurality of words identifying an item and a plurality of search refinement terms associated with the item;

converting the single utterance into a bidirectional sequence of tokens using a natural-language model comprising a set of stacked transformer encoder layers and a conditional random field (CRF) decoder, the bidirectional sequence of tokens representing the plurality of words;

extracting, from the bidirectional sequence of tokens, a plurality of facet-actions corresponding to the plurality of search refinement terms from the plurality of words, a facet comprising an item attribute, wherein a facet-action comprises a set of filter actions corresponding to the item attribute, each facet-action comprising an entity corresponding to the item attribute, an entity-role identifying a semantic refinement value associated with the entity, and a mapped filter action selected from a set of filter operations including exact, exclude, greater-than, less-than, and range;

identifying, based on the plurality of facet-actions, a corresponding plurality of multi-facet filters;

applying the plurality of multi-facet filters to a plurality of items in a catalog, the plurality of multi-facet filters corresponding to the plurality of facet-actions removing items that are irrelevant to the conversational search query;

identifying a plurality of candidate items from the plurality of items remaining after application of the plurality of multi-facet filters;

scoring the plurality of candidate items using a set of scoring criteria comprising a semantic-similarity measure between the conversational search query and item-attribute data associated with each candidate item;

selecting a candidate item from the plurality of candidate items having a highest score; and

generating a search result comprising the selected candidate item, wherein the search result is presented to the user via a user interface.

16 . The one or more computer storage devices of claim 15 , wherein the operations further comprise:

training a machine learning model using a set of customized training data to filter items based on multiple item attributes and multiple filter actions.

17 . The one or more computer storage devices of claim 15 , wherein the operations further comprise:

converting the single utterance into a bidirectional tokenized query string comprising a plurality of bidirectional tokens representing the plurality of words, wherein an extractor analyzes the plurality of bidirectional tokens to extract the plurality of facet-actions from the single utterance.

18 . The one or more computer storage devices of claim 15 , wherein the operations further comprise:

identifying the entity associated with an attribute of the item;

recognizing an entity role corresponding to the entity; and

identifying a set of actions corresponding to the entity and the entity role, wherein the facet-action is associated with the entity, the entity role, and the set of actions mapped to the entity role.

19 . The one or more computer storage devices of claim 15 , wherein the operations further comprise:

generating two sub-scores associated with each candidate item, wherein the two sub-scores are weighted and summed to obtain a score for each candidate item.

20 . The one or more computer storage devices of claim 15 , wherein the operations further comprise:

generating a first facet-action corresponding to a first entity and first role associated with the first entity;

generating a second facet-action corresponding to the first entity and a second role associated with the first entity; and

applying a first filter to a set of items, wherein the first filter corresponds to the first facet-action; and

applying a second filter to the set of items, wherein the second filter corresponds to the second facet-action.

Assignments (3)

Continuity (1)

References Cited (29)