Systems and methods for determining a response to a user query

Patent metadata

Field	Value
Publication number	US-11741139-B2
Application number	US-202117317805-A
Country	US
Kind code	B2
Filing date	May 11, 2021
Priority date	May 12, 2020
Publication date	Aug 29, 2023
Grant date	Aug 29, 2023

Abstract

Systems and methods are presented for providing a response to a user query. Reception of a user query is detected. An augmentation machine learning model is utilized to determine one or more variations of the user query that correspond to a semantic meaning of the user query. A plurality of response candidates is determined that correspond to the user query by comparing the user query and the one or more variations of the user query to a plurality of documents. A final response candidate is determined from the plurality of response candidates based on utilizing a semantic machine learning model to perform a semantic comparison between the plurality of response candidates and at least the user query.

First claim

What is claimed is:

1. A system comprising:
a non-transitory memory having program instructions stored thereon; and
one or more hardware processors coupled to the non-transitory memory and configured to execute the stored instructions to cause the system to perform operations comprising:
detecting reception of a user query by a chatbot;
utilizing an augmentation machine learning model to determine multiple different variations of the user query that are semantically related to the user query;
storing the multiple different variations of the user query in a first asynchronous decoupling buffer;
accessing a set of documents, wherein a given document includes one or more query-response pairs, including:
querying the set of documents by comparing queries of query-response pairs in the documents with both:
the user query; and
the multiple different variations of the user query from the first asynchronous decoupling buffer;

determining a plurality of response candidates based on the querying and storing the plurality of response candidates in a second asynchronous decoupling buffer;
determining a final response candidate from the plurality of response candidates from the second asynchronous decoupling buffer, wherein the determining utilizes a semantic machine learning model to perform a semantic comparison between the plurality of response candidates and at least the user query; and
outputting, by the chatbot, a response to the user based on the determined final response candidate.

2. The system of claim 1 , wherein determining the plurality of response candidates that correspond to the user query is based on a term frequency-inverse document frequency process comparing the user query and the multiple different variations of the user query to the set of documents.

3. The system of claim 1 , wherein the operations further comprise:
monitoring user browsing activity; and
generating the user query and directing the user query to the chatbot based on the browsing activity.

4. The system of claim 1 ,
wherein the utilizing the augmentation machine learning model includes:
performing a most-similar word search to generate words similar to words in the user query;
mapping the user query and results of the most-similar word search to a vector space;
performing a distributed word-embedding technique to compare learned representations of semantically-related words as multi-dimensional vectors in the vector space; and
outputting the multiple different variations based on the distributed word-embedding;

wherein the utilizing of the semantic machine learning model includes:
providing token sequences that include tokens from the user query and tokens from a given response candidate;
processing the token sequences using a Bidirectional Encoder Representations from Transformers (BERT) model to generate score values that reflects a semantic relationship between the user query and a given response candidate;
generating an initial ranking of the score values;
adjust the initial ranking, using a ranking loss model, to generate a final ranking, wherein the determining the response candidate is based on the final ranking; and
propagating the final ranking to the BERT model.

5. The system of claim 1 , wherein the determining the final response candidate includes:
determining scores for the plurality of response candidates, and
identifying the response candidate having the highest score as the final response candidate.

6. The system of claim 5 , wherein, when determining scores for the plurality of response candidates, semantic similarities of the response candidates to the user query are weighted more than syntactic similarities of the response candidates to the user query.

7. The system of claim 1 , wherein the first asynchronous decoupling buffer is a push/pull buffer.

8. The system of claim 1 , wherein the operations further comprise utilizing one or more normalization techniques to normalize the user query; and
wherein the augmentation machine learning model determines the multiple different variations of the user query that correspond to a semantic meaning of the normalized user query.

9. The system of claim 8 , wherein utilizing one or more normalization techniques to normalize the user query includes utilizing a text normalization machine learning model to normalize the user query.

10. A computer-implemented method comprising:
detecting, by a computing system, reception of a user query by a chatbot;
utilizing, by the computing system, an augmentation machine learning model to determine multiple different variations of the user query that are semantically related to the user query;
storing, by the computing system, the multiple different variations of the user query in a first asynchronous decoupling buffer;
accessing, by the computing system, a set of documents, wherein a given document includes one or more query-response pairs, including:
querying, by the computing system, the set of documents by comparing queries of query-response pairs in the documents with both:
the user query; and
the multiple different variations of the user query from the first asynchronous decoupling buffer;

determining, by the computing system, a plurality of response candidates based on the querying and storing the plurality of response candidates in a second asynchronous decoupling buffer;
determining, by the computing system, a final response candidate from the plurality of response candidates from the second asynchronous decoupling buffer, wherein the determining utilizes a semantic machine learning model to perform a semantic comparison between the plurality of response candidates and at least the user query; and
outputting, by the chatbot, a response to the user based on the determined final response candidate.

11. The computer-implemented method of claim 10 , wherein determining the plurality of response candidates that correspond to the user query is based on a term frequency-inverse document frequency process to compare the user query and the multiple different variations of the user query to the set of documents.

12. The computer-implemented method of claim 10 , wherein the user query and the multiple different variations of the user query are syntactically different and semantically similar.

13. The computer-implemented method of claim 10 , wherein the determining the final response candidate includes:
determining scores for the plurality of response candidates, and
identifying the response candidate having the highest score as the final response candidate.

14. The computer-implemented method of claim 10 , further comprising:
monitoring, by the system, user browsing activity; and
generating, by the system, the user query and directing the user query to the chatbot based on the browsing activity.

15. A non-transitory machine-readable medium having stored thereon machine-readable instructions executable to cause a machine to perform operations comprising:
detecting reception of a user query by a chatbot;
utilizing an augmentation machine learning model to determine multiple different variations of the user query that are semantically related to correspond to a semantic meaning of the user query;
storing the multiple different variations of the user query in a first asynchronous decoupling buffer;
accessing a set of documents, wherein a given document includes one or more query-response pairs, including:
querying the set of documents by comparing queries of query-response pairs in the documents with both:
the user query; and
the multiple different variations of the user query fr

Assignees

• Paypal Inc

Inventors

• Zhuo Yuzhen
• Cavallari Sandro
• Nguyen Van Hoang
• Bui Kim Dung
• Neo Rey
• Singalreddy Harsha
• Xu Lei
• Wang Hewen
• Tang Quan Jin Ferdinand
• Ho Chun Kiat

Classifications

• G06F16/3322 Primary

  using system suggestions (G06F16/3325 takes precedence) · CPC title

• G06F16/3334

  Selection or weighting of terms from queries, including natural language queries · CPC title

• G06F16/3338

  Query expansion · CPC title

• G06N20/00

  Machine learning · CPC title

• H04L51/02

  using automatic reactions or user delegation, e.g. automatic replies or chatbot-generated messages · CPC title