CN-122029532-A - Method and system for generating customized replies to conversational interactions

Abstract

Methods and server systems for generating custom reply suggestions are described herein. The method performed by the server system includes generating, by a Large Language Model (LLM), a set of dialog-specific embeddings for a plurality of dialog interactions based on a plurality of dialog-specific features corresponding to each dialog interaction. A set of dialogue clusters is then generated from the plurality of dialogue interactions based on the set of dialogue-specific embeddings by a clustering machine learning model. A representative dialog instance for each dialog cluster in the set of dialog clusters is then determined. A plurality of vector stores is then generated for the set of dialogue clusters associated with the LLM based on the representative dialogue instance for each dialogue cluster. A set of proxy replies is then generated for client conversational interactions initiated by the client and the proxy based on the plurality of vectors associated with the LLM.

Inventors

• S.K. Chinum Gary
• M. Duns
• A. Y. Puladan
• G. S. Mohanty

Assignees

• 马士基股份公司

Dates

Publication Date

20260512

Application Date

20240930

Priority Date

20240207

Claims (20)

1. 1. A computer-implemented method, comprising: Accessing, by a server system, a plurality of dialog-specific features corresponding to each of a plurality of dialog interactions from a database associated with the server system; Generating, by a Large Language Model (LLM) associated with the server system, a set of dialog-specific embeddings for each dialog interaction based at least in part on a plurality of dialog-specific features corresponding to the dialog interaction, each dialog-specific embedment in the set of dialog-specific embeddings corresponding to each dialog interaction; Generating, by a clustering machine learning model associated with the server system, a set of conversational clusters from the plurality of conversational interactions based at least in part on the set of conversation-specific embeddings; determining, by the server system, a representative dialog instance for each dialog cluster in the set of dialog clusters, the representative dialog instance being indicative dialog interactions of individual dialog clusters, and Generating, by the server system, a plurality of vector stores for the set of dialog clusters associated with the LLM based at least in part on the representative dialog instance of each dialog cluster, each vector store of the plurality of vector stores including the representative dialog instance for the corresponding dialog cluster and metadata corresponding to the representative dialog instance.
2. 2. The computer-implemented method of claim 1, wherein determining the representative dialog instance comprises: Determining, by the server system, a centroid of each dialog cluster based at least in part on the set of dialog-specific embeddings, and The representative dialog instance is selected by the server system for each dialog cluster based at least in part on the centroid of each dialog cluster.
3. 3. The computer-implemented method of claim 1, further comprising generating a set of proxy replies for client conversational interactions initiated by a client and a proxy based at least in part on the plurality of vectors stored in association with the LLM.
4. 4. The computer-implemented method of claim 3, wherein generating the set of proxy replies comprises: receiving, by a communication interface associated with the server system, the customer conversational interaction from the customer, the customer conversational interaction comprising a plurality of customer interaction features; Generating, by the server system, a dialog prompt based at least in part on the plurality of client interaction features, the dialog prompt including a dialog domain field, a dialog topic field, a client query field, and a proxy reply field; Determining, by the server system, a set of similar vector stores of the plurality of vector stores based at least in part on at least a threshold similarity between the dialog prompt and the plurality of vector stores, and The set of proxy replies is generated by the server system for the proxy reply field based at least in part on the set of similarity vector stores.
5. 5. The computer-implemented method of claim 4, further comprising: A set of custom proxy replies are prepared by the LLM based at least in part on the set of proxy replies and the dialog prompt.
6. 6. The computer-implemented method of claim 5, further comprising ranking the set of custom proxy replies based at least in part on metadata present in each similarity vector store used to generate the set of custom proxy replies.
7. 7. The computer-implemented method of any of claims 5 to 6, wherein generating the set of proxy replies comprises: validating, by the server system, a subset of similarity vector stores in the set of similarity vector stores based at least in part on the dialog field and the set of similarity vector stores, and The set of proxy replies is generated by the LLM for the proxy reply field based at least in part on the similarity vector store subset.
8. 8. The computer-implemented method of any of claims 5 to 7, further comprising: Display of a Graphical User Interface (GUI) on an electronic device associated with the agent is facilitated by the server system, the GUI indicating the set of custom agent replies to the agent.
9. 9. The computer-implemented method of claim 8, further comprising: receiving, by the server system, a selection of a custom proxy reply from the proxy from the set of custom proxy replies, and A selected custom proxy reply is transmitted by the server system to the client.
10. 10. The computer-implemented method of claim 1, further comprising: Accessing, by the server system, a historical dialog data set from the database, the historical dialog data set including the plurality of conversational interactions between a plurality of clients and a plurality of agents, and The plurality of dialog-specific features corresponding to each dialog interaction are determined by the server system based at least in part on the plurality of dialog interactions.
11. 11. The computer-implemented method of claim 10, wherein determining the plurality of dialog-specific features comprises: Determining one or more client session messages and one or more proxy session messages for each session interaction; Generating a plurality of conversation pairs from the plurality of conversational interactions based at least in part on concatenating one or more client conversation messages and one or more proxy conversation messages for each conversational interaction, and The plurality of dialog specific features are extracted in a predefined format based at least in part on the plurality of dialog pairs.
12. 12. The computer-implemented method of any of claims 1 to 11, wherein the metadata includes a cluster size and a dialog domain.
13. 13. The computer-implemented method of claim 1, wherein the LLM is fine-tuned based at least on the plurality of session-specific features associated with a historical session data set stored in the database, and wherein the historical session data set is updated by new session interactions between the plurality of clients and the plurality of agents.
14. 14. The computer-implemented method of claim 13, further comprising: generating, by the server system, a plurality of masked conversational interactions based at least in part on masking predefined information in the plurality of conversational interactions, and The LLM is trimmed by the server system based at least in part on the plurality of masked conversational interactions.
15. 15. A server system, comprising: A communication interface; A memory configured to store instructions, and A processor in communication with the communication interface and the memory, the processor configured to execute the instructions stored in the memory, thereby causing the server system to at least partially perform: accessing a plurality of dialog-specific features corresponding to each of a plurality of dialog interactions from a database associated with the server system; Generating, by a Large Language Model (LLM) associated with the server system, a set of dialog-specific embeddings for each dialog interaction based at least in part on a plurality of dialog-specific features corresponding to the dialog interaction, each dialog-specific embedment in the set of dialog-specific embeddings corresponding to each dialog interaction; Generating, by a clustering machine learning model associated with the server system, a set of conversational clusters from the plurality of conversational interactions based at least in part on the set of conversation-specific embeddings; determining a representative dialog instance for each dialog cluster in the set of dialog clusters, the representative dialog instance being indicative dialog interactions of an individual dialog cluster, and A plurality of vector stores are generated for the set of dialog clusters associated with the LLM based at least in part on the representative dialog instance for each dialog cluster, each vector store of the plurality of vector stores including the representative dialog instance for the corresponding dialog cluster and metadata corresponding to the representative dialog instance.
16. 16. The server system of claim 15, wherein to determine the representative dialog instance, the server system is caused, at least in part, to: Determining a centroid of each dialog cluster based at least in part on the set of dialog-specific embeddings, and The representative dialog instance is selected for each dialog cluster based at least in part on the centroid of each dialog cluster.
17. 17. The server system of claim 15, wherein the server system is further caused, at least in part, to generate a set of proxy replies for client conversational interactions initiated by a client with a proxy based, at least in part, on the plurality of vectors associated with the LLM.
18. 18. The server system of claim 17, wherein to generate the set of proxy replies, the server system is caused, at least in part, to: receiving, by the communication interface, the customer conversational interaction from the customer, the customer conversational interaction comprising a plurality of customer interaction features; Generating a dialog prompt based at least in part on the plurality of client interaction features, the dialog prompt including a dialog domain field, a dialog topic field, a client query field, and a proxy reply field; determining a set of similarity vector stores of the plurality of vector stores based at least in part on at least a threshold similarity between the dialog prompt and the plurality of vector stores, and The set of proxy replies is generated for the proxy reply field based at least in part on the set of similarity vector stores.
19. 19. The computer-implemented method of claim 18, wherein the server system is further caused, at least in part, to: A set of custom proxy replies are prepared by the LLM based at least in part on the set of proxy replies and the dialog prompt.
20. 20. A non-transitory computer-readable storage medium comprising computer-executable instructions that, when executed by at least a processor of a server system, cause the server system to perform a method comprising: accessing a plurality of dialog-specific features corresponding to each of a plurality of dialog interactions from a database associated with the server system; Generating, by a Large Language Model (LLM) associated with the server system, a set of dialog-specific embeddings for each dialog interaction based at least in part on a plurality of dialog-specific features corresponding to the dialog interaction, each dialog-specific embedment in the set of dialog-specific embeddings corresponding to each dialog interaction; Generating, by a clustering machine learning model associated with the server system, a set of conversational clusters from the plurality of conversational interactions based at least in part on the set of conversation-specific embeddings; determining a representative dialog instance for each dialog cluster in the set of dialog clusters, the representative dialog instance being indicative dialog interactions of an individual dialog cluster, and A plurality of vector stores are generated for the set of dialog clusters associated with the LLM based at least in part on the representative dialog instance for each dialog cluster, each vector store of the plurality of vector stores including the representative dialog instance for the corresponding dialog cluster and metadata corresponding to the representative dialog instance.

Description

Method and system for generating customized replies to conversational interactions Technical Field The present disclosure relates to the field of generating replies in a dialog ecosystem, and more particularly to an electronic method and complex processing system for generating custom reply suggestions for agents participating in conversational interactions with clients or users. Background For any business organization that wishes to effectively serve its customers, a conversation between the customer (or customer) and the customer support representative (or agent) plays a vital role. Each enterprise needs to listen to their customers' opinion and solve any queries they may encounter while ensuring a good customer experience. To ensure a good customer experience, businesses typically employ a large community of agents to be responsible for interacting with customers to resolve their queries. Such interactions between the customer and the agent are often referred to as conversational interactions or conversational interactions. The preferred way to perform conversational interactions is written digital communication. Written digital communication is an effective form of communication that allows for better understanding and ease of recording. It allows customers or agents to query at any time for future reference. Examples of written digital communications include electronic mail (email), chat-based conversations, and the like. In a written digital communication-based environment, a customer typically initiates a conversation with an agent seeking a solution to his or her query. After receiving the query, the agent must understand the problem faced by the customer and find a solution to his/her problem. It is critical that agents solve customer queries in time to ensure a good user experience. In general, queries from customers may be received by a server system maintained by an organization, and customer support representatives may obtain notifications or alerts on which they may respond to the queries. As the subscription volume rises (such as during the busy season/peak hours), queries may multiply, which may result in an increase in processing load on the server system. Further, as the number of queries increases, agents may require more time to resolve the queries, and thus the number of outstanding queries in the server system may rise. The increase in the number of outstanding queries may consume additional processing resources and/or memory resources of the server system and adversely affect the processing speed of the server system. Thus, there is a need to find technical solutions that enable agents to timely address customer queries while ensuring that the provided solutions are correct and meet customer needs. Disclosure of Invention Techniques are needed to overcome one or more of the limitations described above, such as adverse effects on the processing speed of the server system and poor customer user experience due to delays in response of the agents of the organization. Various embodiments of the present disclosure provide methods and systems for ensuring timely resolution of queries by generating custom reply suggestions for agents participating in conversational interactions with clients. Various embodiments of the present disclosure describe a computing device and method that enables an agent to select from a set of custom reply suggestions while also providing summaries of those suggestions thereby improving response time to a query. The improvement in response time allows for a reduction in the number of outstanding queries in the server system, thereby reducing processing load and mitigating its adverse impact on the processing speed of the server system. Due to the improved time of response to the query, the customer experience is also optimized as the customer's query is resolved faster. Further, custom reply suggestions alleviate the burden of understanding customer issues faced by agents, thus reducing the complexity of addressing customer queries. To achieve the above and other objects of the present disclosure, in one aspect, a computer-implemented method for generating customized reply suggestions for agents engaged in conversational interactions with a client is disclosed. The computer-implemented method is performed by a server system. The method includes accessing a plurality of dialog-specific features corresponding to each conversational interaction from a database associated with the server system. The method further includes generating, by a Large Language Model (LLM) associated with the server system, a dialog-specific embedding for each dialog interaction based at least in part on a plurality of dialog-specific features corresponding to each dialog interaction. The computer-implemented method further includes generating, by a clustered machine learning model associated with the server system, a set of conversational clusters from the plurality of conversational interac