US-12626267-B2 - Omnichannel data processing and analysis

US12626267B2US 12626267 B2US12626267 B2US 12626267B2US-12626267-B2

Abstract

Natural Language Processing (NLP) techniques are used to facilitate inferring actionable insights from interactions. Customer data from various communication channels can be used to determine, for example, voice-related aspects from sentiment analysis, intent analysis, Semantic Conscious Word Extraction (SCWE), emotion analysis, and contextual summarization. These analysis results can also be used, along with customer profile information, for deriving deep customer insights. The deep customer insight driven analysis can include multivariate customer fragmentation, setback accountability analysis, potential contender analysis, and propulsive business planning.

Inventors

Sai Prasanth Singavarapu
Rhea Thomas
Priyanka Satish
V Vikash Kumaran
Ram Prakash S
Sanjeev Kumar Srinivasan Khannan
KARTHICK S
Hari Bharathi A
Azarudeen Mohamed Ibrahim

Assignees

Zoho Corporation Private Limited

Dates

Publication Date: 20260512
Application Date: 20230109
Priority Date: 20220112

Claims (16)

1 . A method comprising: collecting a dataset, wherein the dataset includes a first pulse from a first channel and a second pulse from a second channel; performing preprocessing on the dataset, wherein word embedding is done for the preprocessed data to convert it into vectors; generating respective results for a plurality of different analysis models, wherein the plurality of different analysis models includes: pre-training a sentiment analysis model, fine-tuned using Customer Relationship Management (CRM) data to predict sentiment, that classifies the first pulse and the second pulse into respective ones of positive, negative, or neutral sentiment, wherein the sentiment analysis model comprises a deep transfer learning model, and wherein the fine-tuning includes tokenizing the CRM data into a readable format accepted by the pre-trained sentiment analysis model; an intent analysis model, fine-tuned using CRM data to predict intent, that classifies the first pulse and the second pulse into respective ones of complaint, feedback, query, request, or purchase-related intention; a Semantic Conscious Word Extraction (SCWE) model trained to extract semantic conscious keywords/key-phrases from pulses, wherein, in operation, the pulses include the first pulse and the second pulse; utilizing the respective results of the plurality of different analysis models to perform deep customer insight driven analysis, wherein the deep customer insight driven analysis includes: propulsive business planning to achieve an agenda; filtering, based on the agenda, customers who have met the agenda; considering the conversations made by a particular type of entity to the filtered customers, wherein the customer conversations include a first conversation and a second conversation, and wherein the first pulse is associated with the first conversation, and the second pulse is associated with the second conversation; determining aspects/information about the filtered customers including activity-related aspects, wherein the activity-related aspects identify a number of times that keywords or key-phrases with the intent are mentioned in the conversations with the filtered customers; increasing or decreasing, based on feedback provided by the customers, any of the sentiment, the keywords, or the key-phrases with the intent; constructing a pattern from the determined aspects/information to achieve the agenda.
2 . The method of claim 1 , wherein the plurality of different analysis models includes the sentiment analysis model and the intent analysis model, and wherein the sentiment analysis model is trained by distilling Bidirectional Encoder Representations from Transformers (BERT) base.
3 . The method of claim 1 , wherein the plurality of different analysis models includes an emotion analysis model that classifies the first pulse and the second pulse into respective ones of happiness, enthusiasm, discontentment, frustration, trust, confusion, gratitude, or neutral emotion.
4 . The method of claim 1 , wherein the plurality of different analysis models includes a contextual summarization model fine-tuned using CRM data to produce a terse summary of the first pulse incorporating customer intention.
5 . The method of claim 1 , wherein the deep customer insight driven analysis is performed using a setback accountability analysis engine that identifies plausible reasons for customer setbacks and influential factors for the customer setbacks.
6 . The method of claim 5 , comprising performing, by the setback accountability analysis engine, Aspect-Based Sentiment Analysis (ABSA) to identify customer sentiments associated with specific aspects of products or services.
7 . The method of claim 1 , wherein the deep customer insight driven analysis is performed using a potential contender analysis engine to provide analysis for a significant factor determination engine to determine, for a keyword associated with the first pulse, a number of contenders represented in a CRM contender list that are influenced by the keyword.
8 . The method of claim 1 , wherein the deep customer insight driven analysis is performed using a potential contender analysis engine to provide analysis for a significant contender determination engine to determine a contender represented in a CRM contender list is a challenging contender due to multiple customers, including a first customer and a second customer, with positive sentiment, wherein the first pulse is associated with the first customer and the second pulse is associated with the second customer, and to identify keywords of the challenging contender from positive sentiment of the multiple customers.
9 . The method of claim 1 , further comprising: utilizing the respective results of the plurality of different analysis models to perform deep customer insight driven analysis, wherein the deep customer insight driven analysis includes a potential contender analysis to identify potential switching customers; identifying Churned out Customers (CoC) due to contender attraction; filtering CoC conversation in which contenders are mentioned; determining a limit on conversation with negative sentiment before churn; identifying keywords that represent factors behind the CoC using the SCWE model; and identifying customers before churn using multivariate customer fragmentation based on CoC characteristic and concentrating on customers by improving identified factors before reaching the limit on conversation.
10 . The method of claim 9 , wherein the deep customer insight driven analysis is performed using a potential contender analysis engine to provide, via contextual summarization that considers negative sentiment customer conversations where a contender is mentioned, a reason why a customer has switched to the contender.
11 . The method of claim 9 , wherein the plurality of different analysis models includes a Semantic Conscious Word Extraction (SCWE) model trained to extract semantic conscious keywords/key-phrases from pulses and prioritize the keywords/key-phrases based on objectives of an organization.
12 . The method of claim 11 , wherein the SCWE model is used to prioritize keywords/key-phrases using customized sorting, wherein the sorting order is determined with respect to the priority of at least one of sentiment, intent and emotion provided by an organization.
13 . The method of claim 1 , wherein the constructing a pattern comprises: determining intent, sentiment and keywords from interactions with filtered customers; determining profile-related aspects of the filtered customers; determining activity-related aspects performed by an organization to the filtered customers; constructing a pattern from the sentiment and keywords, the profile-related aspects, and the activity-related aspects.
14 . The method of claim 1 , wherein the deep customer insight driven analysis is performed using a multivariate fragmentation engine that clusters the customers based on the results from the different analysis models.
15 . A system comprising: one or more processors; and memory storing instructions that, when executed by the one or more processors, cause the system to perform: collecting a dataset, wherein the dataset includes a first pulse from a first channel and a second pulse from a second channel; performing preprocessing on the dataset, wherein word embedding is done for the preprocessed data to convert it into vectors; generating respective results for a plurality of different analysis models, wherein the plurality of different analysis models includes: pre-training a sentiment analysis model, fine-tuned using Customer Relationship Management (CRM) data to predict sentiment, that classifies the first pulse and the second pulse into respective ones of positive, negative, or neutral sentiment, wherein the sentiment analysis model comprises a deep transfer learning model, and wherein the fine-tuning includes tokenizing the CRM data into a readable format accepted by the pre-trained sentiment analysis model; an intent analysis model, fine-tuned using CRM data to predict intent, that classifies the first pulse and the second pulse into respective ones of complaint, feedback, query, request, or purchase-related intention; a Semantic Conscious Word Extraction (SCWE) model trained to extract semantic conscious keywords/key-phrases from pulses, wherein, in operation, the pulses include the first pulse and the second pulse; utilizing the respective results of the plurality of different analysis models to perform deep customer insight driven analysis, wherein the deep customer insight driven analysis includes: propulsive business planning to achieve an agenda; filtering, based on the agenda, customers who have met the agenda; considering the conversations made by a particular type of entity to the filtered customers, wherein the customer conversations include a first conversation and a second conversation, and wherein the first pulse is associated with the first conversation, and the second pulse is associated with the second conversation; determining aspects/information about the filtered customers including activity-related aspects, wherein the activity-related aspects identify a number of times that keywords or key-phrases with the intent are mentioned in the conversations with the filtered customers; increasing or decreasing, based on feedback provided by the customers, any of the sentiment, the keywords, or the key-phrases with the intent; constructing a pattern from the determined aspects/information to achieve the agenda.
16 . A system comprising: a means for collecting a dataset, wherein the dataset includes a first pulse from a first channel and a second pulse from a second channel; a means for performing preprocessing on the dataset, wherein word embedding is done for the preprocessed data to convert it into vectors; a means for generating respective results for a plurality of different analysis models, wherein the plurality of different analysis models includes: pre-training a sentiment analysis model, fine-tuned using Customer Relationship Management (CRM) data to predict sentiment, that classifies the first pulse and the second pulse into respective ones of positive, negative, or neutral sentiment, wherein the sentiment analysis model comprises a deep transfer learning model, and wherein the fine-tuning includes tokenizing the CRM data into a readable format accepted by the pre-trained sentiment analysis model; an intent analysis model, fine-tuned using CRM data to predict intent, that classifies the first pulse and the second pulse into respective ones of complaint, feedback, query, request, or purchase-related intention; a Semantic Conscious Word Extraction (SCWE) model trained to extract semantic conscious keywords/key-phrases from pulses, wherein, in operation, the pulses include the first pulse and the second pulse; a means for utilizing the respective results of the plurality of different analysis models to perform deep customer insight driven analysis, wherein the deep customer insight driven analysis includes: propulsive business planning to achieve an agenda; filtering, based on the agenda, customers who have met the agenda; considering the conversations made by a particular type of entity to the filtered customers, wherein the customer conversations include a first conversation and a second conversation, and wherein the first pulse is associated with the first conversation, and the second pulse is associated with the second conversation; determining aspects/information about the filtered customers including activity-related aspects, wherein the activity-related aspects identify a number of times that keywords or key-phrases with the intent are mentioned in the conversations with the filtered customers; increasing or decreasing, based on feedback provided by the customers, any of the sentiment, the keywords, or the key-phrases with the intent; constructing a pattern from the determined aspects/information to achieve the agenda.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS The present application claims priority to U.S. Provisional Patent Application No. 63/322,679 filed Mar. 23, 2022, entitled “COGNITION ASSISTED OMNICHANNEL DIGESTER”, and Indian Provisional Patent Application No. 202241001721 filed Jan. 12, 2022, entitled “COGNITION ASSISTED OMNICHANNEL DIGESTER”, all of which are hereby incorporated by reference herein. TECHNICAL FIELD Embodiments of the present disclosure are related, in general, to Natural Language Processing (NLP) and more particularly, but not exclusively to analyzing unstructured data from different channels. BACKGROUND Data is generated at a very rapid pace. Data can include customer information in the form of text and speech. Customer interactions and their corresponding data are part of a typical CRM system. Such interactions are received through different channels like e-mail, call, survey, service desk ticket and social media. SUMMARY An automatic analysis on unstructured data to derive deep customer insights is provided in an Omnichannel Intelligent Pulse Digester (OIPD) framework. Actionable insights are derived from customer data obtained through multiple communication channels for calls, ticket management, surveys, and email, to name several. Voice-related aspects can be extracted from the customer data using Natural Language Processing (NLP) models such as sentiment analysis, intent analysis, Semantic Conscious Word Extraction (SCWE), emotion analysis, and contextual summarization. Deep customer analysis that includes multivariate customer fragmentation, setback accountability analysis, potential contender analysis, and propulsive business planning can be performed considering, for example, voice-related aspects. BRIEF DESCRIPTION OF DRAWINGS FIG. 1 depicts an Omnichannel Intelligent Pulse Digester (OIPD) system. FIG. 2 depicts a flow diagram of an example of a data analysis pipeline in an OIPD system. FIG. 3 depicts a flow diagram of an example of Semantic Conscious Word Extraction (SCWE) process flow for Similarity Score (SS) based sorting. FIG. 4 depicts a flowchart of an example of customized keyword sorting. FIG. 5 depicts a flowchart of an example of contextual summarization. FIG. 6 depicts a flow diagram of an example of customer fragmentation analysis. FIG. 7 depicts a flowchart of an example of setback accountability analysis. FIG. 8 depicts a flowchart of an example of Aspect-Based Sentiment Analysis (ABSA). FIG. 9 depicts a diagram of an example of a potential contender analysis system. FIG. 10 depicts a flowchart of an example of significant factor determination. FIG. 11 depicts a flowchart of an example of significant contender determination. FIG. 12 depicts a flowchart of an example of possible switching customer identification. FIG. 13 depicts a flowchart of an example of providing reasoning for potential contender analysis using contextual summarization. FIG. 14 depicts a diagram of an example of a propulsive business planning system. FIG. 15 depicts a flowchart of a method of constructing a pattern to achieve an agenda. LIST OF TABLES Table 1 depicts results of voice-related aspect models for a customer conversation. Table 2 shows analysis results for customer conversations. DETAILED DESCRIPTION Interactions with, for example, a customer contain information in an unstructured format that can includes surveys, issues, and other business process related details. Hence, it is difficult to draw inferences from the data in a superficial manner, such as by looking at it. An Omnichannel Intelligent Pulse Digester (OIPD) framework provides an automated way of analyzing the unstructured data using Natural Language Processing (NLP) techniques to derive significant insights from it. Omnichannel, as used in this paper, is intended to represent multiple different sources of a pulse, such as communication channel (e.g., email, voice, or the like) or a data channel (e.g., a survey, desk data, transaction data, product or service data, industry data, or the like). A pulse, as used in this paper, is an activity associated with an item of interest, such as a communication instance with a customer, a transaction, data about a product or service, data about an industry, or some other data item or activity. Digestion, as used in this paper, is intended to indicate various related pulses (even if the relationship cannot be easily articulated by a human) are considered in the aggregate; and intelligent, as used in this paper, is intended to indicate the aggregation yields an insight related to the aggregated pulses from the various channels. Customer insights, for example, are useful to deliver business information in an anticipated, relevant, and timely manner. It can facilitate better pricing, promotion, and management-related decisions. Advantages include turning customer data into insights and turning insights into loyalty, profitability, and sustainable growth. FIG. 1 depicts a diagram 100 of an example