CATEGORIZATION OF USER INTERACTIONS INTO PREDEFINED HIERARCHICAL CATEGORIES

• US 20140229408A1
• Filed: 02/14/2014
• Published: 08/14/2014
• Est. Priority Date: 02/14/2013
• Status: Active Grant

• Current Assignee: [24]7.ai, Inc.
• Original Assignee: 24/7 Customer Incorporated
• Inventors: VIJAYARAGHAVAN, Ravi, Srivastava, Vaibhav, Sri, R. Mathangi, Hardeniya, Nitin Kumar
• Granted Patent: US 9,626,629 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: G06F 16/353 into predefined classes
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 5/02 Knowledge representation; Symbolic representation
• CPC Class Codes: G06N 5/025 Extracting rules from data
• CPC Class Codes: G06N 5/04 Inference or reasoning models
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks
• CPC Class Codes: G06Q 10/10 Office automation; Time management

Abstract

User interactions are categorized into predefined hierarchical categories by classifying user interactions, such as queries, during a user interaction session by labeling text data into predefined hierarchical categories, and building a scoring model. The scoring model is then executed on untagged user interaction data to classify the user interactions into either action-based or information-based interactions.

Claims

1. A computer implemented method for categorization of user interactions into predefined hierarchical categories, comprising:

providing a processor configured for classifying user interactions during a user interaction session by labeling text data into predefined hierarchical categories, and building a scoring model; and

said processor executing said scoring model on untagged user interaction data to classify said user interactions into either action-based or information-based interactions.

2. The method of claim 1, further comprising:

during scoring model execution, applying a trained model to interaction data to classify and label interaction related documents

3. The method of claim 1, further comprising:

said processor providing a received unlabeled document to a relevant line extractor; and

said relevant line extractor extracting from said unlabeled document only those lines in the text of said unlabeled document that are relevant.

4. The method of claim 3, further comprising:

inputting an output of said relevant line extractor and said scoring model to an information retrieval (IR) engine;

responsive thereto, said IR engine outputting data matrices.

5. The method of claim 4, further comprising

preprocessing the lines of text from said relevant line extractor by any of reshaping data;

masking text patterns which comprise any of different date patterns, digit patterns, phone numbers, credit card numbers, company names, and URLs;

converting to lower case;

removing numbers;

removing punctuation;

tokenization;

stemming;

part-of-speech tagging; and

lemmatization.

6. The method of claim 4, further comprising:

said IR engine evaluating one or more rules to score every document against said scoring model; and

based on rules that hit each text data point, giving each such data point a score for each of said categories;

wherein a data matrix is produced for each level in a category tree.

7. The method of claim 6, further comprising:

a classifier receiving said data matrix for each level in a category tree;

said classifier using at least one classification algorithm to label said data matrices;

wherein said labeled data matrices yield labeled documents that comprise cleansed, transformed, and structured representations of an original, transcribed or typed and unstructured text interaction history.

8. The method of claim 7, further comprising:

combining said representations with additional structured data.

9. The method of claim 7, further comprising:

building any of said classifier and said model using a supervised or an unsupervised approach.

10. The method of claim 1, further comprising:

during a scoring model building phase, said processor receiving labeled response variables and a set of structured input features; and

during a model testing or model validation phase, using said model for model scoring.

11. The method of claim 10, further comprising:

during a model scoring phase, for a given set of structured input features, using said model to predict a response variable.

12. The method of claim 1, further comprising:

using said model across different businesses in a same or similar domain, even if said businesses are not part of a training data set.

13. The method of claim 1, further comprising:

during scoring in a model execution phase, said processor receiving said model and untagged data; and

said processor classifying said untagged user interactions as either action-based or information-based interactions.

14. The method of claim 13, further comprising:

a relevant line extractor receiving untagged data; and

said relevant line extractor outputting data to a model execution module, said model extraction module using said model and a rule extraction system to parse user interaction sessions and extract relevant or important features therefrom to build machine learning models based on tagged data;

wherein said features comprise any of most frequent n-grams;

most discriminatory n-grams;

rules comprising any of AND, OR, or proximity rules;

any structured data comprising any of handle times and delays in response.

15. The method of claim 13, further comprising:

categorizing said user interactions into sub-categories of an information-based or action-based type of chat.

16. The method of claim 14, further comprising:

said rule extraction system extracting different features that qualify text;

said rule extraction system checking distribution of said extracted features across different categories hierarchically; and

based on relevance of said categories, said rule extraction system providing scores.

17. An apparatus for categorization of user interactions into predefined hierarchical categories, comprising:

a processor programmed with machine instructions which, when executed classify user interactions during a user interaction session by labeling text data points into predefined hierarchical categories, segregating issues, and building a scoring model; and

said processor executing said scoring model based upon data within said scoring model and untagged user interaction data to classify said user interactions into either action-based or information-based interactions.

18. A method for query categorization into predefined hierarchical categories, comprising:

an agent receiving from a user either of an information or action oriented request during an interaction session;

providing a processor configured to implement a text filter for filtering relevant sections of text in said interaction session;

said processor configured to implement text parser for parsing said text and for extracting important features from the interaction session;

said processor building a category tree to categorize different kinds of issues in said interaction session to segregate said issues into different categories;

during a training phase;

once a domain specific category tree is built, defining categories and creating rules, wherein said rules help to extract relevant features during said interaction session;

once said rules are created, extracting relevant features;

during a pre-processing session, generalizing said text by replacing specific words with standard words;

creating a scoring model; and

generating a matrix based on said score;

wherein said matrix is given a score based on the relevance of the matrix generated; and

wherein said score obtained is based on a comparison between the rules created and the category tree.