Training Corpus Generation from User Behavior Logs

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The existing speech recognition model iteration period is excessively long due to manual labeling, and resource consumption is severe, with many invalid labels resulting from randomly extracted audios.

Innovation Solution

Automatically mine positive and negative feedback samples from user behavior logs, using correctly and incorrectly recognized audio-text pairs to create a targeted training corpus, reducing the need for manual labeling and improving model efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If manual labeling is used to generate training corpus, then the speech recognition model can be trained, but the iteration period becomes excessively long and resource consumption increases

Engineering Contradiction:
Improvespeech recognition model training qualityVSAvoiditeration period
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The system automatically generates training corpus by mining user behavior logs and speech recognition results without requiring manual labeling. The system serves itself by using its own operational data (user speeches and recognition results) to create training samples, eliminating the need for external manual intervention and significantly reducing the iteration period

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system implements feedback loops where user behavior logs and speech recognition results are continuously collected and used to generate new training corpus. The recognition results feed back into the training process, creating a closed-loop system that automatically improves the model over time without manual intervention

Inventive Principle:
Principle #23Feedback

2Quantity of substance

If random audios are extracted for manual labeling, then training data can be obtained, but many invalid labels are produced reducing effectiveness

Engineering Contradiction:
Improvetraining data volumeVSAvoidlabeling accuracy
Core Design Contradiction:
Quantity of substanceVSManufacturing precision

Solution Approach 1:

Instead of uniformly random sampling, the system selectively mines training data from specific user behavior scenarios where speech recognition actually occurs. This targeted approach ensures that training samples have high quality and relevance, improving labeling precision while maintaining adequate quantity

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The system uses its own speech recognition results from actual user interactions as the basis for generating training labels. Since these are the system's own recognition outputs, they inherently have validated quality and relevance, eliminating the problem of invalid labels that occurs with random audio extraction

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3709295B1Methods, apparatuses, and storage media for generating training corpus
Publication Date: 2022.11.30 BEIJING BAIDU NETCOM SCI & TECH CO LTD
  • EP3709295B1 patent drawingFigure 1~2a
  • EP3709295B1 patent drawingFigure 2b
  • EP3709295B1 patent drawingFigure 2c

AI summary

The present disclosure provides methods, computing devices, and storage media for generating a training corpus. The method includes: mining out pieces of data from user behavior logs associated with a target application, each piece of data including a first behavior log and a second behavior log, the first behavior log including a user speech and a corresponding speech recognition result, a second behavior log belonging to the same user as the first behavior log and time-dependent with the first behavior log; and determining the user speech and the corresponding speech recognition result in each piece of data as a positive feedback sample or a negative feedback sample, based on the first behavior log and second behavior log.