ASR Context Biasing with Text Injection for Named Entity Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based ASR models tend to overfit training data, leading to difficulties in generalizing unseen data, especially when the training data is not extensive, and struggle to accurately recognize named entities without increasing word error rates or latency.
Innovation Solution
A training process that injects contextually biased data into the ASR model by pairing unspoken textual utterances with carrier phrases, generating higher-order feature representations, and using probability distributions to improve model training on unpaired text data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If query expansion techniques are used to improve search recall, then more relevant documents are retrieved, but false positives increase and precision deteriorates
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different queries. Instead of uniformly expanding all queries, the system identifies and expands only those queries that meet specific criteria (e.g., queries with certain characteristics or from specific contexts), thereby locally applying expansion only where beneficial. This selective approach maintains precision while improving recall for targeted queries.
Solution Approach 2:
The system dynamically changes parameters of query expansion based on query characteristics, user behavior, and contextual factors. By adjusting expansion intensity, selection criteria, and boosting factors according to specific parameters, the system optimizes the balance between recall and precision for different search scenarios without uniformly degrading precision across all queries.
2Measurement precision
If contextual information is injected into search queries, then search accuracy and relevance are improved, but system complexity increases
Solution Approach 1:
The patent implements a universal contextual injection mechanism that handles multiple types of context (user profile, session history, device information, location) through a single integrated system. This multi-functional approach consolidates what would otherwise require separate processing modules, improving search accuracy while minimizing the increase in system complexity through consolidation rather than proliferation of components.
Solution Approach 2:
The system automatically retrieves and injects contextual information without requiring manual intervention or complex configuration. Contextual data is self-acquired from available sources (user profiles, session data, device attributes) and self-applied to queries, reducing the operational complexity despite the enhanced functionality.
3Reliability
If multiple context types are integrated into search processing, then user experience and result relevance are enhanced, but processing time increases
Solution Approach 1:
The patent pre-loads and caches contextual information (user profiles, preferences, device attributes) before they are needed for query processing. By having contextual data readily available in advance, the system avoids time-consuming retrieval operations during actual search execution, thereby enhancing result relevance while minimizing additional processing time.
Solution Approach 2:
The system selectively applies contextual injection only to queries that benefit from it, rather than uniformly processing all queries with full contextual expansion. By applying context partially—only where it adds value—the system enhances relevance for targeted queries while avoiding the processing overhead of applying full contextual analysis to every search request.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
A method (600) includes receiving context biasing data (510) that includes a set of unspoken textual utterances (320) corresponding to a particular context (512). The method also includes obtaining a list of carrier phrases (520) associated with the particular context. For each respective unspoken textual utterance, the method includes generating a corresponding training data pair (532) that includes the respective unspoken textual utterance and a carrier phrase. For each respective training data pair, the method includes tokenizing the respective training data pair into a sequence of sub-word units (402), generating a first higher order textual feature representation (203) for a corresponding sub-word unit, receiving the first higher order textual feature representation, and generating a first probability distribution (253) over possible text units. The method also includes training a speech recognition model (200) based on the first probability distribution over possible text units.