Dialog History Embedding for Spoken Language Understanding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional spoken language understanding (SLU) systems lack effective integration of dialog history, leading to ambiguity in context-dependent spoken task-oriented conversations, as they process each conversation turn independently without considering previous information.
Innovation Solution
A spoken language understanding system that encodes dialog history into embeddings using processors and memory devices, integrating this information with speech features to train neural network models for improved understanding, including the use of Bidirectional Encoder Representations from Transformers (BERT) for encoding dialog history and appending embeddings to input acoustic features for enhanced performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If end-to-end SLU systems process each turn independently, then processing speed is improved, but accuracy in context-dependent conversations deteriorates
Solution Approach 1:
The patent adds a temporal dimension by incorporating dialog history embeddings into the SLU model. Instead of processing only the current utterance in isolation, the system processes the current utterance together with encoded representations of previous turns, transforming the input from a single-turn feature vector to a multi-turn contextualized representation that resolves ambiguity while maintaining efficient processing.
2Measurement precision
If dialog history is integrated into SLU systems, then recognition accuracy is improved, but system complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-encoding dialog history into embeddings using a separate encoder model before feeding it into the main SLU neural network. This preprocessing step transforms complex historical conversation data into compact, fixed-dimensional representations, simplifying the main model's task and reducing overall system complexity while maintaining high accuracy.
Solution Approach 2:
The patent implements a nested architecture where the dialog history encoder is embedded within the overall SLU system. The encoder processes historical turns and produces embeddings that are nested as additional input features into the main SLU model, creating a hierarchical structure that manages complexity through modular organization.
3Measurement precision
If dialog history embeddings are appended to acoustic features, then understanding accuracy is improved, but computational requirements increase
Solution Approach 1:
The patent extracts only the essential contextual information from dialog history by encoding it into compact embeddings with fixed dimensionality. Instead of processing the full raw history of all previous turns, the system extracts salient features and compresses them into efficient representations that are appended to acoustic features, reducing computational load while preserving critical contextual information for accuracy.
Data Source
AI summary
Audio signals representing a current utterance in a conversation and a dialog history including at least information associated with past utterances corresponding to the current utterance in the conversation can be received. The dialog history can be encoded into an embedding. A spoken language understanding neural network model can be trained to perform a spoken language understanding task based on input features including at least speech features associated with the received audio signals and the embedding. An encoder can also be trained to encode a given dialog history into an embedding. The spoken language understanding task can include predicting a dialog action of an utterance. The spoken language understanding task can include predicting a dialog intent or overall topic of the conversation.