Bidirectional RNN Spoken Language Understanding Joint Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Spoken language understanding systems for task-oriented virtual agents currently process user utterances sequentially, limiting efficiency and accuracy in classifying speech acts, identifying intent, and extracting semantic constituents, and require separate training for each function, which increases computational resources and time.
Innovation Solution
A system and method using bidirectional and forward attention-based Recurrent Neural Networks (RNNs) to jointly perform the functions of classifying speech acts, identifying intent, and extracting semantic constituents, sharing a single encoder and multiple decoders, which allows for interrelated processing and faster joint training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If sequential processing is used for classifying speech acts, identifying intent, and extracting semantic constituents, then each function can be performed separately, but the processing speed and accuracy are limited
Solution Approach 1:
The patent combines three separate functions (speech act classification, intent identification, and semantic constituent extraction) into a single unified neural network model. The encoder processes the input sequence once and generates a shared hidden state representation that is simultaneously used by three decoders to perform all three functions in parallel, eliminating sequential processing bottlenecks and improving overall processing speed.
Solution Approach 2:
The encoder in the patent serves as a universal component that generates a shared hidden state representation used by multiple decoders for different functions. This multi-functional design allows the same encoded representation to support speech act classification, intent identification, and semantic constituent extraction simultaneously, improving efficiency while maintaining specialized capabilities for each function.
2Measurement precision
If separate training is performed for each function, then each function can be optimized independently, but the training time and computational resources increase significantly
Solution Approach 1:
The patent merges the training processes of three separate functions into a single unified training procedure. The model is trained on a dataset containing annotations for all three functions simultaneously, with a single loss function that combines the objectives of speech act classification, intent identification, and semantic constituent extraction. This allows the shared encoder to learn a comprehensive representation that serves all functions, reducing training time while maintaining optimization quality.
3Reliability
If one function is performed at a time, then the system architecture can be simpler, but the accuracy of each function decreases due to lack of contextual information from other functions
Solution Approach 1:
The shared hidden state representation acts as an intermediary that captures contextual information from the input sequence and makes it available to all three decoders. This intermediate representation serves as a common knowledge base that enriches the processing of each function with contextual cues from the other functions, improving accuracy without requiring direct interaction between the decoders themselves.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A system and method for spoken language understanding using recurrent neural networks ("RNNs") is disclosed. The system and method jointly performs the following three functions when processing the word sequence of a user utterance: (1) classify a user's speech act into a dialogue act category, (2) identify a user's intent, and (3) extract semantic constituents from the word sequence. The system and method includes using a bidirectional RNN to convert a word sequence into a hidden state representation. By providing two different orderings of the word sequence, the bidirectional nature of the RNN improves the accuracy of performing the above-mentioned three functions. The system and method includes performing the three functions jointly. The system and method uses attention, which improves the efficiency and accuracy of the spoken language understanding system by focusing on certain parts of a word sequence. The three functions can be jointly trained, which increases efficiency.