Bi-LSTM Intent Vector Encoding for Voice Assistant Context Disambiguation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Natural Language Understanding (NLU) systems in voice assistants fail to effectively utilize inter-utterance context, leading to difficulties in accurately predicting user intent, especially for context-dependent utterances, as the onus of context disambiguation primarily lies with the dialog manager, which becomes challenging to scale for a large number of such cases.
Innovation Solution
A computer-implemented method that involves encoding intent vectors from user utterances using bi-directional long short-term memory (Bi-LSTM) networks and combining them with previous intent vectors using a gated recurrent unit (GRU) or weighted sum to enhance intent recognition, which is then forwarded to a dialog manager for generating responses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the dialog manager handles context disambiguation for all context-dependent utterances, then intent recognition accuracy can be improved, but the system complexity and scalability deteriorate
Solution Approach 1:
The patent segments the context disambiguation function into two parts: (1) the NLU component extracts and encodes context features from previous utterances into intent vectors, and (2) the dialog manager only needs to combine these encoded vectors with current utterance representations. This segmentation reduces the dialog manager's burden while maintaining accurate intent recognition.
Solution Approach 2:
The patent introduces an intermediary mechanism (context encoder and intent vector representation) that bridges the gap between raw conversational context and the dialog manager's decision-making process. This intermediary transforms unstructured context into structured, reusable representations that are easier to process.
2Reliability
If rule-based context disambiguation is implemented in the dialog manager, then specific context-dependent cases can be handled, but the system cannot scale to handle a large number of different query types
Solution Approach 1:
The patent changes the parameter representation from discrete rules to continuous vector embeddings. By representing intents and contexts as vectors in a continuous space, the system can generalize from seen examples to unseen query types through vector similarity and combination, enabling scalability while maintaining reliability.
Solution Approach 2:
The patent substitutes the mechanical rule-based system with a learned vector-based system. Instead of manually encoding rules for each context scenario, the system learns contextual relationships from data through vector representations, automatically adapting to new query types without requiring explicit rule updates.
3Ease of operation
If inter-utterance context is integrated into the NLU component, then the burden on the dialog manager is reduced, but the NLU component complexity increases
Solution Approach 1:
The patent performs preliminary action by having the NLU component pre-process and encode contextual information from previous utterances into intent vectors before passing them to the dialog manager. This preliminary encoding reduces the complexity of context handling in the dialog manager, as the heavy lifting of context understanding is already done.
Data Source
AI summary
A system and method of improving the Natural Language Understanding of a voice assistant. A first utterance is converted to text and parsed by a Bi-LSTM neural network to create a vector representing the utterance. A subsequent utterance is similarly converted into a representative vector and the two vector are combined to predict the true intent of a user's subsequent utterance in context with the initial utterance.