Code-Mixed Utterance Intent Detection With Token-Level Translation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI chatbots struggle to accurately interpret user intent in code-mixed utterances due to limited training data, language switching, and challenges in pronunciation and word overlapping, especially in multilingual contexts, leading to inaccurate language interpretation and failure in understanding user commands.
Innovation Solution
A system and method utilizing an AI-based Language Translator Network (LTN) to segment and translate code-mixed utterances into a target language, followed by an AI-based Intent Detection Network (IDN) to identify user intent, leveraging contextual information and vocabulary probabilities to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional single language AI chatbots are trained with large amounts of natural language intent data in a selected base language, then the chatbot can accurately understand intent in that specific language, but it fails to interpret code-mixed utterances where users switch between languages
Solution Approach 1:
The code-mixed utterance is segmented into individual words or tokens, and language identification is performed on each segment separately. This allows the system to handle multiple languages within a single utterance by identifying the language of each word independently, then routing to appropriate language models for accurate intent interpretation.
Solution Approach 2:
The system creates a universal intent detection framework that can handle both single-language and code-mixed utterances. By implementing language identification at the word level and using a unified processing architecture, the chatbot gains versatility to adapt to different language scenarios without requiring separate systems for each language combination.
2Measurement precision
If ASR-based systems use language-specific backend pronunciation dictionaries and language models to identify spoken language, then pronunciation accuracy improves, but development time and computational resources increase hugely
Solution Approach 1:
Language identification is performed preliminarily on each word segment before proceeding to full ASR processing. By identifying the language of each token in advance, the system can selectively apply language-specific pronunciation dictionaries and models only where needed, rather than maintaining comprehensive language models for all possible language combinations from the outset.
Solution Approach 2:
Instead of applying uniform language-specific processing to the entire utterance, the system applies language identification and subsequent ASR processing locally to each word or token based on its identified language. This allows efficient resource utilization by activating specific language models only for the relevant segments of the utterance.
3Measurement precision
If language identification methods rely on identifying words in specific languages, then single-language utterances can be processed accurately, but code-mixed utterances with multiple languages or novel vocabulary produce inaccurate interpretation
Solution Approach 1:
The utterance is divided into discrete word segments, and language identification is applied to each segment independently. This segmentation approach enables the system to detect and handle multiple languages within a single utterance, identifying the language of each word even when the overall utterance contains code-mixed or novel vocabulary combinations.
Solution Approach 2:
The system introduces an intermediary language identification layer that processes each word segment before passing it to the intent detection system. This intermediary step translates or adapts code-mixed words into a unified representation that the intent detection model can process, bridging the gap between diverse language inputs and the target language processing pipeline.
Data Source
AI summary
A method for identifying a user intent in a code-mixed utterance of a user is disclosed. The method includes receiving the code-mixed utterance from the user and segmenting the received code-mixed utterance into a set of tokens. The method also includes obtaining contextual information and contextual representations associated with the received code-mixed utterance and generating a fixed dimensional representation for each token of the set of tokens. Further, the method includes generating a set of vocabulary probabilities for the fixed dimensional representation and translating the code-mixed utterances to a translation text in the target language based on the generated set of vocabulary probabilities by using the AI-based LTN. Further, the method includes identifying the user intent in the translation text.