Automated Ground-Truth Corpus Generation for Multilingual NLU Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The development of multilingual natural-language understanding (NLU) models is hindered by the resource-intensive process of creating annotated ground-truth corpora across multiple languages, requiring expert knowledge and extensive manual labeling of slot values and intent labels.
Innovation Solution
An automated method for generating annotated ground-truth corpora in multiple languages, which involves delexicalization, translation, and relexicalization of utterances, allowing for the creation of a multilingual NLU model from a single annotated corpus without the need for expert labeling in each language.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation of slot values and intent labels is performed for each language, then accurate ground-truth corpus is achieved, but resource consumption and time required increase significantly
Solution Approach 1:
The patent creates annotated corpora in multiple languages by copying and translating a single annotated corpus from a source language. The system translates utterances and their annotations automatically, avoiding the need to create annotations from scratch for each language. This copying approach maintains annotation accuracy while significantly reducing the time and resources required for corpus creation across multiple languages.
Solution Approach 2:
The patent performs preliminary annotation of a corpus in a source language (e.g., English), then uses this pre-annotated corpus as the foundation for creating annotated corpora in other languages through automated translation. By completing the annotation task once in the source language and then replicating it through translation, the system avoids repeating the time-consuming annotation process for each target language.
2Measurement precision
If expert knowledge is required for labeling slot values and intent labels in each language, then labeling accuracy is maintained, but device complexity and resource requirements increase
Solution Approach 1:
The system copies the annotation structure and labeling decisions from a source language corpus and applies them to target language corpora through automated translation. This eliminates the need for experts to manually label each language while preserving the accuracy of the original annotations, as the translation process maintains the semantic relationships and slot-value correspondences.
Solution Approach 2:
The system enables itself to perform the annotation task by using automated translation to replicate the annotation process across languages. Instead of requiring external expert knowledge for each language, the system uses its own pre-trained capabilities to translate and annotate, making the annotation process self-sufficient and scalable.
3Adaptability or versatility
If annotated ground-truth corpora are created for multiple languages, then multilingual NLU model capability is improved, but manufacturing cost and resource intensity increase
Solution Approach 1:
The patent creates a universal approach where a single annotated corpus in one language serves as the foundation for generating annotated corpora in multiple other languages through automated translation. This multi-functional system allows one corpus to be reused across multiple languages, making the manufacturing process more efficient and scalable compared to creating separate annotated corpora for each language.
Solution Approach 2:
By copying the annotated corpus structure and using automated translation to replicate it in other languages, the system achieves multilingual capability without proportionally increasing the effort required for corpus creation. The copying mechanism preserves the quality and structure of the original annotations while enabling expansion to multiple languages.
Data Source
AI summary
In one embodiment, a method includes accessing an initial corpus of ground-truth utterances in a first language, each ground truth utterance annotated with one or more intent labels and one or more slot labels, each slot label having an associated slot value. The method includes generating a translated, annotated corpus in a second language by delexicalizing each utterance and translating portions of the delexicalized utterance to the second language. The method includes creating a multilingual natural language model by training a natural-language model using the translated corpus.