Subword Alignment for Low-Resource Language Text Units

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for annotation projection, particularly IBM-based alignment models, fail to effectively relate text units in low-resource languages due to the scarcity of parallel corpora, which are typically characterized by fewer than 10,000 sentences per language pair.

Innovation Solution

The method involves subword sampling using Byte-Pair-Encoding to generate subword representations, followed by iterative Bayesian optimization to select and align subwords, improving word-level alignment through the use of established methods like fast-align and eflomal, and applying learned hyperparameters across language pairs to enhance alignment results.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If IBM-based alignment models are used for word alignment, then alignment performance is improved in high-resource settings, but alignment reliability deteriorates in low-resource language settings

Engineering Contradiction:
Improvealignment performanceVSAvoidalignment reliability
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent segments words into subwords using Byte-Pair-Encoding (BPE) to create a hierarchical representation. This segmentation allows the model to capture both fine-grained character-level patterns and coarser word-level structures, improving alignment reliability in low-resource settings where complete word-level data is scarce. The subword level provides sufficient statistical signal even with limited parallel corpus data.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimension by aligning at the subword level rather than only at the word level. This multi-level alignment approach (character, subword, word) enriches the alignment representation and allows the model to leverage available data more effectively, transforming the low-resource constraint into a manageable challenge through hierarchical representation.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If subword sampling and iterative Bayesian optimization are applied, then alignment accuracy is improved, but processing time increases

Engineering Contradiction:
Improvealignment accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary subword segmentation using Byte-Pair-Encoding before the alignment process. This pre-processing step creates the subword representations in advance, which then serve as the input for the iterative Bayesian optimization alignment process. The preliminary action of segmentation enables more efficient subsequent alignment operations by working with smaller, more manageable units.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent employs iterative Bayesian optimization with feedback mechanisms to guide the alignment process. The feedback from alignment scoring allows the model to iteratively refine its predictions, focusing computational resources on the most uncertain alignments. This feedback-driven approach improves accuracy while managing processing time by avoiding unnecessary computations on clearly aligned or misaligned pairs.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3916615A1Method, computer program and apparatus for relating text units
Publication Date: 2021.12.01 VOLKSWAGEN AG
  • EP3916615A1 patent drawingFigure 1~2
  • EP3916615A1 patent drawingFigure 3~4
  • EP3916615A1 patent drawingFigure 5

AI summary

The present invention is related to a method, a computer program, and an apparatus for relating text units, in particular for relating text units for low-resource languages. The invention is further related to a system for text processing or language understanding, which makes use of such a method or apparatus, and to a vehicle comprising such a system. In a first step, subword sampling (10) is performed on parallel source and target sentences of a source language and a target language to generate subword representations of the source and target sentences. Then subwords are selected (11) from the space of subword representations. Finally, the selected subwords are aligned (12) to generate related text units.