Partitioned Embedding Neural Network for Morphology-Aware Language Modeling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional neural network language models (NNLMs) fail to effectively account for morphology and word shape, which are crucial for part of speech analysis and information extraction, leading to limitations in capturing semantic and syntactic relationships in natural language processing.

Innovation Solution

The proposed solution involves a Partitioned Embedding Neural Network (PENN) framework, which uses a partitioned hidden layer to model the context of words relative to a focus term, optimizing the objective function for continuous bag of words (CLOW) or skip-gram training styles, and incorporates dense interpolation embedding (DIEM) to generate syntactic vectors that capture word structure and relationships.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional neural network language models are used, then the system can process natural language, but it fails to effectively account for morphology and word shape, leading to limitations in capturing semantic and syntactic relationships

Engineering Contradiction:
Improveaccuracy in capturing semantic and syntactic relationshipsVSAvoidability to account for morphology and word shape
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent segments the word representation problem by introducing character-level embeddings that operate independently from word-level embeddings. The character embeddings are organized into buckets based on character positions, allowing the model to separately capture morphological information (word shape) and semantic information, thereby resolving the contradiction between capturing semantics and accounting for morphology.

Inventive Principle:
Principle #1Segmentation

2Reliability

If the model captures detailed word structure information, then it improves part of speech analysis and information extraction, but it increases model complexity and training requirements

Engineering Contradiction:
Improveperformance in part of speech analysisVSAvoidmodel architecture complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent adds a character-level dimension to the traditional word-level embedding space. By organizing character embeddings into positional buckets and concatenating them with word embeddings, the model captures word structure information without fundamentally redesigning the existing architecture, thus improving reliability while controlling complexity.

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

3Adaptability or versatility

If the model uses dense interpolation embedding to generate syntactic vectors, then it achieves better generalization, but it requires additional computational resources and training time

Engineering Contradiction:
Improvegeneralization capabilityVSAvoidtraining time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent performs preliminary action by pre-computing character embeddings and organizing them into positional buckets during the training phase. This preprocessing allows the dense interpolation embedding to efficiently generate syntactic vectors during inference without requiring extensive additional training, thereby improving generalization while minimizing training time overhead.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3259688B1Systems and methods for neural language modeling
Publication Date: 2024.09.25 DIGITAL REASONING SYSTEMS INC
  • EP3259688B1 patent drawingFigure 1~2
  • EP3259688B1 patent drawingFigure 3
  • EP3259688B1 patent drawingFigure 4

AI summary

In some aspects, the present disclosure relates to neural language modeling. In one embodiment, a computer-implemented neural network includes a plurality of neural nodes, where each of the neural nodes has a plurality of input weights corresponding to a vector of real numbers. The neural network also includes an input neural node corresponding to a linguistic unit selected from an ordered list of a plurality of linguistic units, and an embedding layer with a plurality of embedding node partitions. Each embedding node partition includes one or more neural nodes. Each of the embedding node partitions corresponds to a position in the ordered list relative to a focus term, is configured to receive an input from an input node, and is configured to generate an output. The neural network also includes a classifier layer with a plurality of neural nodes, each configured to receive the embedding outputs from the embedding layer, and configured to generate an output corresponding to a probability that a particular linguistic unit is the focus term.