Word Semantic Embedding Using Lexical Network for Homograph Disambiguation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current natural language processing systems face challenges in accurately disambiguating homographs, especially for unlearned patterns, due to limitations in corpus-based learning methods and lexical networks, which result in low accuracy and mistranslation errors.
Innovation Solution
A word semantic embedding method using a lexical semantic network that generates processing data from word lists and semantic information, employing a modified skip-gram model with negative sampling to learn word vectors, allowing for efficient disambiguation of homographs by comparing similarity between homographs and adjacent words.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If corpus-based learning methods are used for homograph disambiguation, then learning data can be obtained, but accuracy is low especially for unlearned patterns
Solution Approach 1:
The patent introduces a lexical semantic network as an intermediary knowledge base that bridges corpus-based learning and homograph disambiguation. The semantic network provides structured semantic relationships (hypernym, hyponym, synonym, antonym) that act as a mediator to guide disambiguation when corpus patterns are insufficient, thereby improving accuracy for unlearned patterns while maintaining the benefits of corpus-based learning.
Solution Approach 2:
The patent combines multiple data sources (corpus, lexical semantic network, dictionary) to create a composite knowledge representation. This composite approach integrates statistical patterns from corpora with structured semantic knowledge, enabling the system to leverage both learned patterns and semantic reasoning for improved disambiguation accuracy across both learned and unlearned patterns.
2Measurement precision
If word embedding learning uses large amount of corpus data, then word vectors can be learned, but learning time and data requirements increase
Solution Approach 1:
The patent performs preliminary action by pre-constructing a lexical semantic network with structured semantic relationships before word embedding learning. This pre-established semantic framework guides the learning process, allowing the system to learn more efficient word vectors with fewer corpus examples by leveraging the preorganized semantic knowledge structure.
Solution Approach 2:
The patent changes the learning parameters by incorporating semantic network guidance into the word embedding learning process. Instead of relying solely on large-scale corpus statistics, the system uses semantic relationships as additional constraints and features, effectively changing the learning dynamics to achieve better accuracy with reduced data and time requirements.
3Ease of operation
If one-hot encoding is used to express words, then words can be represented, but vector dimension becomes very high
Solution Approach 1:
The patent changes the representation parameter from high-dimensional one-hot encoding to low-dimensional word vectors through embedding learning. This parameter transformation maintains the simplicity of word representation while dramatically reducing the vector dimension, enabling efficient processing in deep learning models.
Data Source
AI summary
A word semantic embedding apparatus includes: a data storing unit to store a lexical semantic network including word dictionary data and word semantic data; a word list generating unit to extract vocabularies having a predetermined part of speech from the stored word dictionary data and generate a list of words to be learned; a processing data generating unit to bring the generated list of words to be learned and the word semantic data of a word to be learned included in the list of words to be learned from the data storing unit and process the data suitable for word embedding learning to generate processing data; and a word embedding learning unit to learn the word to be learned through the word embedding learning using a learning model formed of an input/output layer and a projection layer with the generated processing data to generate a word vector.