In the field of Natural Language Processing (NLP), semantic analysis remains one of the most important and relevant challenges. Word Sense Disambiguation (WSD) is a key task within semantic analysis, and the automatic identification of homonymous words requires the application of modern algorithms. Machine learning (ML) algorithms and transformer-based models are among such approaches. In this study, the K-Nearest Neighbors (K-NN), Random Forest, and SenseBERT algorithms were applied to disambiguate homonymous words in the Uzbek language. A dataset containing sentences for diferent senses of Uzbek homonymous words was compiled. The collected data set was subjected to initial pre-processing procedures. After cleaning, the models were trained using Random Forest and Sense BERT algorithms. The trained models were tested, achieving an average accuracy of 92 %. To further improve accuracy, it is recommended to expand the size of the dataset and train the models on high-memory computing resources.
Semantic analysis is itself divided into two categories: sentiment analysis and word sense disambiguation. While sentiment analysis evaluates the emotional tone of a text, word sense disambiguation helps to determine the intended meaning of a word within its context. The task of word-sense disambiguation is considered one of the core areas of NLP and plays a crucial role in the field. Solving the problem of identifying the correct meaning of a word in a natural language is achieved by semantically distinguishing between diferent types of words.
In numerous natural languages around the world, this problem has been explored and addressed
efectively. The commonly adopted approaches to word-sense disambiguation are presented below:
• Knowledge-based approaches
• Machine learning-based approaches
– Supervised learning algorithms
– Unsupervised learning algorithms
– Semi-supervised learning algorithms
• Neural network models
Numerous studies have been conducted across a wide range of the world’s languages to address the
problem of word sense disambiguation. In recent years, modern approaches have been applied to tackle
this task. For example, in Russian, neural network models have been utilized to identify homonyms [
Similarly, in the Uzbek language, a number of studies have been conducted to determine the meaning
of words. The problem of word-sense disambiguation can be addressed using knowledge-based
approaches, though their accuracy is relatively limited. Such approaches are typically applied to determine
the contextual meaning of ambiguous words within a specific domain [
Supervised learning algorithms require a dataset consisting of semantically tagged data that belongs to specific categories (1...). Each data point within the dataset is represented by a set of features (1...2). The goal of the algorithm is to learn and identify the relationships between these features and their corresponding categories, enabling it to accurately classify new, unseen data based on the learned patterns.
To determine word meaning in context, supervised learning requires a semantically tagged dataset
or corpus. For instance, when disambiguating polysemous words in English, the Sem-Cor corpus
is commonly used. Sem-Cor is a subset of the Brown Corpus, containing 234,000 words, where the
lexical units in each sentence are annotated with their corresponding WordNet senses. Naturally, these
WordNet sense annotations for the lexical units were manually labeled by human annotators [
When applying these algorithms, two main types of features are typically used, often combined in various ways: collocational features and co-occurrence features.
Collocational features refer to specific words (along with their part-of-speech (POS) tags) that appear at certain positions immediately to the left or right of the target word. "Traktor matorining shovqini suruvdagi otlarni hurkitib yubordi"
If the target word in this sentence is "ot+larni", its feature vector - consisting of the two words to the left and two words to the right — would be as follows. [shovqin, N, suruv, N, hurkitmoq, V, yubormoq, V].
The feature vector is made up of the lemmas of the two preceding and two following words relative to the current ambiguous word, along with their corresponding part-of-speech (POS) tags.
Co-occurrence features rely on the words surrounding the target word. In this approach, the features are the surrounding words themselves, without considering their part-of-speech (POS) categories. The value of each characteristic indicates how frequently these words appear in the context of the target word.
To apply this method efectively, a small number of meaningful words that frequently occur near the target word are typically selected as features. For example, for the word "ot", the most frequently occurring surrounding words from sentences containing its diferent senses might include the following: tog’, dara, chipor, toy, poyga, dala, ... "Zotdor otlar orasida poyga uchun qabul davom etmoqda. . . "
If the window size is set to 10, this sentence can be represented as the following vector. 2.1. Unsupervised learning algorithm.
In this approach, we start with an untagged training dataset — meaning we do not know which class each data point belongs to. Only the features are available, and the algorithm must independently identify which data points belong to the same class based on the patterns it discovers in the data.
Machine learning approaches and transformer-based models ofer significant capabilities for identifying the contextual meaning of homonymous words. In this study, we utilize machine learning algorithms to semantically distinguish homonymous words within sentences in the Uzbek language.
K-NN is a supervised machine learning algorithm that applies a classification method based on semantic
similarity [
Feature extraction refers to the process known in English as Feature Extraction [
In this method, the most frequently occurring words in the text are selected. Based on TF, the words
that appear most often in the context of a particular sense of an ambiguous word are identified, as they
help determine the intended meaning of the ambiguous word [
In the K-NN algorithm, Euclidean distance is commonly used as a distance measure
3. Finding the k Nearest Neighbors: For a new word, the closest neighboring words are selected
4. Learning and Classification: Based on these neighbors, the meaning of the new word is
determined. In other words, the sense of an ambiguous word in a newly provided sentence is
identified by relating it to the overall meaning of its neighboring words [
In word sense disambiguation, the position and frequency of words within a text play an important role. The weight of each word can be calculated based on the following formula:
The significance (weight) of a word is determined according to its frequency in the context and its relative position within the sentence. This weighting approach helps prioritize more contextually relevant words when identifying the meaning of an ambiguous word.
(, ) = (, ) (1) ()
Here: (, ) – the number of feature words that co-occur with the word in sense , () – the total number of samples for sense . Average accuracy: 76.1
The proposed method outperformed previously developed approaches.
This algorithm uses multiple decision trees and yields good results in identifying complex homonyms. Let’s review the process of determining the meanings of homonymous words in a text using the Random Forest algorithm. This process consists of the following main stages: • Homonyms and their context-based sentences are identified through a dataset (Excel file). • The model is trained and then used to determine word senses in new texts • The model’s accuracy is evaluated through testing.
To build a model using the Random Forest algorithm, sentences are first collected for each meaning of the homonymous words. The dataset for this algorithm was compiled in the following format.
Libraries required for model development (Figure 8).
Model Training and Evaluation
Context Tag_sense Kasr sonlarni o‘nli kasrga aylan- Siniq, parcha. Matematika termini; tirish bo‘yicha dars o‘tdik miqdor birligining qismi Bir litr sutning yarm kasri qay- Siniq, parcha. Matematika termini; natildi miqdor birligining qismi.
Bog‘ning uchdan bir kasri sabzavot Siniq, parcha. Matematika termini; ekish uchun ishlatildi. miqdor birligining qismi U kasr tufayli yugurish mu- Shikastalik, nuqsonlilik sobaqasida qatnasha olmadi Kun bulutlarni yorib, qishloqqa is- Yerga issiqlik va nur taratib turusiqlik berdi vchi planeta, quyosh, oftob U kun chiqishini suratga olish Yerga issiqlik va nur taratib turuuchun dron ishlatdi vchi planeta, quyosh, oftob. 1. Sentences and their corresponding senses were extracted from an Excel file. 2. The sentences were converted into vector representations using the CountVectorizer method. 3. The dataset was divided into training and testing subsets in an 80/20 ratio using the train-test-split function.
4. A RandomForestClassifier() was employed to train the model.
During the model development process, a total of 54,000 semantically annotated sentences were collected. Of these, 43,200 sentences were used to train the model. The resulting model was then tested on 10,800 instances. The test dataset contains 20 sentences for each meaning of Uzbek noun homonyms. In the test results, each homonym was analyzed separately. The results of the analysis can be seen in Table 3.
In this article, the 1 scores are calculated for the words presented in Table 2. Based on the results of the testing phase, the 1 score was calculated to evaluate the performance of the model. The test results are summarized below: • Total phrases recognized as valid expressions: 10800 • Correctly identified synonyms: 9950 • Incorrectly matched synonyms: 50 To assess the overall performance of the model, the 1 score was calculated based on these results.
* 1 = 2 * + = 1000 (3)
1000 + 50 = 1000 (4)
1000 + 9950 Based on the calculated precision values (3) and recall (4), (2) was found to be 0.664.
The data was extracted from a corpus containing 460 million words, which includes news articles, scientific abstracts, spoken-language transcripts, and literary texts. In terms of size, this represents one of the largest training and testing datasets used in the field of word-sense disambiguation. The compiled dataset was used to train a modern transformer-based model, specifically the Sense BERT algorithm, and the resulting model was subsequently evaluated. The model trained with Sense BERT demonstrated significantly higher accuracy compared to traditional approaches. In particular, the same data set size was used for both the Random Forest and Sense BERT algorithms. The Sense BERT model achieved an accuracy score of 92 %. (2)
Modern approaches to word-sense disambiguation rely on several empirically observed linguistic properties, particularly the principle that a word typically conveys a single meaning within a given collocational and discourse context. Algorithms aim to exploit these properties by modeling the diversity of collocational relationships as efectively as possible. In this study, the problem of identifying the contextual meanings of homonymous words in the Uzbek language was addressed using Random Forest, K-Nearest Neighbors (K-NN) and Sense BERT algorithms. Given that homonyms carry diferent meanings depending on the context in which they occur, automatically determining their intended sense is one of the essential challenges in natural language processing. The primary goal was to identify the meaning of a homonymous word in a user-provided text based on its context.
During the model development phase: • A dataset in Excel format was compiled, containing columns for the homonym, the sentence, and its corresponding meaning • The texts were cleaned and lemmatized. • Sentences were converted into numerical representations using vectorization techniques such as
TF-IDF. • The Random Forest algorithm was selected for model training because: – It achieves high accuracy in classification tasks. – It is resistant to overfitting.
– It provides interpretable results.
The final model successfully identified the contextual meaning of homonymous words in user-submitted texts.