The objective of this research is to develop and improve document classification methods in low-resource natural language processing through graph embeddings of linguistic dictionaries.
Considering that 63% of the Internet is written in English, and most of natural languages are represented in less than 1% of all web pages, a lot of natural languages are considered low-resource, and are less researched in the field of natural language processing. This leads to information systems built to work with low-resource languages having lower quality than their English counterparts.
Consequently, improving existing low-resource natural language processing methods and the development of new ones is a relevant research problem.
The results of the research showed that vector representations of dictionaries based on graph node embedding methods can be combined with common vector representations of documents to improve the quality of document classification using machine learning approaches. Each step of the proposed method has a set of parameters and hyperparameters, which the result and effectiveness of the final solution depend on. Therefore, an analysis of these options is additionally given, as well as a comparison of different approaches to the construction of graph node embeddings in the context of dictionaries. To achieve the best results, it is suggested to use random-walk based method - Node2Vec, which converts dictionary elements into vectors in an acceptable time, does not require a lot of resources, and receives higher F1-scores further down the pipeline – during document classification. For the next step, namely the fusion of vector representations of documents and dictionary information, the weighted sum method turned out to be better than concatination. In addition, practical recommendations for working with such data are provided, namely, the process of obtaining, saving and preprocessing documents for each of the proposed methods, saving and processing of a synonyms dictionary, as well as the analysis of statistical significance of the results.
Scientific novelty of the results includes:
For the first time, a method of document classification based on dictionary embeddings during low-resource natural language processing is proposed, which differs from dictionary-based methods of data augmentation in that it fuses vector representations of documents with vector representations of elements of linguistic dictionaries, which allows to increase F1-score of document classification in a low-resource environment;
A vector model of words from the dictionary of synonyms is proposed, which, unlike others, is built on the basis of vector representations of the nodes of the dictionary graph, which makes it possible to reuse it in various tasks of natural language processing through transfer learning;
The methods of concatenation and weighted sum during vector representations of words fusion have been modified by adding a stage of matching words from the document to words from the dictionary of synonyms, which allows for covering word forms missing from the dictionary without building models for part of speech tagging and word form generation, which is significantly complicated in low-resource environments.
The practical significance of the results includes:
The proposed method makes it possible to significantly increase the F1-score of document classification systems in low-resource environments. This way, developers of these systems can reduce development time and costs, because higher system quality will be achieved with less labeling, the process which may not be available or require additional time or financial investment;
Vector representations of words in the dictionary of synonyms of the Ukrainian language were developed, which can be reused with the help of transfer learning when creating software systems in other applied areas;
A data set for the classification of petition topics is presented, aimed at testing low-resource natural language processing methods. The documents are written in Ukrainian and have a narrow urban specialization, which makes the data set different from general-purpose corpora;
It is proposed to apply the developed method to the topic classification of petitions to the Kyiv City Council, which allows for automatic suggestions of topic for the petition during manual labeling. This can significantly reduce the time for their analysis.
