The current stage of development of the digital world requires solving many problems. In particular, the analysis of texts occupies one of the important places. The scope of such analysis is very wide, from the semantic analysis of resumes to the psychological portrait of a social network user based on his posts. The algorithm of tongue twisters detection in Ukrainian using the topological data analysis methods based on the consideration of a twister as a sample in the four-dimensional space with a construction simplicial structure on it, calculating its invariant and classifiers as machine learning methods to realize distinguishing a tongue twister from a simple narrative sentence is obtained by authors. A better rate of detection was obtained by using a support vector machine with a Gaussian RВF kernel. Ukrainian tongue twister, persistent homology, decision tree, support vector machine COLINS-2023: 7th International Conference on Computational Linguistics and Intelligent Systems, April 20-21, 2023, Kharkiv, Ukraine
A tongue twister is a syntactically short, correct phrase spoken in any language with especially complicated articulation. The tongue twisters contain combinations of sounds that sound similar but have different phonemes and are difficult to pronounce. A tongue twister is naive in content, simple, even primitive text, which is built on intricate and difficult phrases and phrases. Often tongue twisters contain rhymes and alliteration. There are several aspects that contribute to the need to study tongue twisters as a cluster of texts in human speech:
They can be called hard talkers. Working on tongue twisters is very good, very useful diction training. It is a mistake to think that it is necessary to achieve an extremely fast pace. It is not always so. When dealing with hard talkers, as in all work on diction, each speaker requires an individual approach. Speech therapists adhere to another name for this type of speech art - pure speaking. Phrases are used to practice pronunciation and diction. Phrases are ideal for differentiating and training automatic pronunciation. For the greater development of the speech apparatus, it is recommended to pronounce colloquialisms with a change in tempo and volume, i.e. quietly, loudly or in a whisper, slowly or quickly. But the main task when reading tongue twisters is to do it clearly and cleanly, with deep, full articulation, regardless of the volume and speed of speech. Politicians, actors and other public figures are well familiar with this type of literary creativity and the miracle that happens to their articulation, diction and manner of speech after a few hard sessions with a good speech therapist. Studying the connections between colloquialisms and the peculiarities of the language zone of the brain, scientists found out why colloquialisms are so difficult to pronounce. The reason is the close location of the groups of neurons needed to pronounce a set of sounds in colloquial speech.
2023 Copyright for this paper by its authors. Experts have identified two main mechanisms of the so-called “double attack” - a phenomenon when the speaker tries to pronounce two sounds at the same time. The first type of error is when a person tries to pronounce two sounds almost simultaneously, the second - when mixing sounds with a slight delay, and then a vowel sound appears between consonants.
Purpose of the work – to propose an algorithm and its realization for tongue twisters detection in the Ukrainian language that provides the understanding of text shape aspects with the possibility to implement it into machine learning.
The aim of research – to propose topological invariants, the calculation of which will be informative for understanding the nature of a tongue twister, to establish a set of data, the integration of which in a certain method of machine learning has the ability to distinguish a tongue twister from a simple narrative sentence.
Major research objectives are: using methods of topological data analysis (TDA), propose an away to encode the Ukrainian tongue twisters that allows to integrate it in machine learning; based on obtained results, propose an algorithm for detecting the tongue twisters among simple narrative sentences.
Practical tasks in which the algorithm of tongue twisters detection in Ukrainian can be used are the following: rating complexity of existing texts for children who have speech problems. We analyze the sentences of the text and check whether they have the feature of a tongue twister. The more tongue twisters there are, the more difficult for such children a text is. artificial generation of a special type of text in the Ukrainian language. It may be senseless, which is allowed by the nature of a tongue twister, but it is aimed at eliminating one or another problem, for example, improving the speech of announcers or developing it in younger children. artificial generation of tongue twisters will make it possible to enrich the existing dataset (which is currently consists of no more than 500 items for the Ukrainian language) in order to apply NLPT for tongue twisters detection.
In this section, we consider some existing works on analyzing text by its figurative shape, text artificiality detection, classification, visualization and methods of machine learning with help of topological data analysis (TDA). Below, the authors focus specifically on existing solutions using topological data analysis.
Note that TDA is a relatively new direction in data analysis, its main advantage is the ability to learn the space characteristics from the point of view of its geometry (for details, see Section 3.3). However, movement in this direction requires an understanding of the basic concepts and aspects of such a science as topology, which often complicates and slows down the process of obtaining new results in this area.
In today's world, social media produces a huge amount of content [
In [
The issue of distinguishing two texts is relevant, the distance between them can be used as a criterion,
that is, a metric. The authors [
For artificial text detection, the experimental setup also highlights the applicability of the features
towards the TGM architecture, TGM’s size and the decoding method. Notably, the TDA-based
classifiers tend to be more robust towards unseen GPT style TGMs as opposed to the considered
baseline detectors, see [
By Wlodek Zadrozny, see [
Similarly to bag-of-words, there is persistence bag-of-words, see [
Important aspects of text processing are text classification and visualization [
In this section, vectorizing the words, dataset and main terms of persistent homology are considered.
We have to remark that there are several methods for obtaining vector representations for words:
unsupervised learning algorithm (GloVe, see [
The main disadvantage of all is the fact that the larger the sample (dataset), the better the results. Moreover, the amount of training data is expressed in thousands of units. That is why the authors propose the approach described in this section. 3.1.
It is known that there are two classes of letters: consonants and vowels in Ukrainian. Every letter corresponds to a vector ⃗( 1, 2, 3, 4), where: 1 be the ordinal number of the letter in the text; 2 be the ordinal number of the word in the text which contains letter ; 3 be the ordinal number of alphabetical ordered set of vowels. If a letter is consonant, 3 is equal to a zero;
4 be the ordinal number of alphabetical ordered set of consonants. If a letter is vowel, 4 is equal to a zero.
For any tongue twister, there is a map into non negative real four-dimensional space ℝ4+.
Since the mapping is carried out in a four-dimensional space, any visualization is complicated by the human perception so it is necessary to reduce the dimensions.
In Fig. 1 and Fig. 2, there are three projections in different spaces of the same tongue twister “Babyn bib rozcviv u doshch, bude babi bib u borshch”. Common to all of them is the fact that the points are grouped into certain clusters, which in the future can ensure the presence of a certain cycle in passing from point (letter) to point (letter) in the space that is coded the tongue twister. 3.2.
For research, we propose a dataset that consists of 100 Ukrainian tongue twisters, see [
In Fig. 3, there are two histograms of dataset parts used in this research. These histograms show that most tongue twisters have up to 50 letters. They are short and aimed at the quick pronunciation of sounds. in the text and a histogram of the non-twisters dataset [right], where an axis X is the quantity of the letters in the text
This dataset contains 130 short texts, most of which are sentences. The non-twisters were chosen arbitrarily, they are meaningful and are parts of some stories. 3.3.
Let consider the terms of computational topology. The exact mathematical definitions are in [
, : 1i → 1 , i<j, be a map. On other words, 1i,j= 1i /( 1j ∩ 1i ), where 1i is 1-cycles of and 1j is 1-boundaries of (a set { } =1 of Vietoris-Rips complexes is the filtration for any finite set { 1, 2, …, }, where < , i<j.). There is a method of their calculation based on the matrices algebra, the persistence barcode and the persistence diagrams. The 1-cycle is a 1-chain with empty boundary. The group of 1-cycles is the kernel of the 1-th boundary homomorphism, 1=ker 1. The 1boundary is a 1-chain that is the boundary of a 2-chain. The group of 1-boundaries is the image of the 2-nd boundary homomorphism, 1=Im 1+1. A 1-chain is a formal sum of 1-simplices in a simplicial complex K and its standard notation is c=∑ , where is p-simplex in K and is either 1 or 0.
For next calculation, we used the GUDHI library, which is a generic open source C++ library with Python interface, for Topological Data Analysis (TDA) and
Understanding, see [
In Fig. 4, there are two diagrams that are constructed for the tongue twister “Babyn bib rozcviv u
and 2 = 0 (Betti numbers), see [
So, if the maximum length of an edge of constructed Vietoris-Rips complex is equal to 0.6 and the minimum persistence is equal to 0.36 then “Babyn bib rozcviv u doshch, bude babi bib u borshch” can be considered as a circle.
Step 1. Every letter of the text is coded according to Sec. 3.1. If the text consists of symbols (letters) { 1, 2, … , }, then it corresponds to a set = {( 11, 21, 31, 41), … , ( 1 , 2 , 3 , 4 ) }. In other words, the text is considered as a cloud of the points in ℝ4+. We also normalize it by standard function and map a cloud of the points into 4, where 4 = [0; 1]4 be a four-dimensional unit cube. Let denote it by ̃.
Step 2. To construct on ̃ the filtration by Vietoris-Rips complexes, where { 1, 2, …, } is a finite set, < , i<j, and computing 1=rank 1i,j for every fixed , = ̅1̅̅,̅̅. So, then we obtain a set { 11, 21, …, 1}.
Step 3. For every text of dataset, we apply the previous steps and obtain a set which consists of vectors with − coordinates. For set some methods of classification can be used.
In Fig.5, there is a pipeline of an algorithm. A coding corresponds to Step 1, Computation 1 – Step 2 and Classifier – Step 3. We remark the following:
The output of Coding is a cloud of the points into 4, where 4 = [0; 1]4 be a four-dimensional unit cube.
The output of Computation 1 is N numbers of ordered sets of k positive numbers. The output of Classifier is “Yes or No” answer to “Is this text a tongue twister?”
We have to remark that the structure of the filtrations does not depend on the min_persistence parameter.
To construct a classifier we use the parameter test_size = [0.1,0.2, …, 0.9]. In Table 1, the accuracies of considered classifiers are presented with values of considered parameters.
Tongue twisters detection is stable with below average score with the following parameters – Dataset 2 with random sate =1 by Decision tree classifier.
Tongue twisters detection turned out to be the most unstable with the following parameters – Dataset 2 with random sate =0 by Decision tree classifier.
In Fig. 6 and Fig.7, there are two trees for classifying twisters. Both of them are realized on Dataset 2 with test size=10%.
However, some differences are noticeable: For a tree in Fig. 6, there is a parameter max_depth=5. It regularized the model and reduced the risk of overfitting accordingly. This effect is clearly visible in the accuracy value, it is higher than in the solution in Fig.7.
In Fig. 7, the proportion of nodes in which the Gini indicator is close to 0.5 is bigger, which is not a good indicator for Decision tree classifier.
In Fig. 7, a tree classifier contains many nodes with 1 text.
Both trees contain a branch with length less than 20 letters. Moreover, the similarity of parameter values for their longer classification of such short tongue twisters 80 and 40(45) should be noted. A small tree (Fig. 6) has only two threshold values for the number of characters that are 35 and 50. Whereas, a large tree (Fig. 7) generates new nodes with a step of 2-3 characters in the text, which makes it very difficult for understanding.
The authors remark that the application of the classical statistical approach with the formulation of statistical hypotheses on this sample and determination of the level of significance is complicated. Since the sample can not be expanded to increase the level of significance, due to the presence of established expressions in the language that are tongue-in-cheek, it can only be expanded with the help of artificial methods. For example, if you apply to this sample two-sample t-test to estimate the sample sizes of an experimental group and a control group that are of equal size, then statistical power is equal to 0.25. On the other hand, using the same assumptions, the presence of more than 1000 instances in the sample does not guarantee 0.99, since models can have hundreds of thousands of data, but cannot have high training.
That is why the selection of machine learning models and data mining is an art starting from the formation of a dataset and ending with the selection of models and their parameters. There are no theorems and statements that would clearly establish how and on what to teach the module in order to obtain a high level of recognition.
An algorithm of tongue twisters detection in Ukrainian based on the consideration of a text as a sample of points in four-dimensional space, calculation of Betti numbers as a topological invariant of the existence of one-dimensional circles in it and implementation it into machine learning is proposed. This approach makes it possible to see behind the text not only a set of points, but also the possibility of generating a structure in space with a longer calculation of its invariants, the values of which make it possible to see its geometry (a sentence is as "circle", "sphere" or "torus").
In addition, there were realized two methods of machine learning as classifiers for detecting tongue twisters among simple narrative sentences. According to research, more accuracy is provided by Support Vector Machine Classifier with Gaussian RВF kernel.
In further research, there are several ways to improve the accuracy of detection, as well as to improve the process of encoding words or letters taking into account the analysis of the sounds, as well as the complexity in the pronunciation of sounds and calculating of topological invariants that contain cyclomaticity of higher orders.