This article deals with the scientific and practical task of automatically detecting significant keywords and rubricating Ukrainian content in Internet systems based on the method of linguistic analysis of text information. The article presents theoretical and experimental substantiation of the method of linguistic analysis of Ukrainian content using Porter's stemming. The method is aiming to automatically detect significant keywords of Ukrainian content on the basis of the proposed formalization of components of analysis - grammatical (grapheme), morphological, syntactic, semantic, referential, and structural.
In practical terms, the analysis of the symbolic level of the organization of naturalistic text is limited to the separation of syntactic punctuation from the word itself, the allocation of abbreviations, abbreviations, etc. Analysis of existing texts shows that at the level of sign organization of the text one is using the descriptive capabilities of the semiotic system to encode knowledge about fragments of real reality. For instance, the use of quotation marks (for example, the movie theater “Star”) indicates that the token in quotation marks cannot be considered in the meaning given in the dictionary. The proper names given in the text may coincide with the spelling of common words, but have different meanings (for example, the group Black September, Black Friday, student Sophia Vovk (wolf), assistant Andriy Krolik (rabbit), teacher Nadiya Kohut (cock), singer Katya Chile, singer Vinnytska Alona, Actor David Duchovny as Fox William Mulder, Liberty Avenue, May 1 Street, Actress Sarah Gabriel, Actress Nastya Zadorozhna, etc.). In addition, a number of tokens in the text are not subject to the grammatical rules of the language, but act as semantic units of sign level (for example, the number 30, the percentage value of 15%, the reduction of millions, thousand or kg, etc.). These features of naturalistic text make it necessary to develop a significant level of text organization as the initial stage of building a model for understanding text. A linguistic method of processing textual information to automatically detect meaningful keywords consists of six steps. 1. Grammatical (graphemic) analysis of textual content that is, parsing text with regard to the features of graphs of different languages. 2. Morphological analysis of textual content. 3. Syntax analysis of textual content. 4. Semantic analysis of textual content. 5. Reference analysis for the formation of interphase unities. 6. Structural analysis of textual content.
The input of the graphemic (two-graphemic) analysis or parsing step is the current text file and the a priori reference models (lines and characters). Separation of such units of text as a name, designation, title, etc. allows us to identify at this stage some functional elements of the structure of concepts. Therefore, it is advisable to begin with an analysis of character-level text to solve the urgent problem of forming effective domain-specific knowledge recognition procedures. The electronic component of this is electronic dictionaries of abbreviations, geographical names, names. This approach is caused by the diversity of sign (grapheme) representation of lexical units in the text, which defines their different semantic functions in one context or another. For the automated processing of naturalistic information, it is also essential to define the structure of the text - to separate service information, highlight paragraphs, headings, and more. The text is considered as a sort of organized sequence of lines and graphemes. 2
Relate the Highlighted Issue to Important Scientific and Practical Work The article deals with the scientific and practical task of automatically detecting significant keywords and rubricating Ukrainian-language content in Internet systems based on the method of linguistic analysis of textual information. The work was performed within the framework of joint scientific researches of the Department of Information Systems and Networks of the Lviv Polytechnic National University on the topic «Research, development and implementation of intelligent distributed information technologies and systems based on database resources, data warehouses, data spaces and knowledge in order to accelerate the formation processes of modern information society», as well as the department of automation and informationmeasuring technique of Vinnytsia National Technical University within spine Research Center of Applied and Computational Linguistics. The results of the research were carried out within the framework of the state budget research works on the topics "Development of methods, algorithms and software for modeling, designing and optimization of intellectual information systems based on Web technologies «WEB» and "Intelligent information technology of image analysis of text and synthesis of integrated knowledge base language content". Scientific research was also carried out within the framework of the initiative topics of the ISM Department of Lviv Polytechnic National University on the development of intelligent distributed systems based on an ontological approach to integrate information resources. 3
Analysis of Recent Research and Publications
Text content (article, commentary, book, etc.) contains a considerable amount of data
in natural language, some of which is abstract [
Algorithm 1. Linguistic analysis of textual content.
Stage 1. Grammatical (graphemic) analysis of textual content С1 .
Step 1. Divide textual commercial content С1 С2 into sentences and paragraphs.
Step 3. Allocate numbers, numbers, dates, unchanged turns, and content cuts С2 .
Step 5. Formation and analysis of linear content enhancement technology for content С2 .
Stage 2. Morphological analysis of textual content С2 .
Step 1. Obtained the basic (word form with cut offs).
Step 2. A grammatical category is formed for different words (collection of grammatical meanings: rarity, deviation, deviation).
Step 3. Formation of linear ability of morphological structure.
Stage 3. Syntax analysis α4 : С2 ,U K ,T С3 of textual content С2 .
Step 1. The word matches the semantic classes in the dictionary.
Step 2. Selection of morphosemantic alternatives required for this review. Step 3. Cut the words into a single structure.
Step 4. Generate an orderly number of superposition entries with basic lexical functions and semantic classes. The accuracy of results is the most commonly used / corrective dictionary.
Stage 5. Reference analysis for interphase unities.
Step 1. Contextual analysis of text commercial content С3 . With it, the resolution of local references (the one that is, his) is realized and the expression of the expression is the kernel of unity.
Step 2. Thematic analysis. Separation of statements on a theme and a rheum allocates thematic structures which are used, for example, at formation of a digest.
Step 3. Determine the regular repetition, synonymization and re-nomination of keywords; the identity of the reference, that is, the ratio of words to the subject of the image; presence of implication based on situational connections.
Stage 6. Structural analysis of textual content С3 . The prerequisites for use are a high degree of coincidence of terms of unity, a discursive unit, a sentence in a semantic language, utterance, and an elementary discursive unit.
Step 1. Identify the basic set of rhetorical connections between content unities.
Step 2. Building a nonlinear unity network. The openness of a link set involves its extension and adaptation to analyze the structure of the text С3 .
Let us consider in detail each of the stages of the proposed algorithm.
Step 1. Grammatical (graphemic) analysis of textual content. The grapheme is called the minimum content unit of written text. The objective of this level of recognition is to build a formalized representation of the grapheme structure of the text and to develop a formal apparatus for separating and classifying text units on multiple lines and graphs. Generalized recognition algorithm works with certain restrictions on the input text: formatted width; does not contain hyphenation; Does not contain objects as a table, figure, formula or graphic symbol; submitted in known languages, such as English, Ukrainian, and German, rather than Ancient Egyptian, Mongolian, or Elven. The ultimate goal of recognizing the graphemic level of text representation is to build a grapheme structure of text, which includes separating on a plurality of lines and graphemes of input such semantically independent units of text as fragments (discourses), sentences, syntagms, tokens, and defining the types (classes) of enumerated units of text and units the relationship between them in a specific input text. The process of recognition at the grapheme level of the text representation involves two stages, as shown in Fig. 2.
Input data Input text information (line,
character)
Row classifier Character classifier Reference models
Dictionary of names Dictionary of geographical
names Glossary of abbreviations
Software modules The procedure of grapheme
analysis The procedure of pragmatic
analysis Procedure for data
formation for morphological analysis
Output data Marked text Fragments Tokens: - language, - non-linguistic, - conventionally linguistic Grapheme structure
Fragments Sentence Syntagms Tokens Relation Fig. 2. Structural-logistical scheme of recognition of knowledge from the subject area at the graphemic stage of textual information analysis The purpose of the first stage is to separate substantively separate fragments in the text, tokens in each fragment of the text and determine the language of the input text and / or fragments of the text. The input of the first stage is the current text file and a priori reference models of rows and graphs. The string classifier includes the following significant classes: empty string (EmpStr), full string (full string, FulStr), incomplete right (IncRgt), incomplete left (IncLgt), symmetric incomplete (SmtInc). The rules for recognizing lines in the text are given in Table. 1. Many reference models of graphems are conveniently presented in the normal Backus-Naur (BPF) form, the abbreviations of which are given in the table. 2. Consider grammar G = < V, T, S, P >, where the alphabet is V = < Gr, T >; terminal symbols are T : = < A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, Q, P, R, S, T, U, W, V, X, Y, Z, Ä, Ö, Ü, Ą, Ć, Ę, Ł, Ń, Ó, Ś, Ź, Ż, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t, u, w, v, x, y, z, ä, ö, ü, ß, ą, ć, ę, ł, ń, ó, ś, ź, ż, А, Б, В, Г, Д, Е, Ж, З, И, Й, К, Л, М, Н, О, П, Р, С, Т, У, Ф, Х, Ц, Ч, Ш, Щ, Ь, Ю, Я, Є, І, Ї, Ґ, Ы, Э, Ъ, а, б, в, г, д, е, ж, з, и, й, к, л, м, о, н, п, р, с, т, у, ф, х, ц, ч, ш, щ, ь, ю, я, є, і, ї, ґ, ы, э, ъ, ‘> and a list of tuples for those defined in Table 2 elements of the set of reference models of graphems: 1. Gr : = < Sb Sp > is recognized text content as a set of characters and spaces; 2. Sp : = <_> is space as terminal symbol; 3. Sb : = < Ltr Dgt Ssb Ssg > is plural letters, numbers, special characters, and syntax characters; 4. Ltr : = <LatCyrEngGerPolUkrRusCplSmlCnlVwl‘> is a set of Latin, Cyrillic, English, German, Polish, Ukrainian, Russian letters, including uppercase and lowercase letters of the respective languages, consonants and vowels, as well as apostrophe as a terminal symbol; 5. Dgt : = < 0 1 2 3 4 5 6 7 8 9 > is a set of numbers; 6. Ssb : = < Osb Bsb Msb > is service characters set, brackets and math symbols; 7. Ssg : = <«»“”,.:;-?!> is terminal symbols of syntactic characters; 8. Cpl : = < Lcp Ccp Ecp Gcp Pcp Ucp Rcp > is a set of capital letters of the respective languages; 9. Sml : = < Lsm Csm Esm Gsm Psm Usm Rsm > is a set of lowercase letters of the respective languages; 10. Lat : = < Lcp Lsm > is a set of Latin letters both capital and small; 11. Cyr : = < Ccp Csm > is a set of Cyrillic letters both capital and small; 12. Eng : = < Ecp Esm > is a set of English letters both capital and small; 13. Ger : = < Gcp Gsm > is a set of German letters both capital and small; 14. Pol : = < Pcp Psm > is a set of Polish letters both capital and small; 15. Ukr : = < Ucp Usm > is a set of Ukrainian letters both capital and small; 16. Rus : = < Rcp Rsm > is a set of Russian letters both capital and small; 17. Osb : = <№%/@#$&*\> is a set of terminal service characters; 18. Bsb : =< [ ] { } ( ) > is a set of terminal characters of brackets; 19. Msb : = < + < > = > is a set of terminal mathematical symbols; 20. Cnl : = < Ecc Esc Gcc Gsc Pcc Psc Ucc Usc Rcc Rsc > is a set of uppercase and lowercase letters of the respective languages; The following production rules are offered to recognize the language of the text: P := < SS Gr, S, GrGr Sb, GrGr Sp, Gr, Sp_, SbLtr, Sb Dgt, SbSsb, SbSsg, LtrLat, LtrCyr, LtrEng, LtrGer, LtrPol, LtrUkr, LtrRus, LtrCpl, LtrSml, LtrCnl, LtrVwl, Ltr‘, SsbOsb, SsbBsb, SsbMsb, CplLcp, CplCcp, CplEcp, CplGcp, CplPcp, CplUcp, CplRcp, SmlLsm, SmlCsm, SmlEsm, Sml Gsm, SmlPsm, SmlUsm, SmlRsm, LatLcp, LatLsm, CyrCcp, CyrCsm, Eng Ecp, EngEsm, GerGcp, GerGsm, PolPcp, PolPsm, UkrUcp, UkrUsm, Rus Rcp, RusRsm, LcpLcc, LcpLcv, LcpQ, LcpV, LcpX, LsmLsc, LsmLsv, Lsm q, Lsmv, Lsmx, CcpCcc, CcpCsv, CcpЬ, CcpЙ, CsmCsc, CsmCsv, Csm ь, Csmй, EcpLcc, EcpLcv, EcpQ, EcpV, EcpX, EsmLsc, EsmLsv, Esm q, Esmv, Esmx, GcpLcc, GcpLcv, GcpÄ, GcpÖ, GcpÜ, GcpQ, GcpV, GcpX, GsmLsc, GsmLsv, Gsmä, Gsmö, Gsmü, Gsmß, Gsmq, Gsmv, Gsm x, PcpLcc, PcpLcv, PcpĄ, PcpĆ, PcpĘ, PcpŁ, PcpŃ, PcpÓ, PcpŚ, PcpŹ, PcpŻ, PsmLsc, PsmLsv, Psmą, Psmć, Psmę, Psmł, Psmń, Psmó, Psmś, Psmź, Psmż, UcpCcc, UcpCcv, UcpЄ, UcpІ, UcpЇ, UcpҐ, Usm Csc, Usm Csv, Usmє, Usmі, Usmї, Usmґ, RcpCcc, RcpCcv, RcpЫ, RcpЭ, RcpЪ, RsmCsc, RsmCsv, Rsmы, Rsmэ, Rsmъ, LccB, LccC, LccD, Lcc F, LccG, LccH, LccJ, LccK, LccL, LccM, LccN, LccP, LccR, LccS, Lcc T, LccW, LccZ, LcvA, LcvE, LcvI, LcvO, LcvU, LcvY, Lscb, Lscc, Lscd, Lscf, Lscg, Lsch, Lscj, Lsck, Lscl, Lscm, Lscn, Lscp, Lscq, Lsc r, Lscs, Lsct, Lscw, Lscx, Lscz, Lsva, Lsve, Lsvi, Lsvo, Lsvu, Lsvv, Lsvy, CccБ, CccВ, CccГ, CccД, CccЖ, CccЗ, CccК, CccЛ, CccМ, CccН, CccП, CccР, CccС, CccТ, CccФ, CccХ, CccЦ, CccЧ, CccШ, CccЩ, CsvА, CsvЕ, CsvИ, CsvО, CsvУ, CsvЮ, CsvЯ, Cscб, Cscв, Cscг, Cscд, Cscж, Cscз, Cscк, Cscл, Cscм, Cscн, Cscп, Cscр, Cscс, Cscт, Cscф, Cscх, Cscц, Cscч, Cscш, Cscщ, Csvа, Csvе, Csvи, Csv о, Csvу, Csvю, Csvя >.
These production rules are used to identify meaningful units of analysis U C ,U G text commercial content X (phrase, sentence, theme, idea, author, character, social situation, part of the text, clustered in the content of the category of analysis) (STEP 1 parsing based on the language of the text fragments) by a modified Potter algorithm (STAGE 2 stemming). We have the following requirements for choosing a linguistic unit of analysis: great for interpreting value; small in order not to interpret many meanings; easily identified; the number of units is large to isolate the sample.
Stage 2. Morphological analysis of textual content is to find the basics of words,
for example [
Porter’s Stemmer is a Stemming algorithm published by Martin Porter in 1980.
The original version of Stemmer was designed for English and was written in BCPL.
Subsequently, Martin created the Snowball project and, using the basic idea of the
algorithm, wrote Stemmer for common Indo-European languages, including Russian
[
All of these techniques are applied by groups that form the rules of stemming. But
this significantly complicates the algorithm for finding keychains. Therefore, it is first
suggested to consider common endings - not traditional endings, as part of a word, but
the sequence of letters that end a word (Table 3-4). In the Table 3-4 endings of words
1 to 4 letters long are given. Five or more letters are not taken into account, since
there are not enough such words (for the maximum of 5 йтесь (6837), for 6 - ванням
(4656), etc.). This has created a kind of map for the project of stemming. The purpose
of the project is to build a static termination tree and to capture the algorithm of all
branches of the tree. Generally, a more detailed tree can be built [
Vowels letters are а, е, і, ї, о, у, и, е, ю, я. RV is part of the word after the first vowel. It is empty if there are no vowels in the word. R1 is part of the word after the first combination is vowel-consonant. R2 is part of R1 after the first combination is vowel-consonant.
For example, in the word інформаційний: RV = нформаційний, R1 = формаційний, R2 = маційний. Now let's define several classes of word endings, leaving their original names in the original description of the algorithm.
Class 1. PERFECTIVE GERUND Group 1: в, вши, вшися. The ending should be preceded by the letter а or я. Group 2: ив, ивши, ившися.
а, е, і, и, ими, іми, ій, ий, їм, ім, им, ього, ого, ьому, ому, їх, их, ую, юю, ая, яя, ою, єю. Class 3. PARTICIPLE Group 1: вш, юва, ува, уч, юч, л. The ending should be preceded by letter а or я. Group 2: нн, н, ячи, ачи, ова, ову, єм.
Class 5. VERB Group 1: ла, є, єте, йте, ли, люю, й, в, єм, ємо, ний, ло, ть, но, ють, ні, ть, єш. The ending should be preceded by the letter а or я. Group 2: ила, ела, ена, йте, ите, єте, юй, уй, їй, ай, ало, ив, или, имо, ений, ило, їло, ено, ють, ать, ені, ять, іть, ить, иш, ую, ю.
Class 6. NOUN are а, ев, ов, і, тя, е, ами, іями, ями, єї, єю, ями, ям, ії, и, ою, ій, ой, ий, й, им, им, ім, ам, ом, о, у, ах, ях, ую, ю, ія, я.
Class 7. SUPERLATIVE (найдовший, миліший, більший) are ш, іш. Class 8. DERIVATIONAL (милість, щедрість, малість, крайність) is ість. Class 9. ADJECTIVAL defined as ADJECTIVE or PARTICIPLE + ADJECTIVE. For example: падюча = пада + юч + а.
Rules. When looking for an ending, the longest one is chosen. For example, in the word інформація ія needs to be chosen, not я. All inspections are conducted on a part RV. So, when checking for PERFECTIVE GERUND the previous letters а and я should also be inside the RV. Letters before RV don’t take part in inspections at all.
Step 1. To find and ending of PERFECTIVE GERUND. If it exists, then delete it and complete step. In other words, delete the ending of REFLEXIVE (if exists). Then in the following order check and if there is an ending delete: ADJECTIVAL, VERB, NOUN. Once one of them is found, then the step is completed.
Step 2. If the word ends with і – delete і.
Step 3. If in the Step 2 there will be an end DERIVATIONAL, then delete it.
Step 4. One of three options is possible:
Stage 3. Syntax analysis of textual content. Syntax is known to be a set of rules that
make it possible to construct formulas and recognize the correct formulas in the
sequence of characters. It is important for the symbolic computation system that all but
one of the expression logic operations are binary. This will be based on a parser. We
will consider the process of revising the input sequence of characters in order to parse
the grammatical structure according to the given formal grammar. A parser is a
program or part of a program that performs parsing [
Step 4. Semantic analysis of textual content. Latent-semantic analysis is a method of processing information in natural language that allows you to analyze the relationship between a collection of documents (messages, articles, i.e. textual content) and terms (keywords) that occur in them. Compares some factors (topics) to all documents and terms. Initially, words correlate with semantic vocabulary classes. Then, the morphosemantic alternatives needed for this sentence are selected. What follows is the linking of words into a single structure and the formation of an ordered set of superposition entries of basic lexical functions and semantic classes. The accuracy of the result is determined by the completeness / correctness of the dictionary.
Step 5. Reference analysis for the formation of interphase unities. A contextual analysis of textual content is carried out. With it, the resolution of local references (the one that is, his) is realized and the expression of the expression is the kernel of unity. The following is a thematic analysis. Separation of statements on a theme and a rheum allocates thematic structures which are used, for example, at formation of a digest. Define regular repeatability, synonymization and re-nomination of keywords; the identity of the reference, that is, the relation of the words to the object of reflection; presence of implication based on situational connections.
Step 6. Structural analysis of textual content. The prerequisites for use are a
high degree of coincidence of terms of unity, a discursive unit, a sentence in a
semantic language, utterance, and an elementary discursive unit. A basic set of rhetorical
relationships between content unities is identified and a nonlinear unity network is
constructed [
Similarly, a frequency vocabulary is used to define many keywords.
Experimental studies. 100 scientific publications of the Bulletin of the Lviv Polytechnic National University of the series "Information Systems and Networks" (http://science.lp.edu.ua/sisn) from 783 and 805 (http://science.lp.edu.ua/SISN/SISN2014, http://science.lp.edu.ua/sisn/vol-cur-805-2014-2) were selected as the linguistic base for the experimental study of the proposed method. The analysis of statistics of functioning of the system of detection of a set of keywords from 100 scientific articles was carried out in two stages, in particular: 1. Analyze all articles by checking common block words and thematic vocabulary. 2. Analyze all articles by checking refined blocked words and refined thematic vocabulary (with more startup, a set of unknown words (missing both in the thematic dictionary and in many blocked ones) is formed.
In addition, at each stage, the review was performed in two steps for each article: analysis of the entire article (http://victana.lviv.ua/index.php/kliuchovi-slova) and analysis of the article without beginning (title, authors, editors, annotations in two languages, authors keywords in two languages, authors' place of work), and no literature list to determine errors in the accuracy of multiple keyword formation (Fig.5). а) In Fig. 5a) the diagram of analysis of statistics of formation by set of sets of all potential keywords in comparison with the set defined by authors of articles is presented. The first column is the average number of keywords identified by the author (4.77), and the second is the average number of words that make up those author keywords (9.82). The third column is the arithmetic mean of the potential keywords defined systematically in Step 1, Step 1 (5.46); the fourth is in step 1, step 2 (6.51); the fifth in step 1, step 1 (7.43); sixth - in step 2, step 2 (8.35). Label these columns accordingly A1 A6 . The value A3 is different from the value A1 by 0.69 (in quantity but not in content); accordingly A4 is different from A1 by 1.74; A5 from A1 by 2.66; A6 from A1 by 3.58. The value A2 is different from the value A3 by 4.36; accordingly A2 from A4 by 3.31; A2 from A5 by 2.39; A2 from A6 by 1.47. Therefore, on average, the author of an article defines fewer keywords than is actually present in this work. Adjusting the system parameters increases the number of defined keywords by almost 2 times (with a similar comparison the value A1 with he value A3 more at 1.144654; A4 – at 1.36478; A5 – at 1.557652; A6 – at 1.750524). The total increment of the value obtained by the system depending on the moderation of the dictionaries is accordingly A3 – 14.46541; A4 – 36.47799; A5 – 55.7652; A6 – 75.05241. If consistently compare A2 with A3 A6 (in how many times the value A2 is greater), then we get accordingly the range 1.7985; 1.5084; 1.3217; 1.176.
In Figure 5, b) shows a chart that contains statistics on more detailed textures in the analyzed articles, where 1 is an analysis of the page of the article page (husband's manual and average reference), 2 is paragraphs in the article, 3 is lines of text, 4 is words. 5 is characters, 6 is characters and spaces, 7 is words on the page, 8 - characters on the page, 9 - characters and prints on the page.
In Figure 6 shows the distribution diagram of the set of sets of all potential keywords for each article compared to the set defined by the authors of the articles. 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94 97 100
Keyword accuracy is enhanced during the dictionary moderation process. The difference between the number of keywords identified by the author and the system defined in step 1, step 1 is 44.39919% (the percentage difference). The accuracy is improved in step 1, step 2 is 33.70672%, it is significantly improved in step 2, step 1 is 24.33809%, and in step 2, step 2 is already 14.96945%.
In the Table 5 the results of an analysis of the statistics of the set of sets of all potential keywords for each article compared to the set defined by the authors of the articles are presented, where A is for the author's keywords, B is for the systemdefined keywords in stage 1 (step 1), C is for the key words system-defined keywords in stage 1 (step 2), D is for system-defined keywords in stage 2 (step 1), E is for system-defined keywords in stage 2 (step 2). In the Table 6-7 the statistics of the analysis of the text of the articles in the formation of sets of keywords for the construction of appropriate histograms for groups A-E are presented. 20 15 10 5 0
Author's keywords
System Defined - Stage 1
System Defined - Stage 2 The author of a scientific article usually chooses at his discretion the number of keywords in the range of 2 to 8 words (most often 3-5 keywords). The system also defines a different number of words, depending on the writing style of the particular author (there are articles in which the system does not find any keywords by Zipf law). For group B, most often the system determined the number of keywords 5, 7 and 3 (more than 10), although the distribution of found keywords ranged from 1 to 18 words (except 17). For Group B, the system most commonly identified keywords as 5, 7, and 3, although the distribution of keywords found ranged from 1 to 18 words (except 17), but the number of keywords found increased and the highest reliability was achieved. For Group C, the system most often determined the number of keywords 7, 6, 5, 10 and 8, although the distribution of found keywords ranged from 2 to 14 words (significantly narrowed the range). For Group D, the number of keywords 8, 5, 7, and 10 was most often determined by the system, although the distribution of the keywords found ranged from 3 to 16 words (improved accuracy). The article presents theoretical and experimental substantiation of the method of linguistic analysis of Ukrainian-language commercial content using Porter's stemming. The method is aimed at automatic detection of significant keywords of Ukrainianlanguage content on the basis of the proposed formalization of components of analysis is grammatical (graphemes), morphological, syntactic, semantic, referential and structural. To implement grammatical analysis, the rules of line recognition in the text are proposed, the set of standard graph models for 5 languages according to the normal Backus-Naur form is determined, and the corresponding grammar G = <V, T, S, P> for identifying meaningful units of textual commercial content analysis. The morphological analysis was implemented by adapting Stemming M. Porter's algorithm to the Ukrainian language, in particular, a static end-tree was constructed and a weighted level of detail was selected - from 500 words with common endings, rules for truncation of endings and suffixes were substantiated. The basic requirements and procedures of syntactic, semantic, referential and structural analysis of Ukrainian-language commercial content are defined. An experimental study of the method of linguistic analysis was conducted on the materials of 100 scientific publications from two issues (783 and 805) of the Bulletin of the Lviv Polytechnic National University of the series "Information Systems and Networks" (http://science.lp.edu.ua/sisn). Based on the proposed method, the keyword search system demonstrated the ability to improve itself by forming and refining a number of common blocked words and a thematic dictionary with the participation of moderators. It is found that, for the technical scientific texts of the experimental base, the authors of articles usually define fewer keywords on average than are actually present in this work. Numerous statistics show that debugging your system's keywords nearly doubles the number of defined keywords, without compromising accuracy or reliability. Further experimental research will require testing the proposed method to identify keywords from other categories of texts is scientific humanities, artistic, nonfiction, etc.