1. Introduction

The task of automatic keyphrase extraction (KE) is gaining increasing relevance amidst the exponential growth in both textual and audiovisual information. Effective keyphrase extraction methods enable the structuring and analysis of large volumes of data, automating processes of search, classification, and summarization. This has broad applications across various fields, including computational and cognitive linguistics, psycholinguistics, philology, communication theory, computer science, economics, and more. The selection of extraction methods, in this context, is considered a complex and debated issue. Within the domain of text linguistics, this task is regarded as one of the most challenging (Papusha, 2008). Areas where KE is essential include natural language processing, information retrieval, machine learning (ML) and artificial intelligence (AI), education, databases, marketing and advertising, business analytics, translation, and social media analysis.

The significance of this field is further evidenced by the substantial and growing body of research. A search of the Dimensions[1] engine reveals 24,338 publications on keyphrase processing and applications over

the past 50 years (see Fig. 1). The number of papers devoted to keyphrase extraction, assignment, search, etc. has increased exponentially since the 2000s, reaching a maximum of 1137 items in 2023. Distribution of research papers over fields of knowledge indicates the dominance of three domains: Biomedical and Clinical Sciences (12,485 papers), Information and Computing Sciences (9,836 papers), Language, Communication and Culture (974 papers), cf. Figure 2.

VOS Viewer data visualization[2] allows users to navigate in co-authorship and citation networks. For top 500 researchers we registered 531 co-authorship links in 2739 total co-authorships distributed over 200 clusters, cf. Fig. 3. VOS Viewer indicates high citation density and allows to point cross-reference network with fundamental publications on the topic at its core, cf. Fig. 4. For top 500 researchers the network includes 8994 citation links, 17593 total citations distributed over 52 clusters. However, the problems of keyword annotation and spoken text processing are not sufficiently covered in publications, so links to them are not included in the top results of scientometric tools. Our work aims to restore the interest of the academic community in these aspects.

Figure 1. Publication activity: papers on keyphrase analysis and processing via Dimensions over 50 years

Рисунок 1. Публикационная активность: статьи по анализу и обработке ключевых выражений в Dimensions за последние 50 лет

Figure 2. Distribution of research papers on keyphrase analysis and processing over fields of knowledge (top 10 domains)

Рисунок 2. Распределение научных работ по анализу и обработке ключевых выражений по областям знаний (топ-10 областей)

Figure 3. VOSViewer co-authorship network in papers on keyphrase analysis and processing

Рисунок 3. Сеть соавторов статей по анализу и обработке ключевых выражений в VOSViewer

Figure 4. VOSViewer network of citations in papers on keyphrase analysis and processing

Рисунок 4. Сеть цитирований в статьях по анализу и обработке ключевых выражений в VOSViewer

Different scientific disciplines exhibit significant divergences in their approaches to studying and defining keyphrases. For instance, information retrieval focuses on the practical application of keyphrases for information extraction, while psycholinguistics concentrates on their cognitive and linguistic properties. The issue of approach unification is even proposed to be addressed through musicological analysis, based on common patterns in the semantic structure of poetic texts and musical compositions (Antipina and Prokhorenkova, 2020). The multiplicity of approaches leads to variations in terminology and methodology.

This article examines contemporary theoretical approaches and practical solutions for the task of automatic keyphrase extraction. Within this review, we analyze the stages of automatic KE, the features characterizing them in written and spoken language, and the terminological specifics that have emerged across different scientific domains. We also discuss various methods of automatic KE, including statistical, hybrid, machine learning-based, and structural approaches, and analyze their advantages and limitations. Particular attention is paid to the specifics of KEs in spoken language, including their acoustic characteristics and their impact on perception, as the extraction of these units from speech presents unique challenges and remains significantly underexplored compared to written text analysis. Furthermore, an overview of current trends in automatic KE from spoken language is presented, encompassing the use of deep learning methods and automatic annotation. The analysis of these methods will allow us to evaluate the effectiveness of different approaches and identify promising directions for future research.

This study is framed by the following research questions aimed at advancing our understanding of keyphrase extraction:

RQ1: How do different research disciplines (information retrieval, psycholinguistics, speech processing, etc.) define and operationalize keyphrases in their methodologies?

RQ2: How do these methodologies contribute to keyphrase extraction procedures across disciplines?

RQ3: What opportunities arise from combining psycholinguistic theories with information retrieval metrics for improved keyphrase extraction?

RQ4: What are limitations of current automatic keyphrase extraction methods as regards text genres, domains, written or spoken texts?

RQ5: What are the most significant language-specific challenges in keyphrase processing?

RQ6: How LLMs may improve automatic keyphrase extraction and generation?

RQ7: What are the perspectives of integration of multimodal information into keyphrase processing?

Our review is based on the following methodological principles: in choosing theoretical foundations, we focus on those approaches that have been implemented or are applicable to Russian-language data; in choosing datasets, corpora and models we take into account their accessibility to the scientific community, in choosing technical solutions we consider their scalability and reproducibility.

2. Theoretical foundations of keyphrase extraction

2.1. Terminology and Definitions

The terms “keyphrase,” (“ключевое выражение/словосочетание”) and “keyword,” (“ключевое слово”) can often be used synonymously in most studies, including those mentioned in the next paragraph. In this context, “keyword” frequently refers to both single words and multi-word expressions, driven by the necessity of using collocations rather than isolated words for a more accurate and precise reflection of content (Bolshakova et al., 2011).

Additionally, the Russian-language literature includes terms such as “ключевые термины” (“key terms”) (Grineva and Grinev, 2009), “слова-концепты” (“concept-words”), “слова-лейтмотивы” (“leitmotif-words”) (Svetozarova and Shtern, 1989), “лексические доминанты” (“lexical dominants”), “семантические доминанты” (“semantic dominants”), and “смысловые вехи” (“semantic milestones”) (Shekhtman, 2005). In English-language literature, “keyword,” “keyphrase,” and “key item” are used (Wilson, 2013). The concept of “keyness” is also employed in English literature, referring to a statistical measure that reflects the most content-rich words in a text. Its analysis is used to identify elements that occur statistically significantly more or less frequently in one corpus compared to another, and can focus on differences (keyness-D) or similarities (keyness-S) between corpora (Kilgarriff, 2009; Gabrielatos, 2018; Evert, 2022).

The variety of terminology used to identify keyphrases is illustrated in Figure 5.

Рисунок 5. Терминологическое разнообразие для обозначения ключевых выражений

In this work, we will use the term “keyphrase,” which, in our view, leads to fewer ambiguities. When citing the work of other researchers, their original terminology may be preserved.

We proceed to an examination of the core definitions of keyphrases presented in the Russian-language domain literature. In text perception studies, a psycholinguistic approach is often adopted, where a set of keywords is considered a condensed version of the original, expanded text (“текст-примитив”), possessing wholeness and coherence (Sakharny and Shtern, 1988). Keywords convey the text’s theme, and their ordering refects the rheme (Sakharny, 1982). It is believed that any written or spoken text can be compressed into a set of keywords. Consequently, the content structure, defined by the distribution of keyphrases within a text, can be described as semantically significant, as it enables one to understand the main meaning of the text. This is particularly important for text perception. Keyphrases extracted from a text form its linguistic image, which can be linearly expanded back to reconstruct the content of the compressed text.

The foundational research conducted by Sakharny and Shtern (1988) warrants particular scholarly attention within the context of this review. Their seminal investigation remains highly relevant as a cornerstone for contemporary KE studies, and crucially, it constitutes one of the relatively few methodologies explicitly built upon the empirical analysis of keyphrases functioning within spoken discourse.

The work by Sakharny and Shtern (1988) presents a review of a series of experiments using various texts (fiction, scientific and technical, spontaneous dialogues, etc.), which were used to extract the keywords, reconstruct text from the sets of keywords, and reconstruct spoken text presented against a background of white noise. In the latter case, experiments were conducted both with and without prior familiarization of participants with the sets of keywords. The researchers investigated patterns of text compression into the sets of keywords, their structure, their connection to the original text’s structure, and the influence of the keywords on the noise resistance of the original text. Typically, words from the initial part of the text are prioritized for the set of keywords, with further expansion occurring through refinement of the characteristics of the selected keyword(s). Nouns (in their base form) and/or noun-adjective combinations denoting the themes of text fragments often serve this role. From a syntactic perspective, a chain-like structure is characteristic of the set of keywords (where the preceding rheme becomes the theme in the subsequent fragment). Disruptions arise when the original text has a more complex structure. Experiments also demonstrated that prior familiarization of subjects with the set of keywords can improve the perception of the original text under both external and internal interference (white noise, poor language proficiency, missed words in the text). This confirms that keywords create contextual connections and associations that facilitate text comprehension, and the mechanism underlying improved perception after familiarization with the set of keywords is universal, operating for both Russian speakers in noisy conditions and foreign learners of Russian.

In scientific texts, a set of keyphrases is defined as an invariant of content, reflecting the second stage of scientific text generation and perception (after the title, before the abstract and full text) (Kamshilova, 2013). Keyphrases are found at all these stages but with varying quantitative and substantive content, and they are typically present in strong positions within the text (first sentences and paragraphs).

Since keyphrases indicate “the subject area of knowledge, space, and time, equally understood by members of the same professional community, facilitating dialogue between them” (Moskvitina, 2009, p. 279), they can attract attention to texts, which aligns with the goals of information retrieval. In this context, the set of keyphrases includes the most significant units of a text and serves to form the search profile of documents (indexing), allowing for their concise characterization and retrieval (Jacquemin and Bourigault, 2003; Dostal, 2011). The applicability of keywords as descriptive metadata is universal, extending across diverse information objects, such as text, image, audio file, video file, and digital resources, disseminated online, thereby establishing their relevance for core information retrieval tasks (Abramov, 2011). Their use enhances search efficiency, creates compact indexes, improves ranking, and optimizes user queries. In a general sense, the approach to keyphrase extraction from an information retrieval perspective aims to create content that maximally matches user queries and increases website visibility in search results.

Internet texts are presented as HTML pages containing information about the website’s content. Search engines index these pages, using keyphrase extraction to rank search results. Most information retrieval systems are based on a vector data model, where text is represented as a vector (set) of keyphrases. Search engines compare user query vectors with the vectors of indexed web pages. Web pages containing keyphrases that match the query will have more similar vectors and, consequently, will be considered more relevant (Bolshakova et al., 2011).

Keyphrases are also defined as an “information portrait” of a topic, obtained as a result of searching for a specific query (Bolshakova et al., 2011). Their relevance and pertinence are determined by statistical and semantic algorithms. Evaluation criteria can include features such as the frequency of occurrence, density, and distribution of keyphrases within a document, as well as the use of semantic indexing algorithms (specifically, LSI, Latent Semantic Indexing), i.e., the presence of keyphrases semantically related to the target keyphrases. The final portrait can be presented, for example, as a “semantic map” or a table demonstrating search results (Bolshakova et al., 2011).

From the perspective of semantics, a branch of linguistics, keyphrases are defined as “семантические доминанты” (“semantic dominants”) (Shekhtman, 2005), which provide access to the hidden meaning of a text, i.e., they are considered in conjunction with the situational contexts they represent. With this approach, keyphrase types are identified at three levels of semantic structure organization: at the content level – “опорные слова” (“supporting words,”) at the meaning level – “смысловые вехи” (“semantic milestones,”) and at the intent level – (“слова-концепты”) (“concept-words.”) Based on this, it can be stated that keyphrases convey the meaning of a text at all levels of its comprehension (Moskvitina, 2009).

2.2. Features of Keyphrases

Traditionally, the primary features of keyphrases are considered to include their presence in high-frequency vocabulary; semantic coherence; and part-of-speech belonging to nouns, verbs, and adjectives (Vanyushkin and Grashchenko, 2016). However, the identification of these features is quite conventional. In reality, keyphrases are not always frequent, may not be semantically connected (e.g., in conversational text characterized by rapid topic shifts), and can belong to other parts of speech.

The features characterizing keyphrases can also depend on the functional style of the text. For instance, in literary texts, pronouns are often identified as keyphrases (a prime example being the pronoun “я” (“I”) in texts narrated in the first person) (Svetozarova and Shtern, 1989).

For scientific texts, the following features are important (Moskvitina, 2009).

• High frequency. The occurrence frequency within the studied text is higher than in the language in general.
• Semantic proximity to the text’s theme. Connection to the text’s meaning, e.g., synonymy, hyponymic/hypernymic relationships.
• Informational density. Maximum amount of information with minimal volume; consequently, terms often function as keyphrases in scientific texts.

Synthesizing the various interpretations proposed by researchers (e.g., the previously mentioned Svetozarova and Shtern, 1989; Shekhtman, 2005; Moskvitina, 2009), it can be concluded that generally keyphrases inherently possess a set of common features, which typically encompass the following.

• Significance for text comprehension.
• Semantic integrity (cohesion/wholeness).
• Mutual intelligibility across a broad readership (or professional community).
• Conciseness and the capacity to represent substantial volumes of information compactly.

The aforementioned attributes are relevant to both written and spoken textual modalities.

2.3. Specific Features of Keyphrases in Spoken Text

Keyphrases in spoken text can possess all the features they exhibit in written text, but with the addition of several peculiarities related to the production and perception of spoken language. Consequently, in spoken text, keyphrases are characterized by “prosodic specificity” (“просодическая специфичность”) (Yagunova, 2004). Significant features of keywords in spoken text include phonetic attributes such as their position within a syntagma and the presence of a pause before and after the word.

Research findings (Yagunova, 2004) indicate that keywords are 40% more likely to occur at the end of a syntagma than other text components. The probability of a pause before other words is 70% higher than before keywords. Concurrently, the probability of a pause occurring after a keyword is 1.5 times greater than the probability of it occurring after other words, and almost twice as great for pauses exceeding 750 ms (a duration corresponding to the end of a sentence). An experiment was also conducted with 8 participants whose task was to reconstruct a text after listening with interference (under white noise conditions). Participants recognized keywords 20% better than non-keywords. Furthermore, their recognition accuracy in the final syntagmatic position was 33% higher. Pauses of less than 750 ms, implemented before words, increased keyword recognition accuracy by 23%, and by 55% when such pauses occurred after keywords. Analogous effects for other words were practically absent.

The acoustic (including prosodic) features inherent in keyphrases are not sufficiently studied. A significant number of studies focus on extracting keyphrases from text, but far fewer works are based on spoken texts, and even fewer studies utilize information from audio signals, such as acoustic features. Automatic extraction of keyphrases from spoken language will be discussed in more detail in Section 2.4.

Chen et al. (2012) propose using the following groups of acoustic features:

• Duration-related features. The authors postulate that the duration of words when functioning as keywords may exceed their duration when they are not keywords. Based on this, the average duration of each phonetic unit from a corpus is calculated, and then the duration is normalized by the mean value. Subsequently, for each keyword, its maximum, minimum, average, and range of normalized values are used as duration characteristics.
• Fundamental frequency (F0) features. It is validated that keywords may be distinguished by a wider fundamental frequency range compared to non-keyphrase segments. As with duration, maximum, minimum, average, and range of F0 values are calculated. This measurement is highly relevant because the fundamental frequency is an acoustic correlate of vocal fold vibration and functions as a crucial marker reflecting both the speaker’s affective state and the language-specific contours observed in spoken discourse.
• Spectral energy features. It is also confirmed that keywords are characterized by higher energy levels. In this regard, the authors (Chen et al., 2012) rely on cepstral coefficient data. The maximum, minimum, average, and range values are also calculated as characteristics.

Although the authors generally report positive results confirming their hypotheses, it is important to note that the study was conducted on English and Mandarin Chinese. Consequently, results may differ for other languages. Furthermore, acoustic features were examined not in isolation but in conjunction with lexical and semantic characteristics.

Among the phonetically significant features for the Russian language, Yagunova (2004) reliably confirms only the role of pauses as potential indicators of keyphrases in a text, while the connection between keyphrases and other acoustic features remains unclear.

Guseva et al. (2024) investigate the acoustic characteristics of keyphrases manually identified by participants in spoken popular science texts. An analysis using the OpenSMILE library and the eGeMAPS v0.2 feature set reveals the presence of acoustic correlates associated with keyphrases identification. However, only weak positive or negative correlations are found between individual acoustic features and the keyphrases identified by participants. No significant correlations are found with the keyphrases’ position in the sentence or pauses before/after them. A weak but consistent correlation is noted with spectral and frequency characteristics, suggesting that Russian keyphrases might be characterized by a wider pitch range and greater energy, which aligns with findings from studies on other languages (Chen et al., 2012).

The features that may be significant when identifying keyphrases in a spoken text are illustrated in Figure 6.

Рисунок 6. Характеристики ключевых выражений в устном тексте

Research has described several primary functions of keyphrases, which are listed below in a progression that moves from macro-level text organization and system utility toward individual user memory and comprehension.

• Highlighting essential information. This allows distinguishing a text from others, classifying it, and assigning it to a specific thematic area (Gulyaev and Lukashevich, 2013).
• Information structuring. Information becomes more understandable and easier to perceive (Mitrofanova and Gavrilik, 2022).
• Document clustering. This facilitates grouping documents into relevant categories and simplifies information retrieval (Popova and Danilova, 2014).
• Content evaluation of documents. This information can be applied in algorithms for indexing, categorization, summarization, and document compression (Dubinina, 2020).
• Memory storage. The capacity of human working memory corresponds to the number of keyphrases typically comprising a set. This facilitates memorization and recall of information (Moskvitina, 2018).

Among the functions of keyphrases accompanying scientific articles, the following are particularly highlighted.

• Determining the subject area of a scientific publication.
• Enabling retrieval of a desired article from a database via a search query.
• Gaining an understanding of article similarity without detailed content review (Abramov, 2011).
• Communicative function. Keyphrases help organize subject catalogs and scientific journals, simplifying the search for necessary information and accelerating the process of scientific communication (Moskvitina, 2009).

Thus, keyphrases perform several important functions, which are schematically illustrated in Figure 7.

Рисунок 7. Основные функции ключевых выражений

The selection of keyphrases is based on their frequency within a specific text, not the language as a whole (Bolshakova et al., 2011). For example, in an experiment using linguistic articles, participants without linguistic education who poorly understood the meaning of the presented texts identified linguistic terms that appeared most frequently in the texts as keyphrases (Sakharny and Shtern, 1988). In literary texts, words denoting the subject of description and/or characterizing characters and main plot events are often noted as key (Grudeva and Churilina, 2019; Yagunova, 2010).

In perceptual experiments, informants tend to include expressions in their sets that are not actually present in the text but convey its main idea (Grudeva and Gubushkina, 2020). Nevertheless, the study by Svetozarova and Shtern (1989) confirms that, despite the high variability in the sets of keywords, a core emerges from the participants’ responses, consisting of words chosen by all participants.

The uneven distribution of keyphrases within the structure of the original text also warrants consideration (Sakharny and Shtern, 1988). Words and collocations present in the title, abstract, introductory, and concluding sections of a text possess particular informativeness. This distribution is linked to the fact that these sections contain key information that facilitates rapid comprehension of the text’s main idea and content (Kodzasov and Krivnova, 2001; Vasilyeva and Konkov, 2015). This is particularly characteristic of short texts, which are often used in perceptual experiments, where the theme is usually established at the beginning. Cases where keyphrases first appear at the end or in the middle of the text are less frequent. However, if substitute words (synonyms, pronouns, etc.) and the fact that words can appear repeatedly in the text are taken into account, their distribution becomes more even (Svetozarova and Shtern, 1989).

2.6. Keyphrase set size

Figure 8. Keyphrase set size

Рисунок 8. Объём набора ключевых выражений

Many researchers, when deciding on the necessary volume of keyphrase sets for a text, rely on the approach proposed in (Sakharny and Shtern, 1988). Conducted experiments show that participants can work with virtually any number of expressions, but the most convenient predefined range, especially when dealing with spoken text, is a set of 7-10 words or collocations. Based on this, researchers recommend extracting 5 to 15 expressions depending on the text volume (Sakharny and Shtern, 1988). An extension of this approach is the identification of small, medium (sometimes), and large sets of keyphrases (Svetozarova and Shtern, 1989). Results obtained from analyzing 20 texts of varying sizes (40-400 words), structures, and genre-stylistic affiliations indicate that the sets identified by auditors range from 5 to 20 expressions.

Various publishers and conference organizing committees typically recommend authors to provide 3-5 to 15 terms (Kamshilova, 2013). For example, when submitting an article to the international scientific journal Terra Linguistica^[3] authors are required to provide 5 to 7 keyphrases.

While each additional keyphrase offers an opportunity to attract readers to an article (Abramov, 2011), accuracy in their selection is equally important. The chosen elements must carry fundamental information about the document (Kamshilova, 2013).

Thus, the boundaries for keyphrase sets are generally established within the range of 3 to 15 expressions (see Figure 8).

2.7. Manual and Automatic Keyphrase Extraction

Keyphrases can be extracted manually or automatically. In the former case, extraction can be performed directly by the text’s authors (e.g., when submitting an application for a scientific conference), by subject matter experts, or by participants (Vanyushkin and Grashchenko, 2018). Schematically, the methodology for manual keyphrase extraction involves the following: after familiarizing themselves with a document, a participant selects a few words that, in their opinion, convey the main content of what has been read (or heard) (Sakharny and Shtern, 1988).

Discrepancies can arise where keyphrases identified by participants or even authors are either sparsely represented or entirely omitted within the source text (Mitrofanova and Gavrilik, 2022). In such cases, we speak of “present” and “absent” keyphrases in the text (Chen et al., 2020).

Manual keyphrase extraction is a highly labor-intensive process. Consequently, it is poorly suited for processing large volumes of text and working with corpus data. Furthermore, human selection is always subjective, as it depends on the reader’s/listener’s interpretation of the text.

Automatic keyphrase extraction, in a simplified sense, involves selecting words and collocations that most accurately convey the content of the document under consideration. In essence, this refers to the automatic identification of non-randomly occurring elements that are significant for the sample in the context of the overall dataset (Bolshakova et al., 2011). The extraction process occurs without human intervention and depends on the chosen model (Zhang et al., 2008). The significance level of an expression as a keyphrase is determined by its relative frequency of occurrence in experimental participants’ responses (for manual extraction) or is calculated based on a specific algorithm (for automatic extraction) (Bolshakova et al., 2011).

Automatic keyphrase extraction methods are analyzed in detail in Section 3.

3. Automatic Keyphrase Extraction

3.1. Overview and Limitations of Automatic Keyphrase Extraction in Written Text

Automatic keyphrase extraction from text plays a significant role in Natural Language Processing (NLP). It is applied in tasks such as creating automatic information retrieval systems, document classification and categorization, content summarization and annotation, and others.

In recent decades, the volume of research in the field of automatic KE has significantly increased due to the greater availability of computational and informational resources (Vanyushkin and Grashchenko, 2017). Various automatic methods for keyphrase extraction exist, including statistical, hybrid (linguostatistical), machine learning-based, and structural approaches (classifications will be discussed in Section 3.3.). However, despite the diversity of approaches, the process of automatic keyphrase extraction still faces several challenges.

Hasan and Ng (2014) note the following factors influencing the complexity of keyphrase extraction.

• Volume of the source text. The longer the text, the more potential keyphrases there are, and the more difficult it is to extract them.
• Structural consistency. While keyphrases typically anchor to specific sections within structured documents, the disruption of this expected structural integrity challenges extractive algorithms dependent on positional cues.
• Topic shift within the document. If the subject of discussion changes, the keyphrases associated with the topic also change.
• Interconnectedness. Keyphrases in scientific and news articles are usually interconnected, but in informal texts, topics within a single document may be unrelated, complicating extraction.

The automatic extraction of multi-component keyphrases, particularly of certain types of lexical categories, presents a particular challenge (Sheremetyeva and Osminin, 2015). For example, noun phrases can include various parts of speech (nouns, adjectives, numerals, or their combinations). Consequently, correctly extracting keyphrases requires consideration of the morphosyntactic and semantic characteristics of the language.

Furthermore, automatic methods often require a large amount of training data. For the correct operation of ML algorithms, thematic text corpora with annotated keyphrases are necessary. Although the use of corpora enhances the accuracy of the keyphrase extraction procedure, applying such methods is often complicated by the absence of a corpus for a specific subject area or language.

Statistical algorithms also encounter different kinds of limitations in the task of automatic keyword extraction. While they do not require specialized linguistic knowledge bases and operate relatively quickly, they often yield unsatisfactory results. The scope of effective use for statistical methods is limited to morphologically poor languages, where the frequency of different word forms for a single lexeme is relatively high due to a comparatively small number of grammatical changes and inflections. These languages include, for example, English, Danish, Modern Chinese, and Vietnamese (Tsarfaty et al., 2013). Due to this characteristic, statistical methods may not be suitable for working with languages that have rich morphological systems (Russian, French, German, Arabic, Turkish, and others), where each lexeme has numerous word forms with low frequency in any given text.

Methods for automatic keyphrase extraction are developed for specific tasks, application domains, and languages. For instance, many available algorithms are designed for the English language and therefore require adaptation when working with other languages. Moreover, the choice of method depends on the functional style of the text. For example, literary texts may contain a large number of metaphors, which are difficult to extract using statistical methods. This suggests that there is likely no universal method that can effectively work in all situations.

Thus, the applicability of various automatic keyword extraction algorithms is limited by a number of factors. To create effective methods, it is necessary to consider the typological characteristics of natural languages, subject domains, as well as the availability of linguistic and software resources.

3.2. Stages of Automatic Keyphrase Extraction in Written Text

The study by Vanyushkin and Grashchenko (2016) outlines three sequential stages for keyphrase extraction algorithms.

The first stage is text preprocessing. Preprocessing should be performed at all levels (graphemic, morphemic, lexical). Its goal is to represent the text in a format conducive to subsequent keyphrase extraction (Troshina, 2025). Differences arising at this stage are related to the necessity of working with different languages and/or are dictated by the specifics of a particular task. The following procedures may be applied to the text: tokenization (identifying word forms in the text), lemmatization (reducing various word forms to a single representation – the base form or lemma), markup removal, lexical normalization, part-of-speech tagging, and the use of a stop word list (removing a predefined list of stop words that lack semantic load, either generally or for a specific text). Performing text preprocessing procedures primarily involves the use of linguistic databases and dictionaries, which complicates the automation of this stage.

The second stage involves the recognition (classification) of keyphrase candidates obtained in the previous stage, or the filtering of candidates. The recognition procedure, as well as the use of various corpora and dictionaries, depends on the chosen automatic algorithm. Consequently, the procedures performed at the second stage may or may not be language-dependent.

The third stage consists of the postprocessing of the obtained set of keyphrases. At this stage, the set is brought into the required format – truncated, ranked, ordered, etc.

A similar universal scheme for text keyphrase extraction is proposed in the work by Vinogradova and Ivanova (2016). The first stage involves text preprocessing. The second and third stages involve the selection of keyphrase candidates and their filtering according to predefined criteria (which depend on the chosen automatic method). The fourth stage is the final selection of keyphrases from the set obtained after completing the previous steps.

In the study by Sheremetyeva and Osminin (2015), the procedures of forming and subsequently filtering a list of potential keyphrases are identified as common to various automatic algorithms. The extraction and ranking of keyphrase candidates are also discussed in the research by Wienecke (2020).

Thus, in general terms, the principle of keyphrase extraction using various automatic methods follows a unified logic, involving the sequential execution of text preprocessing procedures, keyphrase classification, and their presentation in a format determined by the task (see Figure 9).

Рисунок 9. Этапы автоматического извлечения ключевых выражений из письменного текста

This section reviews several classifications of automatic keyphrase extraction methods in written text. A number of researchers divide automatic algorithms into just two main groups, including statistical and machine learning-based (Chen and Lin, 2010), or statistical and hybrid, with the latter further considering sub-groups of graph-based and ML-based methods (Sheremetyeva and Osminin, 2015).

Vanyushkin and Grashchenko (2016) propose a more granular classification based on keyphrase recognition systems. Such systems categorize input data into two classes based on whether an element belongs to the keyphrases. Based on this, they suggest classifying methods along three facets by:

• type of learning (unsupervised, self-learning, and supervised)
• type of linguistic resources required (without usage, dictionary-based, ontology-based, based on annotated and unannotated corpora)
• type of mathematical apparatus of the keyphrase recognition system (statistical, neural network-based, hybrid, structural; the latter are further divided into graph-based and syntactic).

Morozov et al. (2023) also discuss classification by type of learning. Keyphrase extraction algorithms are divided into unsupervised and supervised learning methods. The first group of methods uses features to rank words in a text and select the required number of words with the highest rank. Statistical and graph-based methods operate in this manner. Supervised learning methods require a training set consisting of texts and a gold-standard list of keyphrases extracted from them (the set of standards is usually compiled by an expert) for preliminary algorithm tuning.

The authors of a previous study (Koloski et al., 2021), followed a similar approach, using supervised machine learning techniques to extract keyphrases from materials not only in the Russian language, but also in several low-resource languages (Estonian, Croatian, and Latvian).

In the subsequent analysis, we adhere to the classification of automatic keyphrase extraction methods into statistical, hybrid (linguostatistical), machine learning-based, and structural (graph-based). We then examine each of these groups in more detail and provide examples of specific algorithms.

It is important to acknowledge that the categorization of KE methods into distinct groups is inherently somewhat arbitrary, as many modern approaches combine elements from multiple paradigms. The categories we employ should therefore be understood as useful heuristics highlighting dominant characteristics rather than mutually exclusive boxes.

3.3.1. Statistical Methods for Keyword Extraction

As previously indicated in Section 3.1, statistical extraction methods are subject to inherent limitations. These methods are based on analyzing the frequency of words and their combinations within a text. While this approach does not necessitate prior training and allows for facile application to novel data, rendering them versatile, frequency alone is not always the most decisive factor in identifying the most important expressions for a text, which means these methods may be less accurate than those considering semantics and context (Sheremetyeva and Osminin, 2015). Nevertheless, the relevance of statistical methods increases when studying new subject areas and their terminology (Vinogradova and Ivanova, 2016).

In its simplest form, the overall frequency of word forms, reduced to their root (lemma) through stemming, is calculated. Word forms that are ranked by frequency and exceed a given threshold are considered to be keywords (Sheremetyeva and Osminin, 2015).

Chronologically, the first methods for automatic keyword extraction, discussed for instance in works by Piotrovsky et al. (1977) and Luhn (1957), were based on ranking all word forms (lexemes) found in a text by their frequency and selecting keyword candidates based on this principle.

Statistical methods include TF-IDF, Chi-squared, Log-Likelihood, PMI-test, T-test, and others.

The TF-IDF (Term Frequency-Inverse Document Frequency) measure is one of the most common statistical methods (Jones, 1972). The TF-IDF weight depends on how frequently a given expression appears in a specific document relative to its frequency in other documents within a corpus. The higher the frequency in the document under investigation and the lower it is in others, the greater the TF-IDF weight. Consequently, this approach allows for assessing the significance of a word or phrase for understanding a document’s content within the analyzed corpus.

The Chi-squared criterion quantifies the statistical significance of frequency differences. Unlike TF-IDF, it can be applied to a single text or a small sample of texts without requiring additional corpora or a list of comparison documents. The method’s principle involves comparing observed frequencies with expected frequencies based on a hypothesis. A significant difference between observed and hypothesized frequencies indicates that the hypothesis is not supported. Constructing a co-occurrence matrix of expressions within a text enables a more precise estimation of the statistical significance of frequency differences. It also reduces the influence of low-frequency words that might receive disproportionately high scores due to matrix sparsity (Krasavina and Mirzitova, 2015).

The Log-likelihood association measure is typically used to quantify the degree of systematic association between parts of collocations (Zakharov and Khokhlova, 2014; Brezina et al., 2015). Based on the actual parameter values of data and their probabilistic model, this method yields parameter values that most accurately correspond to reality (Mitrofanova and Gavrilik, 2022). When applied to keyword extraction, the Log-likelihood association measure assesses how often words actually co-occur in a text compared to their random co-occurrence probability. If the probability of joint occurrence is significantly higher than expected by chance, it may indicate that these words form a keyword expression.

The PMI-test (Pointwise Mutual Information) metric measures the degree of association between an expression and a text or corpus. A high PMI score suggests that the expression frequently appears in the document or corpus and may therefore be a keyword.

The T-test method determines whether there is a significant difference between the average frequency of an expression’s occurrence in the studied text and a control text. A significant difference indicates that the expression is a keyword for the text.

3.3.2 Hybrid (Linguostatistical) Methods for Keyword Extraction

Hybrid methods for keyword extraction combine statistical and linguistic approaches, leveraging the advantages of both. Such algorithms can produce more accurate results than purely statistical methods because they consider various features for assigning keyword status to expressions:

• semantic (e.g., analysis of semantic relationships between words, such as synonymy, antonymy, and hyponymy);
• syntactic (e.g., analysis of sentence syntactic structure);
• morphological (e.g., part-of-speech tagging);
• lexical (e.g., frequency of expression usage in a specific domain or text type).

Furthermore, some hybrid methods can incorporate features specific to particular domains or text types. For instance, when extracting keywords from medical texts, the presence of medical terminology can be considered. However, hybrid methods are more complex and computationally intensive. Their effectiveness can also strongly depend on the quality of the training data used.

The RAKE (Rapid Automatic Keyword Extraction) method is domain-independent and does not require prior model training or corpus usage. It assigns weights to each word in a text based on its frequency and position. The original RAKE algorithm was developed for English (Rose et al., 2009). Versions adapted for Russian exist, accounting for features, such as cases and declensions (Moskvina et al., 2017). The algorithm performs syntactic analysis and tokenizes the text by punctuation and stop words. In the Russian-adapted algorithm, grammatical rules of a Russian-language syntactic parser based on NLTK (NLTK4RUSSIAN) can be used to define syntactic group boundaries. For each token, the following metrics are calculated: its frequency in the text and its degree of co-occurrence with other tokens (co-occurrence measure). After calculating these metrics, each token is assigned a weight, computed as the ratio of its occurrence frequency to its co-occurrence degree. Tokens with the highest weights are considered semantically more significant for the document and are identified as keyphrases.

The YAKE (Yet Another Keyword Extractor) algorithm (Campos et al., 2020) is based on RAKE principles. However, it does not use stop words for tokenization and considers a greater number of metrics for calculating keyword weights:

• normalized word frequency;
• word position in the text;
• number of sentences containing the token;
• number of capitalized occurrences (casing);
• word relatedness to context; it is assumed that if a token has a high degree of similarity, it is less important for the context.

Keyphrases are selected based on weight ranking. Since YAKE extracts ready-made and frequent fragments from a text, it is suggested for use in summarization and similar tasks (Morozov et al., 2023).

For English texts, the Topia algorithm is used, extracting keywords based on text tokenization and morphological analysis of a corpus. For Russian, the analogous library RuTermExtract has been developed. The primary difficulties encountered when working with it are incomplete rules and ambiguity during morphological parsing.

Hybrid methods also include SpaCy, an advanced library for Natural Language Processing. This library enables keyphrase extraction by considering word frequency or the weight of distributed vector embeddings. Additionally, there is the linguistic toolkit PullEnti, which extracts keywords based on rules that consider morphological (morphological parsing is performed for each token separately) and semantic features.

Overall, the application of hybrid KE methods, combining statistical and linguistic approaches, allows for accounting for a greater number of factors and improving the accuracy and completeness of results compared to using only one of these methods.

3.3.3. Machine Learning-Based Methods for Keyphrase Extraction

Machine learning-based algorithms are sometimes categorized as a subgroup of hybrid methods (Sheremetyeva and Osminin, 2015). Such methods typically require a corpus with pre-labeled keyphrases to create a training set and build a classifier. During training, keyphrases are marked as positive examples, and all other elements are marked as negative. Word vectors are generated based on specific parameters (e.g., TF-IDF measure, word length, word position in the title/first paragraph/last paragraph). The probability of a word combination belonging to the set of keyphrases is calculated, and a threshold for inclusion is set. After training, the model computes word relevance based on parameter vectors and determines the probability that a word combination expression is also a keyphrase.

ML-based methods include decision trees (Moskvina et al., 2018), Bayesian methods (Rose et al., 2009; Sheremetyeva and Osminin, 2015), Support Vector Machines (Vanyushkin and Grashchenko, 2016), and neural networks (Abrosimov and Mosyagina, 2022).

KeyBERT is a contextualized predictive model of distributed vectors, based on the original BERT model (Devlin et al., 2019). The method relies on a bidirectional transformer that converts sentences and texts into vectors to reflect their meaning. KeyBERT determines the cosine similarity of potential keyphrase vectors relative to the text in which they are presented and uses pre-trained models to calculate word relevance in context (Grootendorst, 2020). First, the text is processed using a pre-trained BERT model. Each word in the text is transformed into a feature vector representing its semantic meaning. Next, the algorithm calculates word relevance by comparing their feature vectors with the feature vectors of other words in the text. Relevant words are grouped into clusters, each representing a potential keyword expression. From each cluster, the most representative word is selected and combined with other relevant words to form a keyphrase.

In addition to KeyBERT, other ML-based methods for keyphrase extraction exist. The KEA algorithm is quite common, based on a probabilistic classification model (Moskvina et al., 2017; Moskvina et al., 2018). Its effectiveness for the Russian language has been clearly demonstrated in a study by Sokolova and Mitrofanova (2018).

3.3.4. Structural (graph-based) Methods for Keyphrase Extraction

Structural methods, sometimes considered within hybrid KE approaches (Sheremetyeva and Osminin, 2015), are based on analyzing the text’s structure and the relationships between its elements. These methods employ various algorithms to construct word graphs based on their relative positions in the text and identify keyphrases based on their significance within the graph. The vertices of the graph represent keyphrase candidates, and the edges connecting them are weighted according to the proximity of these candidates.

Methods differ in their candidate selection and proximity determination techniques, which are based on statistical, morphological, syntactic, and sometimes semantic analysis. One common structural method is TextRank. It is based on the PageRank algorithm, used by Google’s search engine for ranking web pages (Mihalcea, 2004). The core idea of TextRank is to form a weighted graph where nodes represent tokens, lemmas, or phrases, and edges reflect textual relationships, with weights indicating the strength and/or type of semantic connections. Keyphrase candidates with high ranks are considered more significant.

Other graph-based approaches include, for instance, the language-independent algorithm DegExt. In DegExt, stop words are filtered before graph construction. In the graph, edges between vertices connect only words that are adjacent in any sentence and are not separated by punctuation (Litvak, 2013).

Some structural algorithms can utilize hypergraph models. In a hypergraph, any subset of vertices can form an edge. Such methods can be effective for large texts but are complex to implement and require significant computational resources.

3.3.5. Keyphrase generation

Keyphrase generation is a new approach that has emerged with the advent of Large Language Models (LLMs). LLMs can enrich and expand the keyphrase sets assigned to the text and predict topically determined lexical units that are observed or unseen in a given text. The applicability of keyphrase generation was evaluated primarily for the English datasets (Boudin, Aizawa 2025). Figure 10 shows that pretrained LLMs efficiently predict keyphrases present or absent in the texts, the core difference from keyphrase extraction being operations on embedding vectors.

Рисунок 10. Процесс предсказания ключевых выражений (Umair et al., 2024)

In recent works promising results were obtained by zero-shot and few-shot prompts for keyphrase generation by models from the GPT family. The paper (Martínez-Cruz et al., 2023) focuses on GPT models advantages in keyphrase generation taking into account domain adaptation and long document processing. The development of prompt engineering methods has shown advantages of zero-shot and few-shot prompts in keyphrase generation (Song et al., 2023). It was shown that the choice of prompt type and design improves the results, with a few-shot approach being more effective than zero-shot prompting.