Today, the rapid globalization of the information space leads to the rise of huge arrays of text data on information resources, including unstructured data. Therefore, developing new and improving existing methods and techniques for finding necessary and relevant information from this text data is important. This article is devoted to solving an urgent and important task related to conceptualization and further formalization in the form of a network of terms of unstructured data contained in thematic information flows distributed on the Internet.
With the beginning a rapid development of IT-technologies and globalization of an information space the huge amount of text data is produced on information resources every day. Of course, the part of unstructured data is growing among these data. This fact complicates the search for necessary and relevant information. So, the huge volumes of information flows and dynamic text arrays, which are related to some problem subject domain, determine the relevance of the process of data conceptualization and their subsequent formalization in the form of a certain ontological model. Therefore, it is important to develop new and improve existing methods that are used to solve the task described above.
Many tasks arise when working with textual information flows lie at the intersection between mathematical sciences and linguistics. This fact opens wide opportunities to apply a powerful mathematical and linguistic theory. For example, the application of knowledge in the field of discrete mathematics makes it possible to present a text data in the form of a network model that is convenient and effective to use. In terms of a complex network, the texts of a certain thematic orientation can be presented in the form of a network of words and phrases that are connected by a formal semantic connection. The network built from key terms (hereinafter network of terms) is one of the forms of this network model. In this network, the nodes are related with the single key terms of some subject domain, and the edges correspond to the links between these terms. Analysis of such networks can be a basis for decisions making in chosen problem subject domain
But while building the network of terms, the identification and extraction the key object (the key words and phrases) are open and unsolved problem. Due to the sparsity of text data and complex semantics of natural language, the determining of the syntax and semantic connections between nodes that correspond to the terms in the text, and the determining the direction of these connections (links) and their weight values are also open problem of conceptualization. The automatization of processes described above and their further visualization are no less important.
The aim of this work is to propose a new method for extracting the key terms of thematic text corpus and determining the directions of links between nodes in the undirected network of key words and phrases to build the terminological ontologies in the form of a directed weighted network of terms. Further these networks can be used to make constructive conclusions about the network structure and its parameters, and on this basis make effective decisions in the accordingly considered problem subject domains that are related to the texts. 2
Text data is a part of natural language. While using a natural language many problems arise first of all connected with ambiguity, non-compositionality and self-application.
This is due to the natural language contains a different forms of words (word forms that have a common basis) and linguistic expressions that are used for expressing different content; thus, in specific situation the meaning of these word will be depend on the context. This language is also called an Inflected Language. Non-compositionality is caused by the lack of rules in natural language that allows to determine the accurate meaning of a complex statement without knowing its context but knowing the meaning of all other components of words in the statement. This is because in the statement some phrases can be interpreted ambiguously.
While building terminological ontologies of subject domains based on a certain
thematic text document [
In this work, the most common approaches for preprocessing of text data such as tokenization and stop-word removal are used.
Tokenization is used for preliminary lexical analysis and segmentation the text on
elementary units (tokens). As independent unit, a token is some form of word. The
token is also considered in conjunction with its possible forms and meanings. The
tokenization is usually the initial stage of text processing because makes it possible to
work with the word as a separate entity while knowing its context [
To remove from the text, in particular, all preposition (for example, „an“, „the“ etc.), a stop-word removal is used. The stop-words can be considered as a source of the noise. Prepositions are common stop words particles, exclamations, conjunctions, adverbs, pronouns, introductory word, numbers from 0 to 9 (single digits), other frequently used, auxiliary and independent parts of speech, symbols, punctuation marks. Relatively recently, this list has been supplemented by sequences of characters as www, com, HTTP, etc. that such frequently used on the Internet.
All mentioned above pre-processing methods can be easily applied to different types
of texts. It can be done using the standard Python libraries such as Python NLP (Natural
Language Processing), in particular, NLTK [
Besides that, the Part-of-Speech tagging approach (or just tagging) (fig. 1) is proposed to use to extrapolation of syntax and structure of text. Tagging is usually the next step of natural language processing. Tagging is used after tokenization, and is to refer the word in the text (corpus) to a certain part of speech. This step is based on both a definition of the word and a context of the word. In other words, it is based on the connection of the word with adjacent and related words in a phrase, sentence, or paragraph. Also, Part-of-Speech tagging is one of the main and basic components of almost any NLP task. The collection of tags assigned to each word in the sentence is used for this task. PoS tagging can be used for word indexing, information retrieval and also for many other applications. PoS tagging can be especially useful if some words or tokens can have multiple tags. And most importantly, tagging simplifies the context that refers to some subject domain.
Parts of speech are also known as word classes or lexical categories (which are based on the syntactic context of a phrase). Then, we tag each word according to its lexical category using the above method of classifying words by parts of speech.
E. Brill's parser [
To extract keywords from the text it is necessary to assign them a certain numerical
assessment (in other words, a statistical indicator of importance). In [
The building of a directed network of terms is made within every separate sentence.
In this work, separate functions such as “word_tokenize” and “pos_tag” of a specialized Python add-in, the module NLTK (Natural Language Toolkit that is an opensource library), are used to automatic tokenization and Part-of-Speech tagging to assign the tag to every word, accordingly.
Also in this paper, in addition to the standard sets of stop words, which are available
by reference [
The method for determining of key words and phrases, and also directions of links
in the undirected network of terms proposed in this work is based on using the process
of classification of words on parts of speech and the corresponding tagging of parts of
speech (Part-of-Speech tagging). The practice research shows [
Then, with the help of the global frequency of the term GTF, the idea of which is described above, the statistical weighing of words and phrases that is a part of the sequence formed at the previous stage is carried out.
The so-called tuple is formed for each word, in the order of its occurrence in the text. Each element of the tuple consists of three values: the first one is a term (a word or phrase); the second is a tag that is assigned to a word depending on its belonging to a certain part of speech; the last one is numerical value of a GTF.
It should be noted that the GTF is calculated considering two previous values – the word or phrase and part of speech to which this word or phrase relates. The number of similar tuples in the whole text, which normalized by the total number of formed terms in this text defines the value of the third element.
At the next step, it is proposed to determine the undirected links between terms in
the text. For this goal, the Horizontal Visibility Graph algorithm (HVG algorithm) that
transform time series into a visibility graphs is used [
If there is an undirected link, determined by the above algorithm, between the nodes from ti to tj of the time series then: ─ it is proposed to determine the link from node ti to tj, if in a sentence the word (not a phrase) to which the node ti corresponds occurs earlier than the term (word or phrase) to which the node tj corresponds; ─ it is proposed to determine the link from node tj to ti, if in a sentence the phrase (not a word) to which the node tj corresponds occurs earlier than the term (word or phrase) to which the node ti corresponds (fig. 2).
Given the above-described principle of forming the sequence of the terms and the
proposed rules for determining links, it can be noticed that words and phrases will be the
part of the corresponding phrases that have more words. In other words, a significant
part of phrases with more words is only an extension of the corresponding phrases and
words. A similar principle of the building of directed networks of words, the building
of networks of natural hierarchies of terms, proposed in the work [
The weight of links between nodes of the directed network of terms is determined
by the principle proposed in the work [
The proposed methodology for computerized processing of text corpus was tested on the example of a children’s allegorical story-tale, “The Little Prince” by Antoine de Saint-Exupéry.
According to the methodology proposed above, the selected text document was processed and key terms were identified (Table 1).
If we arrange all the key terms in descending order of their numerical value GTF, then
the graph (Fig. 3) shows Zipf's law [
Also, the following parameters of the built network were obtained applying the Gephi software: the number of nodes is 79; the number of links is 117; the average degree is 1.48; the average path length is 3.74; the average clustering coefficient is 0.012; the network density is 0.019; the number of connected components is 4.
The list of the most important links between the corresponding nodes in the network of terms is presented in Table 2. In this work, a new method for extracting of key words and phrases from thematic information flows and a new method for determining the directions of links between nodes in undirected networks of terms with using Part-of-speech tagging were proposed. Also, the holistic methodology of computerized text corpora processing and building the directed weighted networks of terms (of key words and phrases) is presented. Using previous words' classification process into parts of speech (Part-ofspeech tagging) the key words were extracted. The proposed methodology for computerized processing of text corpus was tested on the example of a children’s allegorical story-tale, “The Little Prince” by Antoine de Saint-Exupéry. The most important links between the corresponded nodes in the network of terms corresponding to certain key concepts in the considered text were revealed after analyzing the results of the methodology. The terms such as “little”, “prince” and “little_prince” turned out the key within the proposed ontological model. These terms also correspond to the name of the considered text document. As expected, the most important links between the key terms are "little → little_prince" and "little → prince".