=Paper=
{{Paper
|id=Vol-2258/paper44
|storemode=property
|title=Multimodel Method of Rubricatng the Unstructured Electronic Text Documents
|pdfUrl=https://ceur-ws.org/Vol-2258/paper44.pdf
|volume=Vol-2258
|authors=Maxim Dli,Olga Bulygina,Pavel Kozlov
}}
==Multimodel Method of Rubricatng the Unstructured Electronic Text Documents==
https://ceur-ws.org/Vol-2258/paper44.pdf
Multimodel method of rubriсating the unstructured electronic
text documents
M I Dli1, O V Bulygina1 and P Yu Kozlov1
1
The Branch of National Research University “Moscow Power Engineering Institute” in
Smolensk, Smolensk, Russia
Abstract. Analysis of electronic text documents written in natural language is one of the most
important tasks implemented in systems of automated analyzing the linguistic information. It is
known that text documents can be characterized by different parameters: size, presence of a
structure, frequency of references keywords, etc. Depending on these parameters, a procedure
of selecting the type of the rubrication model is proposed. As a result, a multi-model approach
to document classification, characterized by the combined use of artificial neural networks,
growing pyramidal networks and probabilistic-statistical methods, is proposed. Its application
will improve the accuracy of attributing the electronic text documents to concrete rubrics,
taking into account their specificity and various objectives of practical application in the
organization.
1. Introduction
The task of unstructured information processing is one of the priority problems, since its solution is
necessary to achieve the goal of transition to a digital economy, formulated in Strategy of the
Information Society Development in the Russian Federation for 2017-2030. The program "Digital
Economy", approved by the Government in July 2017, is considered as one of the most important
tasks to improve accessibility and quality of public services for citizens. To solve this task it is
necessary to improve the procedures of electronic information exchange between citizens and public
services. The effectiveness of this interaction is assessed using such indicators as operativeness of
reaction to the citizen requests and satisfaction with the results of provided services. Both indicators
directly depend on the processing speed of incoming information requests and the accuracy of
message recognition.
The fact that information in the citizens' queries is presented in unstructured form does not allow
using classical methods of processing relational, object, hierarchical and marked text data. The
methods of semantic search can be recommended for the analysis of unstructured text information. But
their application is complicated by the specifics of information messages containing citizens' requests
and the peculiarities of functioning of the state services processing these messages.
The foregoing stipulates the need to develop methods of analyzing the electronic unstructured text
documents (EUTDs), taking into account the specifics of its content and usage in the system of
electronic public services.
2. Selection of an approach to the EUTD rubrication
2.1. The concept of an electronic unstructured text document
The main features of EUTD are the following:
366
the text is written in natural language;
there is no explicitly defined structure;
there are the grammatical and syntactic errors in the text;
automatic allocation of a structure can’t be performed in an unambiguous way.
The examples of unstructured text document types can be: books; magazines; metadata; medical
records; audio and video materials; analog data; Images; files based on unstructured text (e-mail
messages, web pages, documents created using word processors) 0.
EUTDs can include documents in such formats as html, doc, rtf, etc. Usually the content of such
EUTDs is presented to the user for examination by the Internet browser or any text editors (e.g.
Microsoft Word).
2.2. Analysis of the EUTD rubrication methods
Currently, information systems can use a sufficiently developed methodological apparatus of the
theory of classifying the objects of various types (in this case a document is regarded as an object of
classification or rubrication) 0.
By the criterion of learning stage existence in the EUTD rubrication model, all methods can be
conditionally divided into two groups: with and without training.
The EUTD rubrication methods with a teacher suggest the availability of complete information
about the rubricating field (number and characteristics of columns). In this case, a "teacher" is a
prepared sample of EUTDs (training set) with the concrete degree of belonging to one or another
rubric.
Currently, the rubrication methods with a teacher are widely used in information systems and are
suitable for solving a large number of emerging tasks. However, these methods require a large amount
of preliminary information for teaching the rubrication model.
The rubrication methods without a teacher analyze the collection of EUTDs for classifying them in
such a way that each rubric contains documents closest to the chosen metrics. In these methods there is
no need for initial training of the models "with a teacher", since the characteristics of the rubric field of
text documents are not known in advance.
In general, the rubrication algorithms without a teacher (clustering algorithms) group EUTDs in the
feature space (document parameters) with the subsequent interpretation of the result. The EUTD
classification is based on the hypothesis that text documents close in meaning are relevant to the same
queries and selected rubrics.
The papers 0,0 consider a wide class of various modifications of the EUTD rubrication algorithms
based on the decision rules, artificial neural networks, growing pyramidal networks and probabilistic-
statistical methods.
Analysis of the prospects of using the automated ENDD rubrication methods has showed that a
multimodel approach is required to solve this task. It involves the EUTD classification using various
intellectual analysis algorithms and taking into account the specificity of each document.
2.3. Criteria of selecting the ENDD rubrication method
To select of the rubrication method, it is advisable to take into account the following features: the
rubric field characteristics, the amount of accumulated statistical information and the rubric thesaurus
characteristics.
The rubric field characteristic shows the degree of interconnection of their thesauri. The presence
of interrelated rubrics is the most difficult condition for the rubrication.
The amount of accumulated statistical information determines the possibility of constructing the
various rubrication models. Thus, if it is possible to correctly use the probabilistic-statistical methods,
the amount of EUTD can be considered "sufficient".
The rubric thesaurus characteristic determines whether the composition and influence degree of
significant words of their dictionaries change. Thus, some rubrication models can’t be used in the
367
conditions of dynamically changing thesauri because of the complexity of retraining and restructuring
of the rubrication model structure.
The degree of rubric thesaurus intersection K∩ depends on the number of unique words for all the
rubrics and defined as:
1 J J Count ( Ri R j )
K ,
J j 1 i 1 Mj
where J – total number of rubrics, Count ( Ri R j ) – number of matching significant words in the
thesauri of the rubrics Ri and Rj, Mj – total number of significant words in the thesaurus of the rubric
Rj.
The following criteria levels of the thesaurus intersection have been taken on the basis of research
results: K∩ < 0,15 – insignificant level; 0,15 ≤ K∩ ≤ 0,4 – medium level; K∩ > 0,4 – significant level.
In addition, it is advisable to take into account the EUTD size: short (contain less than 10% of
unique significant words), medium (10-20% of unique significant words), large (more than 20% of
unique significant words). Concrete values of the EUTD sizes are specified by experts when setting up
the automated rubrication system.
3. Multimodel method of the EUTD rubrication
3.1. Typical situations of the EUTD rubrication
In accordance with the indicated EUTD characteristics, the typical situations of their rubrication have
been outlined and methods of their conducting have been provided.
Typical situation 1. The analysis of short or medium-size EUTDs is performed in conditions of
medium or significant degree of rubric intersection and in case of statistical data insufficiency for the
effective use of traditional mathematical models of the EUTD rubrication. In this situation, it is
advisable to use fuzzy pyramidal networks that allow classifying objects without a teacher in the
absence of statistical data 0. In contrast to the known 0, this apparatus allows taking into account the
degree of rubric cohesiveness based on the fuzzy-logic algorithm use that increases the accuracy of the
EUTD rubrication.
Typical situation 2. The analysis of medium-size EUTDs is performed in conditions of insignificant
or medium degree of rubric intersection and in case of statistical data insufficiency for the effective use
of traditional mathematical models of the EUTD rubrication. In this case it is advisable to apply the
model using weight coefficients that takes into account, in addition to statistical characteristics, expert
information in the processing of the EUTD rubrication. In contrast to the known 0, this model should
take into account the degree of significance of words in the EUTDs, depending on the appearance of
new significant events affecting on the rubric thesauri.
Typical situation 3. The analysis of medium-size EUTDs is performed in conditions of medium or
significant degree of the rubric intersection and in case of statistical data insufficiency for the effective
use of traditional mathematical models of the EUTD rubrication. This situation determines the
expediency of applying the rubric selection model based on a fuzzy decision tree that will improve the
accuracy and efficiency of the rubrication with the use of additional expert information. In contrast to
the known 0, this model should use syntax links and word roles, the fuzzy evaluation of document
differences in the n-dimensional space of the text attributes in the construction and application of the
fuzzy decision tree for referring the document to specific rubric in conditions of the interdependence of
their thesauri.
Typical situation 4. The analysis of short-size EUTDs is performed in conditions of insignificant
degree of rubric intersection and in case of sufficient amount and statistical data quality of this
document type (for training the hybrid fuzzy models 0,0). For this situation, it is advisable to use a
neural-fuzzy classifier based on a well-known approach 0, but using expert information to determine
the keyword relevance in the formalization and the subsequent rubrication of the EUTD.
368
Typical situation 5 is characterized by the presence of large-size EUTDs, sufficient amount of
statistical data on this document type and small degree of rubric intersection. In this situation, it is
advisable to use the known probabilistic methods of the textual information analysis 0,0.
Typical situation 6. The analysis of large-size EUTDs is performed in conditions of medium or
significant degree of rubric intersection and in case of sufficient amount and statistical data quality
about this document type. In this case, it is advisable to use the voting method taking into account all
models that will reduce the error probability in the EUTD rubrication.
Table 1 contains a brief description of the typical situations for the EUTD rubrication, based on the
EUTD size, the degree of rubric intersection, the statistical data sufficiency for the application of
probabilistic analysis methods and the corresponding models for the effective rubrication.
Table 1. Choice of the EUTD rubrication model.
Situation EUTD Degree of rubric Statistical data EUTD rubrication
size intersection sufficiency model
Model based on fuzzy
medium, significant insignificant
1 short, medium pyramidal networks
Model using the
insignificant, medium insignificant
2 medium weighting coefficients
Model based on fuzzy
medium, significant insignificant
3 medium decision tree
Cascade neural-fuzzy
insignificant significant
4 short model
5 large insignificant significant Probabilistic classifier
6 large medium, significant significant Classifier voting method
3.2. Algorithm of implementing the multimodel method of the EUTD rubrication
The analysis of typical situations shows the impossibility of applying a unified model of the automated
EUTD rubrication and allows justifying the advisability of creating a multimodel method for solving
this task. The choice of a particular model of the EUTD rubrication should be based on the results of
identifying the corresponding typical situation.
The algorithm of multimodel method of the EUTD rubrication based on the combined use of
intellectual analysis models is given in Figure 1. Directly choosing a particular model or their
ensemble is carried out using the base of fuzzy product rules (R1-R6).
The method involves analysis of the set of EUTDs V={V1,…,Vk,…VK}, where each document Vk is
represented as a set of meaningful words:
Vk {v1( k ) ,...,vl(kk ) ,...,vK( k ) }, k 1..K , lk 1..Lk ,
where vl( k ) – the word, Lk – number of words in EUTD k .
k
At the preliminary analysis stage, each document Vk is registered and presented as XML linguistic
markup using a set of tags (descriptors).
After registration, the modified text document is analyzed:
Vkr Vk h (k ) ,
where h(k) – header information of document k that described by the rules.
369
Begin
Set of EUTDs
Selection of the first EUTD
Is Yes
Registration
selected EUTD Finish
k=k+1 the last?
No
Registration of documents
Preliminary
«Segmentator»
analysis
Morphological analysis
Fuzzy production system of
selecting the rubrication method
Is R1 No Is R2 No Is R3 No
satisfied? satisfied? satisfied?
Yes Yes Yes
Rubrication
Model based on fuzzy Model using weighting Model based on
pyramidal networks coefficients fuzzy decision tree
Is R4 No Is R5 No Is R6 No
satisfied? satisfied? satisfied?
Yes Yes Yes
Cascade neural- Probabilistic Classifiers Manual
fuzzy model classifier voting method rubricarion
Figure 1. Algorithm of multimodel method of EUTD rubrication.
Document Vkr is analyzed to extract the words, sentences, paragraphs in the markup language form
in a text part (block "Segmentator").
As a result, the modified EUTD is formed at the output of the block "Segmentator":
VkS Vk h ( k ) Sen ( k ) Abz ( k ) ,
where Sen(k) – set of the sentences, Abz ( k ) – set of the paragraphs.
It should be noted the complexity of implementing the segmentation procedure, the high degree of
error influence on the final results of the automated EUTD rubrication. First of all, this is typical for
procedures of the rubrication of short EUTD.
370
After linguistic marking the morphological analysis of the EUTD VkS is carried out, with the
separation of lexical characteristics of words and morphemes (the least meaningful linguistic units).
As a result of the morphological analysis the modified EUTD is formed:
VkM Vk h ( k ) M ( k ) ,
where M(k) – the results of decomposing the document k into words, sentences, paragraphs and
allocating the units and morphological characteristics attributed to the words.
The standard morphological characteristics of the thesaurus are stored in a morphological database,
which is filled up from several sources. The most important source is the set of text documents of the
Russian National Corpus containing the labeled texts 0.
The results of the morphological analysis of the EUTD are stored as a separate linguistic group in
the original document. It means that in addition to the morphological information attributed to each
word of the text, the syntactic structure is given for each sentence.
The next stage is the selection of a particular EUTD rubrication model in accordance with the
highlighted typical situations given in the table 1.
Since the values of the EUTD characteristics given the table 1 (EUTD size Lk, degree of rubric
thesaurus intersection K∩, statistical data sufficiency Vst) are determined by expert information, to
implement the procedure of selecting a specific model of the automated EUTD rubrication Mj from the
model set, it is proposed to use the base of fuzzy product rules of the following form:
IF( Vst is Vsta )and ( K is K b )and ( Lk is Lck )THEN( M is M j ),
a { insignificant, significant },
b { insignificant, medium, significant },
c { short ,meduim ,l arg e }, j { 1,2,3,4,5,6 }.
The proposed multimodel method of analyzing the electronic unstructured text documents makes it
possible to increase the accuracy of attributing text documents to specific rubric, taking into account
their specific subject area and different amount of statistical data.
4. References
[1] Zaboleyeva-Zotova A, Petrovskiy A, Orlova Yu and Shitova T 2016 Automated analysis of
news texts International Journal “InformationContentandProcessing vol. 3 no. 3.
[2] Sebastiani F 2002 Machine learning in automated text categorization ACM Computing Surveys
vol. 34 no. 1 pp 1-47.
[3] Gulin V 2013 Research and development of methods and software of text document
classification: PhD thesis.
[4] Shabanov V 2003 Models and methods of automatic text document classification: PhD thesis.
[5] Bulygina O and Okunev B 2016 Creating fuzzy network tools to analyze prospects of projects
of information and telecommunication infrastructure development Neyrokomp'yutery no. 7 pp 15-20.
[6] Sulima E and Milenin V 2010 Method of using the means of the educational material
structuring International Journal “Information Theories and Applications vol. 17 no. 4 pp 387-395.
[7] Aleksandrov М 2008 Methods of automatic classification and statistical analysis of the input
stream of text information in information systems: PhD thesis.
[8] Lewis D 1992 Representation and Learning in Information Retrieval: PhD thesis.
[9] Gimarov V and Dli M 2004 Neural network algorithm of complex object classification
Programmnye produkty i sistemy no.4 pp 51-56.
[10] Kruglov V, Dli M and Golunov R 2001 Fuzzy logic and artificial neural networks (Moscow:
Nauka, Fizmatlit).
[11] Meshkova E 2009 Development and research of hybrid neural network models for the automatic
classification of text documents: PhD thesis.
[12] Chugreev V 2003 Model of the structural representation of textual information and methods of
its thematic analysis on the basis of frequency-context classification: PhD thesis.
371
[13] Kozlov P 2015 Comparison of frequency and weight algorithms of automatic document analysis
Nauchnoye obozreniye no. 14 pp 245-250.
[14] Kozlov P 2017 Automated analysis method of short unstructured text documents Programmnye
produkty i sistemy no. 1 pp 100-105.
Acknowledgments
The reported study was funded by RFBR according to the research project № 18-01-00558.
372