Great model a model for the automatic generation of semantic relations between text summaries based on the relevant information to the user

International Journal of Computational Science and Information Technology (IJCSITY) Vol.1, No.4, November 2013

GReAT Model: A Model for the Automatic
Generation of Semantic Relations between
Text Summaries based on the Relevant
Information to the User.
Claudia Gomez P. and Alexandra Pomares Q.
Pontificia Universidad Javeriana, Carrera 7 No. 40 - 62 Colombia

Abstract.
The large available amount of non-structured texts that belong to different domains such as
healthcare (e.g. medical records), justice (e.g. laws, declarations), insurance (e.g. declarations),
etc. increases the effort required for the analysis of information in a decision making process.
Different projects and tools have proposed strategies to reduce this complexity by classifying,
summarizing or annotating the texts. Partic- ularly, text summary strategies have proven to be
very useful to provide a compact view of an original text. However, the available strategies to
generate these summaries do not fit very well within the domains that require take into
consideration the temporal dimension of the text (e.g. a recent piece of text in a medical record
is more important than a previous one) and the profile of the person who requires the
summary (e.g the medical specialization). To cope with these limitations this paper presents
”GReAT” a model for automatic summary generation that relies on natural language
processing and text mining techniques to extract the most relevant information from narrative
texts and discover new information from the detection of related information. GReAT Model
was implemented on software to be validated in a health institution where it has shown to be
very useful to display a preview of the information about medical health records and discover
new facts and hypotheses within the information. Several tests were executed such as
Functional- ity, Usability and Performance regarding to the implemented software. In addition,
precision and recall measures were applied on the results obtained through the implemented tool,
as well as on the loss of information obtained by providing a text more shorter than the original.

Keywords:
Topic Identification, User’s Profile, Text Mining, Automatic Summary Generation Techniques,
Semantic Relations
This work was supported by the project “Extraccion semi-autom´atica de metadatos de fuentes
de datos estructuradas: Una aproximacion basada en agentes y miner´ıa de datos” funded by
Banco Santander S.A. and Pontificia Universidad Javeriana.

DOI : 10.5121/ijcsity.2013.1405

61


1.Introduction
During the last thirty years the information systems have stored huge amounts of
information in different formats, in some areas or domains such as healthcare (e.g. medical
records), justice (e.g. sworn declarations), assurance (e.g. declara- tion) and insurance (e.g.
Research articles and reports). A lot of this information is stored as narrative texts,
hindering its use for the decision making processes. The process of discovering the
knowledge contained in these texts, or creating new hypotheses according to them include
a lot of time and effort [14],[24] that cannot be afforded by most organizations.
Generally, this problem remains a constant challenge due to the difficulty of organizations
to absorb and use the information they need. It should be noted that, the limits on the
reading speed of a human being make it impossible to capture the key information in a
short time when there is a large amount of text [18]. Therefore, organizations need an
increasingly intelligent use of information and technology to achieve a more efficient
management of data they accumulate [7].
Given that the information overload on an issue creates difficulties to understand it,
and also, as Feldman mentioned [33], you can find the previously unknown information
not only contained in a single text but in a collection of texts, or within related
information for example, in electronic medical records you can find a lot of information
that is contained in the annotations made by doctors such as links of interest or
comparable groups exposed to a risk factor, being very useful in medical research or
administrative analysis.
To achieve the proper management of this unstructured information (text) the typical
approach is to use information retrieval techniques as search engines do. However, the
amount of information obtained is still above what a person can handle and manage
[36]. A more efficient process for the discovery of new information or knowledge known as
Text Mining, allows the automatic extraction of different collections of documents for the
discovery of patterns or trends that are not known [15].
To cope these problems this paper presents GReAT a model for automatic generation of
semantic relations between text summaries based on the identification of relevant topics
that relies on natural language processing and text mining techniques to extract the most
relevant information from narrative texts focused on the requirements and profile of the
end user. GReAT is an extraction based model of summary generation which principle is
to identify the relevant categories for the user through the text and to identify the
information that is relevant to his profile generating an adequate, concise and high quality
text summary. In addition, it is a model that generates semantic relations between these
text summaries to provide a comprehensive overview of existing and related
information. Experimental results show the relevance and applicability on the tasks of the
users.
This paper is organized as follows: Section 2 compares GReAT with important proposals
and illustrates the main strategies of text summary generation and generation of
semantic relations, Section 3 shows the General Process of the Proposed Model 62


GReAT for the automatic generation of semantic relations between text summaries,
Section 4 describes the prototype implemented to validate the proposed model, Section 5
evaluates the functionality of GReAT through its application in a study case in the
healthcare domain and finally, Section 6 presents the Conclusions and Future Work.

2
2.1

State of the Art
Generation of Summaries

To face these narrative information overload problems, different summary generation
approaches have been proposed. There are mainly two approaches to perform this task:
Statistical and Linguistic Methods. The statistical methods are independent of the
language and are based on the frequency of words or take into account the title,
headings, position and the length of the sentence. On the other hand, linguistic methods
include discourse structure and lexical chains [41]. Some proposals of this method use
Clustering based strategies that group together phrases with similar characteristics,
however, they have had accuracy problems due to the ambiguous terms of a language.
Some commercial tools are based on basic statistical approaches, and rely heavily on a
particular format or writing style, such as the position of the text or some lexical
words[30]. Similarly, most of the tools rely on methods or features such as centroidbased, position-based, frecuency-based summarization or keywords to extract relevant
sentences to the summary such as MEAD, Dragon ToolKit [43] , LexRank [11]. There are
methods based on abstraction, which are more complicated compared to approaches based
on extraction, so there are still problems regarding semantic representation, inference and
natural language generation [42]. ML (Machine Learning) and IR (Information Retrieval)
algorithms are needed to achieve a good similarity measure between documents, since there
are many problems while computing the similarity between documents due to
ambiguities in the used vocabulary. That is, if two documents or publications dis- cuss the
same topic, they may use different vocabulary while being semantically similar.
Some aspects have been obtained through the analysis of various works that have used the
most common techniques within the area of text mining for text summary generation.
This analysis is consolidated in Table 1, which shows com- parative features related to how
to select the relevant phrases and build up the summary, including: Frequency, Topics,
Profile, Sequentiality, Duplicity and Noise. The symbol” X” indicates that the
project contains the features represented in the table and the meaning of each one of
these are described below: i) Frequency: Indicates if the project takes into account the
frequency of words to extract relevant sentences to the text summary. ii) Topics:
Indicates if the project takes into account the topics in the text that are important to the
user to generate the text summary. iii) User Profile: Indicates if the project gives
importance to the characteristics of the end-user groups to generate the text summary.
iv) Sequentiality: Indicates if the project maintains the sequentiality of the original
text to form the summary in the chronological order in which it was stored. v)
Duplicity: Indicates if the project eliminates the duplicity of information in the text. vi)
Noise: Indicates if the project eliminates the noise of natural language by bad writing or
bad typing in the text.

63

Table 1. Relevance of Sentences
Pro ject Frequency Topics Profile Sequentiality Duplicity Noise
[20]
X
X
[40]
X
X
X
X
[12]
X
X
[36]
X
[21]
X
[5]
X
X
[22]
X
[19]
X
X
[9]
X
[25]
X
X
X
X
X
[24]
X
X
[32]
X
[30]
X
X
[11]
X
[10]
X
X
X
[16]
X
[41]
X
X
X
[2]
X
X
X

In conclusion, none of the compared projects contains all of the mentioned features and
do have some limitations to be taken into account in this project with the proposed
model. Among the challenges and opportunities identified in the literature to be
considered for the automatic generation of text summaries are: i) To improve the quality
and coherence of the summaries removing the duplicity of information, ii) Adapt
summaries to every user according to their needs of information taking into account the
different topics that are often important to the end-user in a specific domain and
user’s profile. iii) Keep the sequentiality of the original document. iv) Detect noise in
text collections due to the use of natural language.

2.2

Generation of Relations

In the analysis performed on the methods for the generation of relations between terms we
found in most of these works the issue know as ”Problem vocabulary”, since different
languages are ambiguous, the knowledge of the semantic relations between all possible
words and phrases is required. To accomplish this task it is believed that one must first
solve all the other problems related to Natural Language Processing such as the natural
language understanding, common-sense reasoning and logical thinking, however, taking
into account some semantic relations to relate the generated text summaries.
In the table 2 the techniques, characteristics and strategies used in the consulted works
have been consolidated, showing some opportunities and challenges to be addressed by our
proposed model for the detection of semantic relations :
Table 2. Generation of Relations
Pro ject Taxonomy Profile Matrices Frequency Granularity Keywords
[25]
X
X
[28]
X
[17]
X
[34]
X
X
[27]
X
X
X
[26]
X
[14]
X
[1]
X
X
[31]
X

This table contains the following information:Taxonomy, Profile, Ma- trices,
Frequency, Granularity and Keywords. The symbol” X” indicates that the project
contains the features represented in the table and the meaning of each one of these is
described below: i) Taxonomy: Indicates if the project used external knowledge within
64


your solution or not. ii) User Profile: Indicates if the project takes or not into account
the information needs or characteristics of the user. iii) Matrices: Indicates if the project
stores the collection of concepts that are closely related semantically in a matrix of
similarity, inference or interest. v) Frequency: Indicates if the project performs an
analysis of the frequency of words to identify the keywords that relate the content. vi)
Gran- ularity: Indicates if the project takes or not into account or not (besides the
semantic similarity between the contents) one or more degrees of similarity between the
contents. vii) Keywords: Indicates if the project represents or not documents with
keywords to estimate similarity efficiently. In conclusion neither project contains all the
features mentioned, which are taken into account in our project to improve the detection
and visualization of relations with the proposed model. This includes the following
challenges and opportunities: i) To provide various types of relation displays allowing to
cover different information needs of users, ii) Show the relations according to the needs and
information granularities of the user, taking into account the different topics that are
often important to a person in a specific domain, iii) Display the contents which do not
have any relation, in order to find particular cases within the content, iv) Detect the high
or low degree of similarity between summaries, considering the most relevant and
similar words that the summaries possess.

3

GReAT Model

This model is divided into two components, the first one is responsible for ana- lyzing
unstructured content found in the narrative text, allowing the realization of a summary
that consolidates the most relevant information using text mining techniques and the
second one is responsible for generating the semantic relations between these text summaries
generated with the first component, according to the categories or topics of interest prior
predefined by the user. This model is named GReAT (acronym Model for Automatic
Generation of Relations between Text Summaries). This model consists of a series of
processes, initially in order to generate the narrative summaries, and subsequently to
generate the semantic relations between text summaries generated with the first
component of the model. Each one of the components is described in detail:

3.1

Component for Text Summaries Generation

Among the main concepts to keep in mind to understand the first component:
Summary: Its task is to extract a smaller document in size but keeping the relevant
information from the original document, i.e., automatically creates a comprehensible
version of a given text, providing useful information to the user [16]. GReAT is a model
that produces summaries with the following characteristics: Multi-Document, Specific
Domain, Topic Oriented and Summary Based on Extraction. It uses several techniques of
Text Mining, such as: Tokenization, Chunking, Named Entity Extraction, among others.
Multi-Document: The summary is generated from multiple input documents [42].
Specific Domain: The summary is generated from multiple input documents, but
considering the context or domain in which it was written.

65


Summary Based on Extraction: The summary is generated with phrases that are
included literally. This strategy produces a summary by selecting a subset of sentences
from the original document.
Topic Oriented approaches: It focuses on a user’s topic of interest, ex- tracting text
information that is related to the specific topic. Its approach is to identify significant
topics within the data set and generate the topical structure based on these topics [41].
Its principles are based on taking into account some considerations to extract knowledge or
generate an adequate summary such as: quality, consistency, adaptation, duplicity, user
profile, noise and sequentiality, obtained from the analysis of the state of the art detailed
in the section 2.
The processes that are part of the component are presented in Figure 1. It consists of
three main processes: Preprocessing, Identification of Categories and Extraction of
Candidate Phrases. Preprocessing
step is divided into several tasks: Stop-words,
Tokenization, Spelling and Filtering, in the following sections each one of these steps is
explained in detail.
At the beginning of the process, User Profile, Search Filter and Cor- pus(collection
of documents) are received as input information. For instance in a health domain, the
User Profile is comprised of: The keywords that identify the user’s profile, in the case of
a physician anesthesiologist, the keywords could be anesthesia, local, general. The Search
Filter is comprised of: a date range (start date and end date) of the patients’ medical
records, the categories or topics of interest on which the user wants to generate the
summary (including: Diseases, Drugs, Exams ). The keywords associated to a category
allow to specify key words that the user wants to locate, the number of sentences
required be displayed from each category and a Corpus or collection of documents that
will consist on the patients’ medical records in text narrative.
1. Preprocessing: This step is one of the most important in the text mining area, since
its results will affect the performance and quality of the data in the subsequent phases,
given the large number of words, phrases and sentences there might be a large number of
different forms for each of these elements to have in different contexts and combinations,
therefore, this process will be a transfor- mation of the input documents, in order to
extract the most significant terms or ”features” from the text. This process is divided
into five main tasks: Stop- words, Tokenization, Spelling and Filtering, which are detailed
below:
Stop-words: To address the problem of high dimensionality that commonly occurs in a
text mining process, it is proposed to delete the words with little relevance to the
language with the technique known as Stop-Words [6]. They are words that do not provide
relevant information, such as articles, prepositions, among others. This process requires a
dictionary of stops words of the language. High dimensionality means that the size and
scale of the possible combinations of the values of the characteristics of data are large in
text mining systems. The result of this phase will be the phrases of the text without the
words considered as Stop-Words.

66


Tokenization: In order to obtain a text that is shorter than the original for the
summary, it is necessary to divide the text into phrases that allow us to process
information independently [13]. The output of this phase is the set of sentences that
were divided by a tokenizer, in the case study the selected token were the tab spaces
between lines. In addition to this, to get the information as uniform as possible, the Case
folding technique is used, which seeks to convert all characters into the same kind of letter
may be uppercase or lowercase letters, in the case study they becomes uppercase [38].
Spelling: Normally, natural language can make noise, ie, misspelled words. This phase
performs a spelling check of the narrative text, which involves taking the text input and
providing a corrected text. Spell Checking is performed on the Noisy-channel model,
which models user errors (typographical) and expected user input (based on data) [8]. The
errors are modeled by the weights of the Edit Distance technique and the expected input
by the model language characters.
Edit distance technique measures the minimum number of edit operations (insertion,
deletion, and substitution) to transform one string into another. Edit distance is selected
because it can effectively capture typographic errors, words with alternative spellings, and
does not rely on the separation of word boundaries [39]. This process requires a dictionary
of words from the language and from specific domain. The result of this phase will be the
correctly spelled phrases.
Filtering: This process proceeds to remove the redundancy of information found in these
texts. To achieve this, we use a technique called Fixed Weight Edit Distance [23], where the
simplest form of weighted edit distance simply sets a constant cost for each one of the edit
operations: match, substitute, insert, delete, transpose. This general setup subsumes the
Needlman-Wunsch algorithm used in molecular biology. This algorithm will maximize the
number of matches between the sequences along the entire length of sequences. The
output of this phase is the removal of common phrases to each other, leaving only one
instance of them. This step is performed to eliminate duplication of information presented
in the specific domain of the case of study which validated the model GReAT, therefore, it is
an optional step.
2. Identification of Categories: This process consists of two tasks, Verify
Search Filter and Phrases Annotation, which are described below:
Verify Search Filter: At this point of the process the words of specific domain or
Knowlegde Base, the User Profile and Search Filter have been defined. These are
temporarily stored in the knowledge base to be taken into account in the next task. The
output of this phase is knowlegde base with words that are relevant to the user.
Phrases Annotation: Users can have different information search needs, so in this process
we will identify the most relevant information to the user according to the words
associated with the topics of interest defined in the knowledge base, user profile and search
filter defined. For this, it performs a process of annotation of the words contained in the
sentences resulting from the preprocessing phase.
67


An annotation is a layer attached to the text representation, relieving the linguistic
structure in the text which can be coded as such (NN - noun), (IN- preposition) [3]. The
annotations are used to identify important areas of a text that are useful to generate a
summary [32].
To assign the annotations of words containing the phrases in a document to summarize,
there is a technique known as Named Entity Extraction, which involves supervised
training of a statistical model, or more direct methods such as a dictionary or regular
expressions to classify texts or phrases of a document in a category. However,due to the
lack of training data to perform the process of finding the entities or categories in the text,
we will use a dictionary of specific domain and the technique used will be Chunking
based on Dictionary, which aims to find adjacent words in a sentence that make sense
being together. One example is ”diabetes mellitus type I”.
This phase will be based on the implementation of the matching text strings Aho-Corasick
algorithm, which consists
in finding all the
alternatives against a dictionary
independently of the number of matches or the size of the dictionary. This algorithm was
invented by Alfred V. Aho and Margaret J.Corasick, being a kind of dictionary-search
algorithm that locates elements of a finite set of strings (dictionary) in a text input [37].
The output of this phase is each phrase, together with the set of annotated words of the
corresponding category, applying heuristics to eliminate phrases that have no set of
annotated words on the categories selected by the user.
3. Extraction of Candidate Phrases: This phase involves finding key phrases to form
the summary. After the preprocessing phase, which purpose was to de- bug the
information and the phase of identification of phrases were performed, a list of phrases
with words annotated to the categories selected by the user is generated to find the most
relevant and adequate sentences. If the results are higher than the expected, it may
require further filtering of information depend- ing on the size of the summary that the
user has selected. For this, the selection of phrases is performed by three filters: Weighting
by Categories, Weighting by Frequency and Filtering by Categories, Frequency and
Sequencing :
Weighting by Categories: To find the most relevant information in the texts, the
technique known as LDA (Latent Dirichlet Allocation) will be used to automatically
discover relevant topics in the phrases. LDA represents documents as mixtures of topics
containing words with certain probabilities. It calculates the probability that the phrase
belongs to the topic taking into account the total number of words that belong to a
category or topic [4].
Weighting by Frequency: To find the most important information in the texts, we look
the more particular information with particular words in texts. For each one of the
words that were annotated for each sentence, we calculate the ”TF-IDF” frequency on
the complete document.
68


TF-IDF is a well known statistic measure used to evaluate how important a word is to
a document corpus. The importance rises proportionally to the number of times a word
appears in the document but it is countered by the frequency of the word in the corpus
[12].
Filtering by Categories, Frequency and Sequencing: After the process computes the
probability of each sentence into a category, the inverse frequency of the frequent words
in each sentence, the total of annotated words for each phrase and the number of
categories to which the phrase might belong, if the maximum number of phrases
requested by the user is greater than the number of annotated sentences obtained in the
”Identification of Categories” phase, it will get the ranking of each sentence, using the
previously calculated values (probability, the inverse frequency, the number of annotated
words and the number of categories).
This ranking is used to select the best sentences from each category. It is important to
mention the application of several heuristics: 1) Only allowed to select a phrase once,
even if it belongs to more than one category, avoiding duplication of information, 2)
When there is a conflict between the values of the ranking of two phrases, it selects the
newest phrase and 3) Annotated words that are repeated in several sentences are avoided
to disallow duplication of phrases. The result of this phase is the summary of text phrases
selected as the most important ones out of the original text. See Algorithm 1.
3.2

Component for Generating Relations between Text Summaries

The second part of Great Model corresponds to the detection of semantic relations
between summaries generated with the first component described in the
Algorithm 1 Extraction of Candidate Phrases
Require: D: It is a text document, f : It’s a phrase or paragraph that belongs to a document D.
Require: p: It is a word that belongs to a sentence or paragraph, a: It is a word annotated to a category, Pf c: Value of the
probability that a sentence f belongs to a category c, applying the ” LDA technique”.
Require: c: It is a category of information, U : It is the identifier of a user.
Require: U (c) ← c ∈ U : Set of categories selected by the user U , i.e.; U (c) = {c1 , ....cn }.
Require: f (p) ← p ∈ p: Set of annotated a words p belonging to a phrase f , where f (p) = {p 1 , ....p n }.
Require: a(f r): Set of values annotated word-frequency a.
Require: f (a) ← a ∈ f : Set phrases f annotated words a belonging to the categories selected by the user U (c),
where f (a) = {a1 , ....an }.
while f ∃f (a) {While there are phrases f with annotated words a belonging to the categories selected by the user
U (c)}. do
while p∃f (p) {While there are p words a annotated belonging to a phrase f .} do
if a 3 a(f r) {If the word annotated a not exist in the set word-frequency values a(f r)} then
f r = 1 Assign frequency f r of the word annotated a a 1.
Create a value word-frequency f r the set word-frequency values p(f r).
else
f r = f r + 1 Adding 1 to the frequency of the word annotated a. Update frequency-word value f r
the set word-frequency values p(f r).
end if end while
Pf c = LDA(f (p)) Get probability l of the set of annotated words f (p) of the phrase. Add probability Pf c by
category of each sentence f .
end while

previous section. This component is based on the following principles:
adaptation, singularity and granularity which are explained below:
i)

flexibility,

Flexibility: Allows various types of visualization of relations, such as Graph and
Tree, giving the possibility to adapt the visualization to the user’s information
needs, for example in the form of graph is significant for researchers where not known
what information is related and what degree of similarity exists. As well as the form
69


of tree is significant for performing administrative tasks where required to group
related information.
ii)

Adaptation: Allows to adapt relationships to each user according to their needs
and granularities of information, taking into account the different topics that are
often important for a person in a specific domain. For example it as- sociates the
words that correspond to the categories of information selected by the user,

iii) Singularity: Displays content that does not have any relations,
particular cases within the content,

in order to find

iv) Granularity: Display varying degrees of similarity between summaries considering
the annotated words that they contain, the more annotated words they have in
common the higher their degree of similarity.
To begin the analysis it has to be defined what a Semantic Relation is: Semantic
relations between words refers to the relations of meaning between these, there are
different types of semantic relations between words but for this model the semantic
relations that are taken into account are the relations of Hyperonymy, Hyponymy,
Meronymy and Holonym.
* Hyperonymy and hyponymy: Relation ” is-a”. The most general term is the
hypernym for example, Flower, and the most specific one is the hyponym for example a
Rose.
* Meronymy and Holonym: Establishes a hierarchy ” part-of ”. The set is the holonym
for example Boat, and the part is the meronym for example Anchor.

Fig. 1. Component for GenerationText Summarie Fig. 2. Component for Generation
Rela-tions between Text Summaries

In our model the hyponym is a word that is included within the hypernym, ie
Disease (hypernym) and Diabetes (hyponym) and a document identifier (Holonym) and the
word Diabetes (Meronym).
70


The first two relations will be reflected in the visualization in the form of Graph, where
each relation is an edge that will be identified with a color that the user has defined for each
hypernym or category, and each node represented by the identifier of a document will be the
hyponym. The two following relations will be reflected in the visualization in the form of
tree, where each relationship will be a branch of the tree, and each parent node is
represented by a word meronym and each child node of the tree represented by an identifier
of a document will be the holonym. Figures 3 and 4 show an example of each of the types
of visualization of relations.
The processes comprising this component are presented in Figure 2, which were
obtained after analysis to consolidate the processes consulted in literature [1], [28],[34].
As a result the obtained processes are: Extraction of Keywords, Semantic Data
Integration, Mapping Matrix Similarity to Data Structures and Network Development
Summaries / Visualization. Each one of the component processes are described in detail:
1. Extraction of Keywords: At this stage we identify keywords in text summaries
generated by the first component of the GReAT Model, which will be associated
through a similarity measure to reflect the semantic relations between the different
information contents. In this process, the annotated words of each phrase of the text
summaries are grouped in a vector for each summary.
2. Semantic Data Integration: During this phase a conceptual representation of the
data and its relations is performed. Based on the vector of keywords obtained in the
previous step, a similarity matrix is constructed to generate a similarity network to be
displayed. To construct this matrix, relations between summaries are quantified in terms
of distance or similarity between their vector of annotated words. To calculate the
similarity, we use the technique mentioned above Edit-Distance. In the analysis of
keywords, identifying the relations between them takes into account that the presence of a
pattern or similarity implies the presence of a relation [38].
These relations will be stored in a matrix of similarity [34], where each dimension
contains the identifier of the document that was summarized and within each matrix value,
the common keywords between two pairs of identifier text summaries of a document are
stored. The end result of this phase will be the similarity matrix formed between the
identifiers of the documents that were summarized in the first component of the model.
3. Mapping Matrix Similarity: In the previous process we obtained a similarity
matrix where the relations obtained after the process of comparison between the
keywords of each summary are stored, using (dot product - Cartesian product) [25],
between each set of words and each summary. This matrix must be mapped to a data
structure that allows to develop more easily the network between text summaries. For this
mapping, the comparisons between the values stored in the matrix are used to record
properly in a data structure the nodes and edges of the network. The end result of this
phase are two data structures that will allow displaying the two forms of visualization that
the Model GReAT offers Graph and Tree.
71


4. Network Development Summaries / Visualization (Graph and Tree): For the
process in which the network graphic summaries are developed, we found in the literature
several points that are taken into account for this process: [29],[38]. On consolidation,
the tasks that follow this model for network summaries are:
1) Collect data to visualize. 2) Analysis and mapping of the data to be visual- ized. 3)
Deployment of the visualization interface.
In detail, at this process a feature extraction and key words indexing are used to construct a
graphical representation of the collection of documents and to enable the user to quickly
identify the main topics or concepts by its importance in rendering. With the graphical
representation as a graph and tree, it is easy to discover the location of specific documents
and to explore the semantic relations that exist in a large collection of documents and
representing semantic features in order to visualize their meaning. The types of
visualization and a description of their meaning are described below:
1) Type of Visualization - Tree: In the tree view Holonym and Meronymy relations
between words annotated from the categories the user wants to see, are depicted, where
the holonym is an identifier of a document or medical record according to the case
study (see Figure 5) such HC1715059, and the meronym is a word that is part of the
document or medical record, such as WARFARIN (see Figure 4). This type of
visualization was defined to address the needs of users’ groups as administrative,
those requiring to identify comparable groups. For example, detecting comparable
groups exposed to risk factors.

2) Type of Visualization - Graph: The graph view in the Figure 3 de- picts
Hyperonymy and Hyponymy relations between words annotated from the categories
that the user desires to consult, the edge of the graph represents the category which is
the hypernym, for example, MEDICINES, and the hyponym is the node representing
the word that is associated to the category such as WARFARIN. This type of
visualization was defined to meet the needs of users’ groups as researchers, who
require to search and to analyze related information beforehand unknown and the
degree of similarity or disparity between this information. For example, links of
interest or detecting rare cases.

72


4 Prototype
To validate the proposed Model GReAT, it was implemented on a software called
”Great System” . This software was applied to a case study to summarize
information from unstructured electronic medical records in a Colombian Hos- pital.
This chapter initially presents how the GReAT System was implemented and later, its
application in the case study. The prototype was implemented in Java, using the libraries
for Natural Language Processing Lingpipe and JOrtho, as a result of the analysis of the
tools and techniques for the GReAT Model like the most appropriate and complete tools
to implement GReAT Model. The architecture was defined using different views, models
and diagrams to document and support the decisions taken during the architectural
definition of the system. The architecture is multi-layered, with three main layers:
Presentation, Business and Persistence. It uses the frameworks iBatis for data access in
the Persistence layer, Spring for business logic in the Business layer and Swing for the
presentation layer. The main screen of GReAT System can be seen in Figure
6, along with its filter options and menu options:
For the case study there were implemented these filter options: initial date, end date,
categories, total of phrases by category, color, keywords and

Fig. 6. Screens

the types of visualization (Tree and Graph). The prototype implementation allows the
performing by the user of the following functions: 1. Store categories of required
information. 2. Store the knowledge base of the required categories.
3. Store the keywords of user’s profile. 4. Enter data filter to generate text summaries on
unstructured text of electronic medical records. 5. Displaying results of generated
summaries and finally, 6. Build relationships between the results obtained in the form of
Graph or Tree. These options allow the user to perform a search according to his needs and
profile, some of them were created considering the case study of a Hospital in Colombia.

73


5

Application on Case Study

The selected case study for validation of the proposed model is presented in a Hospital in
Colombia, which stores information about the diagnosis, treatment and monitoring of
patients through a narrative text that is found in the patient’s medical records. With the
case study several tests of functionality, usability and performance were conducted.
5.1

Functional tests

In order to evaluate the generated summaries, we wanted to determine how much loss of
information we get with the generation of the summary. Obviously some information to
generate summaries should be removed, so what was intended was to measure what
relevant or irrelevant information was lost because of the summaries. What we want is
that irrelevant information gets omitted, while the relevant information remains, so
what we did was to compare the original document versus the summary, what we won
and what was lost, which was measured in terms of the relevant phrases and sentences
retrieved using precision and recall measures. To do this, we defined the relevance of a
sentence taking into account the hypothesis where information retrieval systems must be
measured indicating the degree of similarity between the question and the answer, in our
case the question will be the search filter, the knowledge base and user profile defined by
the user and the answer will be sentences in the summary, and for that you should
identify any measurable property group to estimate how many of these properties are shared
between the two entities. Measurable properties were defined: Does it have words associated
with the knowledge base defined?, Does it have Keywords associated with search filter? or
Does it have words defined in the User Profile?, what defines that a sentence is relevant or
not, is that at least the phrase comply with any of these properties. In addition to this,
we apply a formula to measure the loss of information. Relevance function is defined as
follows:
1. Representation of sentences: a phrase is considered to be characterized as a set of
one or more topics that represent its content. Following the model popularized by
Salton and McGill [35], if we have a set tcp of terms or words, such that, for example, tcp
= (Diseases, Drugs, Symptoms, T reatments) represents the topics of the sentences,
you can generalize and transform the same expression: T cp = (t1, t2, t3, t4, t5)(1), where
t1, t2, etc.., symbolize a category c or word p belonging to a sentence.
Since this represents a phrase as a vector which adopts for example, as follows: F i =< 1,
1, 1, 0, 0 > which means that phrase i contains the words or categories t1, t2 and t3 of
set tcp, where these are the words or categories that contains a phrase. Each sentence
will result in a vector with a different configuration.
2. Representation of Relevant Entities: defined Relevant Entities are also
represented as a vector. For example: Er =< 0, 1, 0, 1, 1 >, means that a particular
information need has been indexed with the words or categories t2, t4 and t5,
forming relevant entities Er.

74


Now you can see what phrases are more like the features selected by the user or
profile, setting a threshold somewhere below which a sentence would be considered no
longer relevant. For the case study, the established threshold was 1, which means that
each sentence should at least have one relevant entity to be relevant.
3. Calculation: For example, for simplicity we have only two sentences (Fi, Fj), which
present the following vectors of words or categories belonging to each phrase F i =< 1,
1, 1, 0, 0 >, F j =< 1, 1, 0, 0, 1 >.
The selection of the more relevant phrase to the user is performed by cal- culating
which one of the two sentences has more in common with the relevant entities. There
are several ways to perform this calculation (Salton and McGill [35]). One of the
simplest is the sum of products, the two numbers in each column are multiplied
together, and the results are summed, as shown in Table 3. The representation of a
sentence F i is a vector of length n: F i =< ti1, ti2, ...tin > (2a), the representation of
a word or category or relevant entity Er is a vector of length n: Er =< er1, er2, ...ern
> (2b) and the equation that allows the calculation of the degree of similarity
between a sentence F i and words or categories or relevant entities Er is formalized as
P
follows: SI M (F i, Er) = n ti1 ∗ er1
(3), which reads: the similarity (SIM) between the phrase F i and Er relevant
entity is equal to the sum of the products of each pair of respective vector elements.
For example: these are the vectors of a sentence and the relevant entities, respectively:
F i =< 1, 1, 0, 1, 0 >, Er =< 0, 1, 0, 1, 0 >, Then, the similarity is:
SI M (F i, Er) = (1 ∗ 0) + (1 ∗ 1) + (0 ∗ 0) + (1 ∗ 1) + (0 ∗ 0) = 0 + 1 + 0 + 1 + 0 = 2.
Table 3. Example - two sentences
Er = 0, 1, 0, 1, 1
Er = 0, 1, 0, 1, 1
F i = 1, 1, 1, 0, 0
F j = 1, 1, 0, 0, 1
0 +1 +0 +0 +0 = 1 0 +1 +0 +0 +1 = 2

According to this example, the phrase F j is the most similar to the relevant entities
defined by the user. Then, to calculate the relevance of each phrase, we proceed to perform
the calculations of the indices of Precision and Recall to finally get the value of
information loss. The formulas of these indices are shown below:
P recision =

#{phrasesrelevant∩retrievedphrases}

Recall

#{phrasesrelevant∩retrievedphrases}
#phrasesrelevanttrue

The formula for the loss of information is applied with the above results as follows: I L =
T otalrelevantphrases − T otalphrasesretrieved

After running 14 different test cases, the average yielded a precision high value (1.0)
where it is kept constant because the calculations to obtain the weighting of each
sentence to select the most relevant phrases matching function proposed relevance except
in some calculations. And the average yielded a high value on the completeness (0.81)
being dependent on the number of phrases that the user selected to generate the
75


summary, as this number is greater, the completeness increases handing all relevant
phrases. To assess the functionality of generating semantic relations between text
summaries, we successfully executed various test cases.
5.2

Usability tests

Usability tests were divided into two phases: An initial phase where System Great was
introduced to a set of doctors with different specialties (cardiology, pulmonary specialist,
epidemiologist, internist) and with administrative staff to validate the applicability of the
model and receive preliminary comments, from where we obtained several comments. In
the second phase, the comments were implemented and the same procedure was
performed with internists and with the administrative director of Epidemiology and
Demography in the Ministry of Health and Social Protection.
In both phases the instrument or EUCS approach was applied which was designed to
measure the satisfaction of users who directly interact with a specific application called
EUCS (End-User Computing Satisfaction Construct). Response options are in a
range of 1-5, where the value of 1 means the user disagrees with the question, and the
value of 5 means the user is more in line with the question. The results in the first
phase with the mean values were not satisfactory (3.74 / 5) because various comments
were expressed such as: include keywords for each category and focus the summary as a
preview of the information. The results in the second phase were satisfactory (4.54 / 5).
5.3

Performance tests

In order to validate the performance of the application and describe its behavior in terms
of processing speed in the process of generating text summaries several test cases were
executed with historical records of the case study. The results were satisfactory as they
were achieved in less processing time than DragonToolKit free tool as shown in Figure 7
reflecting the total medical records processed with both tools versus processing time
resulting in milliseconds. Why compared to the tool DragonToolKit? The analysis of the
knowledge base was summarized, and we found that it was the most complete tool for
generating summaries, however, there were some drawbacks such as: to fulfill the
maximum size defined for the summary, it trimmed phrases. Also to run the summary,
you must manually create a file for each document to be summarized, so only the
comparison between the two tools was made up to 65 records and Great System analysis
is made up with 600 (equivalent to 1 year medical records).

Fig. 7. Time in milliseconds versus Total medical records
76


6

Conclusions

From the review of the literature about the solutions that currently exist for the
automated generation of text summaries and generation of semantic relations, this model
seeks to address some weaknesses and challenges. The initial results show an improvement
in quality and consistency of obtained summaries, taking into account the user needs
and topics that are often important within the domain, the sequentiality of the
original text, the duplication of information and the noise that natural language
presents. On the other hand, we see the importance to relate summaries of text
obtained to get the information more complete and global for the user. As future work
this model can be reused in their various processes used as pre-processing process,
serving as the basis for other text mining projects and the validity and usefulness of the
proposed model can be extended further considering other semantic relations such as
antonyms and synonyms.

References
[1]. A text-mining-based patent network: Analytical tool for high-technology trend. The Journal of
High Technology Management Research, 15(1):37 – 50, 2004.
[2]. Amjad Abu-Jbara and Dragomir Radev. Coherent citation-based summarization of scientific
papers. In Proceedings of the 49th Annual Meeting of the Association for Computational
Linguistics: Human Language Technologies - Volume 1, HLT ’11, pages 500–509, Stroudsburg,
PA, USA, 2011. Association for Computational Linguistics.
[3]. D. B.Tsujii J. Ananiadou, S.Kell. Text mining and its potential applications in systems biology.
Trends in Biotechnology, 2006.
[4]. Rachit Arora and Balaraman Ravindran. Latent dirichlet allocation based multi-document
summarization. In Proceedings of the second workshop on Analytics for noisy unstructured text
data, AND ’08, pages 91–97, New York, NY, USA, 2008. ACM.
[5]. Aurélien Bossard, Michel Généreux, and Thierry Poibeau. Cbseas, a summarization system
integration of opinion mining techniques to summarize blogs. In Proceedings of the 12th
Conference of the European Chapter of the Association for Computational Linguistics:
Demonstrations Session, EACL ’09, pages 5–8, Stroudsburg, PA, USA, 2009. Association for
Computational Linguistics.
[6]. Senso Jose A. Brun Ricardo Eto. Miner´ıa textual. IEEE, 2004.
[7]. Alonso Martinez Margarita Cobo Ortega Angel, Rocha Blanco Rocio. Descubrim- iento de
conocimiento en repositorios documentales mediante técnicas de miner´ıa de texto y swarm
intelligence. In Rect@ Vol 10 Diciembre 2009.
[8]. Hal Daumé, III and Daniel Marcu. A noisy-channel model for document compression. In
Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, ACL ’02,
pages 449–456, Stroudsburg, PA, USA, 2002. Association for Computational Linguistics.
[9]. C.L. Devasena. Automatic text categorization and summarization using rule reduction. In
Advances in Engineering, Science and Management (ICAESM), 2012 International Conference on,
pages 594–598, 2012.
[10]. Pierre-Etienne Genest and Guy Lapalme. Framework for abstractive summarization using text-totext generation. In Proceedings of the Workshop on Monolingual Text-To-Text Generation, pages
64–73, Portland, Oregon, June 2011. Association for Computational Linguistics.
[11]. Dragomir R. Radev Gunen, Erkan. Lexrank: Graph-based lexical centrality as salience in text
summarization. Journal of Artificial Intelligence Research 22 (2004) 457-479, 22, 2004.
[12]. Jung-Hsien Chiang Heng-Hui Liu, Yi-Ting Huang. A study on paragraph ranking and
recommendation by topic information retrieval from biomedical literature. IEEE, 2010.
[13]. Paass G. Hotho A., Nurnberger A. A brief survey of text mining. IEEE, 2005.
77


[14]. Tasha R. Inniss, John R. Lee, Marc Light, Michael A. Grassi, George Thomas, and Andrew B.
Williams. Towards applying text mining and natural language processing for biomedical
ontology acquisition. In Proceedings of the 1st international workshop on Text mining in
bioinformatics, TMBIO ’06, pages 7–14, New York, NY, USA, 2006. ACM.
[15]. Theo j.d. Bothma Jan h. Kroeze, Machdel c. Matthee. Differentiating data- and text-mining
terminology. ACM, 2003.
[16]. Keivan Kianmehr, Shang Gao, Jawad Attari, M. Mushfiqur Rahman, Kofi Akomeah, Reda
Alhajj, Jon Rokne, and Ken Barker. Text summarization techniques: Svm versus neural
networks. In Proceedings of the 11th International Con- ference on Information Integration and
Web-based Applications y Services, iiWAS ’09, pages 487–491, New York, NY, USA, 2009.
ACM.
[17]. A.Kongthon, C. Haruechaiyasak, and S. Thaiprayoon. Constructing term the- saurus using text
association rule mining. In Electrical Engineering/Electronics, Computer, Telecommunications
and Information Technology, 2008. ECTI-CON 2008. 5th International Conference on, volume 1,
pages 137–140, 2008.
[18]. Dr Richi Nayak Lin Chen. A case study of failure mode analysis with text mining methods. IEEE,
2007.
[19]. Chong Long, Min-Lie Huang, Xiao-Yan Zhu, and Ming Li. A new approach for multi-document
update summarization. J. Comput. Sci. Technol., 25(4):739–749, July 2010.
[20]. M A Zaveri M K Dalal. Heuristics based automatic text summarization of unstructured text.
ACM, 2011.
[21]. Dr. B. Ravindran M. Saravanan, Dr. S .Raman. A probabilistic approach to multi- document
summarization for generating a tiled summary. IEEE, 2005.
[22]. Antnio Horta Branco Marcus V. C. Guelpeli Ana Cristina B. Garcia. The process of summarization
in the pre-processing stage in order to improve measurement of texts), journal = IEEE, year =
2011.
[23]. Yashar Mehdad, Matteo Negri, Elena Cabrio, Milen Kouylekov, and Bernardo Magnini. EDITS:
An Open Source Framework for Recognizing Textual Entailment.
[24]. Saif Mohammad, Bonnie Dorr, Melissa Egan, Ahmed Hassan, Pradeep Muthukr- ishan, Vahed
Qazvinian, Dragomir Radev, and David Zajic. Using citations to generate surveys of scientific
paradigms. In Proceedings of Human Language Tech- nologies: The 2009 Annual Conference of
the North American Chapter of the As- sociation for Computational Linguistics, NAACL ’09, pages
584–592, Stroudsburg, PA, USA, 2009. Association for Computational Linguistics.
[25]. Pradeep Muthukrishnan, Dragomir Radev, and Qiaozhu Mei. Simultaneous similarity learning
and feature-weight learning for document clustering.
In Pro- ceedings of TextGraphs-6:
Graph-based Methods for Natural Language Processing, TextGraphs-6, pages 42–50, Stroudsburg,
PA, USA, 2011. Association for Compu- tational Linguistics.
[26]. S. Nadschlager, H. Kosorus, A. Bogl, and J. Kung. Content-based recommen- dations within a
qa system using the hierarchical structure of a domain-specific taxonomy. In Database and Expert
Systems Applications (DEXA), 2012 23rd In- ternational Workshop on, pages 88–92, 2012.
[27]. S. Nadschlager, H. Kosorus, P. Regner, and J. Kung. Semantic data integration and relationship
identification using the hierarchical structure of a domain-specific taxonomy. In Database and
Expert Systems Applications (DEXA), 2012 23rd In- ternational Workshop on, pages 48–52,
2012.
[28]. Na Ni, Kai Liu, and YaoDong Li. An automatic multi-domain thesauri construction method based
on lda. In Machine Learning and Applications and Workshops (ICMLA), 2011 10th International
Conference on, volume 2, pages 235–240, 2011.
[29]. Zhou Ning, Wu Jiaxin, Wang Bing, and Zhang Shaolong. A visualization model for information
resources management. In Information Visualisation, 2008. IV ’08. 12th International Conference,
pages 57–62, 2008.
[30]. Fukuhara Tomohiro Park Jaehui. Web content summarization using social book marks: A new
approach for social summarization. ACM, 2008.
78


[31]. Aurora Pons-Porrata, Rafael Berlanga-Llavori, and Jose Ruiz-Shulcloper. Topic discovery based
on text mining techniques. Information Processing y Management, 43(3):752 – 768, 2007. Special
Issue on Heterogeneous and Distributed IR.
[32]. Lawrence H. Reeve, Hyoil Han, Saya V. Nagori, Jonathan C. Yang, Tamara A. Schwimmer,
and Ari D. Brooks.
Concept frequency distribution in biomedical text summarization. In
Proceedings of the 15th ACM international conference on Information and knowledge
management, CIKM ’06, pages 604–611, New York, NY, USA, 2006. ACM.
[33]. James Sanger Ronen Feldman. The text mining handbook. Cambridge University Press, 2
edition, 2007.
[34]. Zhao J.Leon Roussinov Dmitri. Automatic discovery of similarity relationships through web
mining. Elsevier Sciense, 2002.
[35]. Gerard Salton and Michael J. McGill. Introduction to Modern Information Retrieval. McGrawHill, Inc., New York, NY, USA, 1986.
[36]. Wen-Feng Hsiao Te-Min Chang. A hybrid approach to automatic text summarization. IEEE, 2008.
[37]. Nhat-Phuong Tran, Myungho Lee, Sugwon Hong, and Minho Shin.
Memory efficient
parallelization for aho-corasick algorithm on a gpu.
In High Performance Computing and
Communication 2012 IEEE 9th International Conference on Em- bedded Software and Systems
(HPCC-ICESS), 2012 IEEE 14th International Con- ference on, pages 432–438, 2012.
[38]. Gurpreet S. Lehal Vishal Gupta. A survey of text mining techniques and applications.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLI- GENCE, 2009.
[39]. Wei Wang, Chuan Xiao, Xuemin Lin, and Chengqi Zhang. Efficient approximate entity
extraction with edit distance constraints. In Proceedings of the 2009 ACM SIGMOD International
Conference on Management of data, SIGMOD ’09, pages 759–770, New York, NY, USA, 2009.
ACM.
[40]. ChengXiang Zhai Bruce Schatz Xu Ling, Qiaozhu Mei. Mining multi-faceted overviews of
arbitrary topics in a text collection. ACM, 2008.
[41]. Jiaming Zhan, Han Tong Loh, and Ying Liu. Gather customer concerns from online product reviews
- a text summarization approach. Expert Syst. Appl., 36(2):2107– 2115, March 2009.
[42]. LI Cun-he Zhang Pei-ying. Automatic text summarization based on sentences clustering and
extraction. IEEE, 2009.
[43]. Xiaohua Zhou, Xiaodan Zhang, and Xiaohua Hu. Dragon toolkit: Incorporating auto-learned
semantic knowledge into large-scale text retrieval and mining. In Tools with Artificial Intelligence,
2007. ICTAI 2007. 19th IEEE International Conference on, volume 2, pages 197–201, 2007.

79

Great model a model for the automatic generation of semantic relations between text summaries based on the relevant information to the user

Recommended

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to Great model a model for the automatic generation of semantic relations between text summaries based on the relevant information to the user (20)

More from ijcsity (20)

Recently uploaded (20)

Great model a model for the automatic generation of semantic relations between text summaries based on the relevant information to the user