Introduction
Since the internet is so widely used, many people can share casual remarks and technical
information about a wide range of issues. Research in a number of natural language processing
(NLP)-related fields is made possible by this, including the importance of providing a concise
overview of the vast amount of information that is available and gleaning valuable information
from it. The primary goals of argument mining, a well-known area of study within natural
language processing, are the identification, extraction, and analysis of argumentative claims. An
argument is a person's particular viewpoint or opinion about something he believes in. For an
argument to be legitimate and a statement to be accepted, it must be backed up by pertinent facts.
Arguments are typically found in scientific studies, debate scripts, argumentative essays, and
user comments in blogs and articles, among many other places.[1]. Although there are many
different types of argumentation structures in the literature, most of them share the presence of a
claim and a justification [5]. Claims are seen as a position, assessment, or conclusion in this
situation. Furthermore, justification can be found in the literature as propositions, premises,
arguments, or proof [5]. Four categories of assertions can generally serve as the foundation for
argument components: main claims, claims, premise, and nonargumentative statements. [E].
Argument mining is a relatively new field of study that requires a combination of linguistic
analysis and artificial intelligence (AI)-based language processing due to its many challenges.
[3].
Even for people, it is difficult to identify arguments because a claim cannot be defined by
established standards. Claims are sometimes given as statements pertaining to pertinent concepts,
and other times they are given as opinions backed up by different facts. The fact that not all
arguments are clearly stated, with premises and claims clearly stated, is another factor
contributing to the challenge. Arguments can occasionally be difficult to distinguish because they
are suggested or concealed within a larger context. The implicit character of the premises allows
one to identify the argument's constituent parts. Although the assertion is clear, the premises or
supporting arguments are not, and revealing them calls for close examination and contextual
knowledge.
Understanding and processing natural language requires further study and advancement in this
area, particularly when it comes to the direct application of and connections with other fields.
With the growing application of scientific data that has demonstrated great promise in healthcare
and model-informed drug development, the medical sector is among the most significant [2].
Drug review websites that offer a multitude of data about the experiences of users of various
products are the subject of a recent development in the field of natural language processing. With
the introduction of new medications to the market, some people could encounter unanticipated
adverse drug reactions. Although there are comprehensive rules for the annotation process, the
medical terminologies are unclear, which leads to inconsistent annotation.
Incomplete medical reviews can occasionally make it difficult to understand the statement
structure, and the system needs domain specialists to make the problems more complex [10].
Accuracy gradually declines as a result of human annotators' fatigue from working on several
medical books [11]. For extensive or difficult assessments, the quality of annotations might be
affected, frequently in a negative way. The development of automatic argument annotation
,methods will make it easier to predict public reaction to medications and to administer them
prudently [4]. Despite the tremendous advancements in the field of argument mining, there is
currently a lack of modern modeling and analytical methods for annotating scientific data [4].
Consequently, the advanced deep learning model employs an effective design to address difficult
problems. [12]. This paper's primary contributions are (1) introducing a novel hybrid strategy
that uses an enhanced QRNN-GRU model to appropriately annotate argument structures. As far
as we are aware, this study is the first to use QRNN for argument mining, showing how its
special architecture can successfully handle the intricate dependencies and subtleties present in
argumentative texts; (2) strengthen argument annotation in intricate datasets with a variety of
complex and nuanced argument forms, like those found in medical drug reviews; and (3)
improve model performance through optimization using the Firefly Algorithm, carefully
adjusting parameters to maximize accuracy and efficiency, thus establishing a new standard in
the field of argument annotation.
To comprehend the procedure, the remainder of the paper is pre-framed. The many researchers'
activities and insights about the classification of argument annotations are described in Section 2. The
operation of the QRNN-GRU-based annotation categorization process and the system's
experimentation, as established in Section 4, are covered in Section 3. Section 5 provides a
description of the work's overall conclusions.
2. Related Works
Researchers' attention has been drawn to the rapidly expanding field of argument mining in
recent years [9]. Argument border identification and extraction, component type prediction (e.g.,
premise and conclusion), or relation categorization as distinct tasks have been the primary focus
of current argument mining research. The pipeline's initial step involves identifying and
extracting arguments, which is a somewhat costly and complex procedure that typically calls for
a lot of effort from human annotators in order to achieve a decent level of inter-rater agreement
[35]. Argumentative components have been manually annotated in numerous research, which
reduces scalability [6], [33].
Human performance and automated processes are contrasted by Stab and Gurevych [28]. As two
distinct classification problems on persuasive essays, the study looked at the identification of
argumentative material (main claim, claim, premise) and the extraction of the relations. They
used support vector machines (SVM), random forests (RF), naïve bayes (NB), and C4.5 decision
trees for argument annotation. Though they did not surpass the accuracy of the human annotator,
the support vector machine (SVM) was the only model that marginally beat the majority of
baseline models because to the class imbalance, reaching 77%.
The goal of Huwaidah and Al Faraby is to determine if a tweet contains an argument or not by
categorizing its content as argument, non-argument, or unknown. They employed SVM,
Multinomial NB, and TF-IDF for feature extraction in this experiment. Three human annotators
annotated the dataset. With an accuracy of 71.42%, SVM utilizing only a unigram feature
yielded the best results. This study demonstrates that the efficiency of the stopwords feature
varies according to the feature combination used in the model. Because the model has not been
able to solve the issue of variance of abbreviations, slang dictionaries, and non-standard word
, equivalents in the dataset, the data prediction results are not optimal. Additionally, one of the
reasons the classification effort has not been at its best is the small amount of data. The
misclassified result was the result of a heavy and unequal distribution between classes.
A study by [Muram] aimed to develop a system, called ATTEST, for automating the assessment
and updating of assurance case arguments. The study used a rule-based methodology, developing
rules based on the sequential dependencies and hierarchical concept tiers of the metamodels.
Regular expressions were used in the development of these rules, which helped identify the
target data for extraction and resolve conflicts between them. These rules were used for defeater
identification, well-formedness, sufficiency tests, argument comprehension, and counterevidence
selection. Overall, the study demonstrated how the ATTEST framework can identify and manage
assertions and supporting information in assurance case arguments, providing useful support for
assurance case updates and reviews. However, a human effort that is essential to the framework's
operation is identifying mappings between process, product, and assurance case metamodels.
Additionally, as part of a broader "hermeneutics of persuasive discourse," a study offers a theory
for interpreting different forms of arguments. This hermeneutics' growth is grounded in
pragmatic discoveries, particularly those found in the philosophy of argument. With an emphasis
on distinguishing argument types using the Periodic Table of Arguments (PTA) as a framework,
the study describes the primary procedures required in analyzing persuasive language. Argument
form, argument substance, and argument lever are the three criteria that the PTA uses to classify
arguments. The author called this approach "rule-based hermeneutics" since it can be explained
in a decision tree for determining the argument form. Identifying the type of argument is
preceded by the stages of argument mapping and detection, and the evaluation of the argument
follows. It was difficult for this study to give a more specific assessment of the link between the
argument's components.
Additionally, Suhartono et al. [1] used a number of deep learning networks, including
convolutional neural networks (CNN), GRU, and LSTM, to examine the arguments and integrate
them with the attention mechanism. This was used on 402 persuasive writings with English
metadata that were translated into Indonesian. The arguments were divided into several groups
by the deep learning models, including argumentized, nonargumentized, preclaim, and claim.
The scholars looked at the arguments to determine whether or not an established argument was
considered sufficient. The study found that the use of the attention mechanism increased the
overall classification process rate, with CNN-attention achieving the best result for argument
annotation at 76.56%.
Additional research has been conducted on the application of fully automated techniques in
argument annotation. Addawood and Bashir [13] developed a theoretical framework in their
study that can recognize the unique arguments found in social media text. A number of
contemporary techniques, including naive Bayes, SVM, and decision trees, were categorized and
contrasted by the authors. With a 0.89 F1 measure, SVM outperformed the other algorithms in
the argument identification challenge, yielding the best results. Additionally, Qing-Chuan et al.
(2023) [24] presented the Modeling Graph Neural Networks (MGNN) technique for document
argument extraction. This work's primary goal is to solve span sentence problems in order to
extract semantic interactions and arguments from the documents. The graph model examines
every object, sentence, and document to adapt the procedure in a dynamic setting and generate
Since the internet is so widely used, many people can share casual remarks and technical
information about a wide range of issues. Research in a number of natural language processing
(NLP)-related fields is made possible by this, including the importance of providing a concise
overview of the vast amount of information that is available and gleaning valuable information
from it. The primary goals of argument mining, a well-known area of study within natural
language processing, are the identification, extraction, and analysis of argumentative claims. An
argument is a person's particular viewpoint or opinion about something he believes in. For an
argument to be legitimate and a statement to be accepted, it must be backed up by pertinent facts.
Arguments are typically found in scientific studies, debate scripts, argumentative essays, and
user comments in blogs and articles, among many other places.[1]. Although there are many
different types of argumentation structures in the literature, most of them share the presence of a
claim and a justification [5]. Claims are seen as a position, assessment, or conclusion in this
situation. Furthermore, justification can be found in the literature as propositions, premises,
arguments, or proof [5]. Four categories of assertions can generally serve as the foundation for
argument components: main claims, claims, premise, and nonargumentative statements. [E].
Argument mining is a relatively new field of study that requires a combination of linguistic
analysis and artificial intelligence (AI)-based language processing due to its many challenges.
[3].
Even for people, it is difficult to identify arguments because a claim cannot be defined by
established standards. Claims are sometimes given as statements pertaining to pertinent concepts,
and other times they are given as opinions backed up by different facts. The fact that not all
arguments are clearly stated, with premises and claims clearly stated, is another factor
contributing to the challenge. Arguments can occasionally be difficult to distinguish because they
are suggested or concealed within a larger context. The implicit character of the premises allows
one to identify the argument's constituent parts. Although the assertion is clear, the premises or
supporting arguments are not, and revealing them calls for close examination and contextual
knowledge.
Understanding and processing natural language requires further study and advancement in this
area, particularly when it comes to the direct application of and connections with other fields.
With the growing application of scientific data that has demonstrated great promise in healthcare
and model-informed drug development, the medical sector is among the most significant [2].
Drug review websites that offer a multitude of data about the experiences of users of various
products are the subject of a recent development in the field of natural language processing. With
the introduction of new medications to the market, some people could encounter unanticipated
adverse drug reactions. Although there are comprehensive rules for the annotation process, the
medical terminologies are unclear, which leads to inconsistent annotation.
Incomplete medical reviews can occasionally make it difficult to understand the statement
structure, and the system needs domain specialists to make the problems more complex [10].
Accuracy gradually declines as a result of human annotators' fatigue from working on several
medical books [11]. For extensive or difficult assessments, the quality of annotations might be
affected, frequently in a negative way. The development of automatic argument annotation
,methods will make it easier to predict public reaction to medications and to administer them
prudently [4]. Despite the tremendous advancements in the field of argument mining, there is
currently a lack of modern modeling and analytical methods for annotating scientific data [4].
Consequently, the advanced deep learning model employs an effective design to address difficult
problems. [12]. This paper's primary contributions are (1) introducing a novel hybrid strategy
that uses an enhanced QRNN-GRU model to appropriately annotate argument structures. As far
as we are aware, this study is the first to use QRNN for argument mining, showing how its
special architecture can successfully handle the intricate dependencies and subtleties present in
argumentative texts; (2) strengthen argument annotation in intricate datasets with a variety of
complex and nuanced argument forms, like those found in medical drug reviews; and (3)
improve model performance through optimization using the Firefly Algorithm, carefully
adjusting parameters to maximize accuracy and efficiency, thus establishing a new standard in
the field of argument annotation.
To comprehend the procedure, the remainder of the paper is pre-framed. The many researchers'
activities and insights about the classification of argument annotations are described in Section 2. The
operation of the QRNN-GRU-based annotation categorization process and the system's
experimentation, as established in Section 4, are covered in Section 3. Section 5 provides a
description of the work's overall conclusions.
2. Related Works
Researchers' attention has been drawn to the rapidly expanding field of argument mining in
recent years [9]. Argument border identification and extraction, component type prediction (e.g.,
premise and conclusion), or relation categorization as distinct tasks have been the primary focus
of current argument mining research. The pipeline's initial step involves identifying and
extracting arguments, which is a somewhat costly and complex procedure that typically calls for
a lot of effort from human annotators in order to achieve a decent level of inter-rater agreement
[35]. Argumentative components have been manually annotated in numerous research, which
reduces scalability [6], [33].
Human performance and automated processes are contrasted by Stab and Gurevych [28]. As two
distinct classification problems on persuasive essays, the study looked at the identification of
argumentative material (main claim, claim, premise) and the extraction of the relations. They
used support vector machines (SVM), random forests (RF), naïve bayes (NB), and C4.5 decision
trees for argument annotation. Though they did not surpass the accuracy of the human annotator,
the support vector machine (SVM) was the only model that marginally beat the majority of
baseline models because to the class imbalance, reaching 77%.
The goal of Huwaidah and Al Faraby is to determine if a tweet contains an argument or not by
categorizing its content as argument, non-argument, or unknown. They employed SVM,
Multinomial NB, and TF-IDF for feature extraction in this experiment. Three human annotators
annotated the dataset. With an accuracy of 71.42%, SVM utilizing only a unigram feature
yielded the best results. This study demonstrates that the efficiency of the stopwords feature
varies according to the feature combination used in the model. Because the model has not been
able to solve the issue of variance of abbreviations, slang dictionaries, and non-standard word
, equivalents in the dataset, the data prediction results are not optimal. Additionally, one of the
reasons the classification effort has not been at its best is the small amount of data. The
misclassified result was the result of a heavy and unequal distribution between classes.
A study by [Muram] aimed to develop a system, called ATTEST, for automating the assessment
and updating of assurance case arguments. The study used a rule-based methodology, developing
rules based on the sequential dependencies and hierarchical concept tiers of the metamodels.
Regular expressions were used in the development of these rules, which helped identify the
target data for extraction and resolve conflicts between them. These rules were used for defeater
identification, well-formedness, sufficiency tests, argument comprehension, and counterevidence
selection. Overall, the study demonstrated how the ATTEST framework can identify and manage
assertions and supporting information in assurance case arguments, providing useful support for
assurance case updates and reviews. However, a human effort that is essential to the framework's
operation is identifying mappings between process, product, and assurance case metamodels.
Additionally, as part of a broader "hermeneutics of persuasive discourse," a study offers a theory
for interpreting different forms of arguments. This hermeneutics' growth is grounded in
pragmatic discoveries, particularly those found in the philosophy of argument. With an emphasis
on distinguishing argument types using the Periodic Table of Arguments (PTA) as a framework,
the study describes the primary procedures required in analyzing persuasive language. Argument
form, argument substance, and argument lever are the three criteria that the PTA uses to classify
arguments. The author called this approach "rule-based hermeneutics" since it can be explained
in a decision tree for determining the argument form. Identifying the type of argument is
preceded by the stages of argument mapping and detection, and the evaluation of the argument
follows. It was difficult for this study to give a more specific assessment of the link between the
argument's components.
Additionally, Suhartono et al. [1] used a number of deep learning networks, including
convolutional neural networks (CNN), GRU, and LSTM, to examine the arguments and integrate
them with the attention mechanism. This was used on 402 persuasive writings with English
metadata that were translated into Indonesian. The arguments were divided into several groups
by the deep learning models, including argumentized, nonargumentized, preclaim, and claim.
The scholars looked at the arguments to determine whether or not an established argument was
considered sufficient. The study found that the use of the attention mechanism increased the
overall classification process rate, with CNN-attention achieving the best result for argument
annotation at 76.56%.
Additional research has been conducted on the application of fully automated techniques in
argument annotation. Addawood and Bashir [13] developed a theoretical framework in their
study that can recognize the unique arguments found in social media text. A number of
contemporary techniques, including naive Bayes, SVM, and decision trees, were categorized and
contrasted by the authors. With a 0.89 F1 measure, SVM outperformed the other algorithms in
the argument identification challenge, yielding the best results. Additionally, Qing-Chuan et al.
(2023) [24] presented the Modeling Graph Neural Networks (MGNN) technique for document
argument extraction. This work's primary goal is to solve span sentence problems in order to
extract semantic interactions and arguments from the documents. The graph model examines
every object, sentence, and document to adapt the procedure in a dynamic setting and generate
