Page 92 - IJEEE-2023-Vol19-ISSUE-1

88 |
response, many agencies have taken steps to improve our
current automated feedback system.
2) The current machine learning approach
Two institutions, GSU and the Learning Institutions
Laboratory, have investigated a machine learning-based
approach to improving automated feedback systems. They
argue that machine learning models can be trained to
accurately classify discourse elements in written works. This
model can then be added to an existing feedback system to
help the system provide better and more constructive
feedback to students. The Learning Institution Lab took the
first steps towards creating this model by creating the corpus
the Persuasive Essays for Rating, Selecting, and
Understanding Argumentative and Discourse Elements
(PERSUADE). The PERSUADE corpus is a collection of
over 25,000 argumentative papers collected from students in
grades 6 to 12 [5]. All articles are annotated by professional
English teachers in order to understand the different elements
of discourse. After creating this dataset, machine learning
researchers have the basic facts they need to start training
models. In particular, they hope to optimize existing natural
language processing (NLP) models for discourse
classification tasks, focusing on Google's Bidirectional
Encoder Representation (BERT) model from Transformers.
I agree with the current approach to fine-tuning this model
for discourse classification; however, I believe some steps
are required to make these models more accurate.
3) The Discourse Elements
This list of discourse elements has been compiled by a
team of teachers and professional writers from The Learning
Agency Lab [6]. They believed that this list contained all the
important discourse elements that make up the students'
writing, and they used this list as a template for creating the Bayesian classifier is small, with only 100 articles, and the
PERSUADE corpus. I will use the same rules when
improving my own models:
• Introduction - an introduction that begins with statistics,
citations, descriptions, or other means of grabbing the
reader's attention and pointing to the paper
• Position - opinion or conclusion on the main issue
• Statement - a statement supporting the position
• Counterclaim - an allegation that refutes another allegation
or justifies the position to the contrary
• Rebuttal - rebutting the assertion of the counterclaim
• Evidence - an opinion or example to support a claim,
counterclaim or refutation.
• Closing Statements - Closing Statements Reaffirming
Statements.
4) My approach & potential outcomes
As with current machine learning approaches, I believe
that transformation-based models such as bidirectional
encoder representations from Transformers (BERT) [7] can
be fine-tuned to successfully address discourse classification
problems. However, in this article, I also want to examine
other transformer models and compare/contrast the different
results. In addition, further improvements can be made to the
corpus of Persuasive Essays for Rating, Selecting, and
Understanding Argumentative and Discourse Elements
(PERSUADE). Current corpora divide articles into
sequences, each sequence corresponding to a different type
of utterance. However, I will restructure the dataset so that
Alkabool, Abdullah, Zadeh, & Mahfooz
the full paper can be provided to the model during training.
In this post, I hope to demonstrate that Transformer-based
models should be trained concurrently throughout the post to
take full advantage of their architecture. I also hope to
demonstrate that it is useful for machine learning researchers
to evaluate models other than BERT (Bidirectional Encoder
Representations for Transformers) models, and I hope to
demonstrate that Long Document Transformer (Longformer)
[8] -Model is better in the following cases The model comes
to discourse classification.
5) Outline
To justify my approach, I first turn to other projects
focusing on discourse classification. I then describe in more
detail my approach to the problem of utterance classification
and what I have done to implement a transformer-based
model bidirectional encoder representation slave transformer
(BERT), long document transformer (Longformer), and
generative pretrained transformer take Step-2 (GPT-2).
Afterwards, I will review some of my promising findings and
explain their implications for discourse classification tasks.
Finally, I'll cover some improvements that can be made for
future fine-tuning attempts.
6) Related Works
Researchers Burstein et al. [7] attempted to use a Bayesian
classifier to identify thesis statements in student written
work. Their model was able to achieve an average overall
accuracy of 43%, but most importantly, they were able to
show that the classification of propositional statements was
generalizable. That is, the model does not need to be
retrained for each new paper prompt, and once trained, the
model can recognize paper statements in all paper topics.
One drawback, however, is that the training set for the
authors admit that their model can hold. Another model to
mention is the Longformer model modified by programmer
Darek Kleczek [8]. Kleczek solves this problem by
optimizing a pre-existing longformer model on the hug face
website [9], which was originally trained by machine
learning engineers at allenai. By fine-tuning the Longformer
model, Kleczek was able to achieve an accuracy of 61.4
Taboada et al. [9] Enters the history of Rhetorical Structure
Theory (RST) and its advantages today. They found that RST
can be used for a variety of applications (including discourse
classification) and is a "robust and well-tested theory". Most
importantly, they found relationships between various
elements of utterances that we hope our model will capture.
Instead of trying to specifically define the relationship
between the models or create a working machine learning
model, the researchers leave it as an open problem for others
to solve.
The machine learning researcher Julian Peller [10] addresses
the problem of classifying utterance elements by improving
Google's BERT model. He approaches the problem as a
tokenized classification problem using Named Entity
Recognition (NER), where articles are lists of tokens and
utterance elements are distinct classes. He also trained on
10,000 articles from the PERSUADE corpus, and he
achieved an overall accuracy of 0.226 on the F-1 score.
Ali Habiby [11] tackled the problem of classifying utterance
elements in a rather unique way. Instead of defining the