A Comparative Study of Text Comprehension in IELTS Reading Exam using GPT-3

International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
DOI: 10.5121/ijcsit.2024.16104 45
A COMPARATIVE STUDY OF TEXT
COMPREHENSION IN IELTS READING
EXAM USING GPT-3
Christopher Le1
, Tauheed Khan Mohd2
1
Dept. of Math and Computer Science, Augustana College, IL USA
2
School of Information Security and Applied Computing, Eastern Michigan University,
Ypsilanti, MI
ABSTRACT
This paper discusses the capabilities and limitations of GPT-3 (0), a state-of-the-art language model, in the
context of text understanding. We begin by describing the architecture and training process of GPT-3, and
provide an overview of its impressive performance across a wide range of natural language processing
tasks, such as language translation, question-answering, and text completion. Throughout this research
project, a summarizing tool was also created to help us retrieve content from any types of document,
specifically IELTS (0) Reading Test data in this project. We also aimed to improve the accuracy of the
summarizing, as well as question-answering capabilities of GPT-3 (0) via long text.
KEYWORDS
GPT, Artificial Intelligence, Natural language processing.
1. INTRODUCTION
In recent years, natural language processing (NLP) has seen significant advances, thanks to the
development of large-scale language models such as BERT (Bidirectional Encoder
Representations from Transformers) from Google, RoBERTa (Robustly Optimized BERT
approach) from Meta, and GPT (Generative Pre-trained Transformer) from OpenAI. These
models have been a breakthrough in enabling machines to understand, interpret, and generate
natural languages in a way that is similar to human communication. Thus, they brought benefits
in a variety of fields, including healthcare, finance, marketing, and education.
However, the success of NLP also raises important ethical and societal concerns, including issues
related to bias, privacy, and automation of tasks that were previously the domain of human
expertise. Therefore, there is a need to explore the capabilities and limitations of GPT-3 and
critically examine its ethical implications.
In this paper, we aim to provide a comprehensive overview of GPT-3’s capabilities and
limitations in the context of text understanding. We will use the dataset of IELTS (0) Reading test
to test GPT-3’s capabilities of summarizing long text, text understanding, and finally question-
answering. In the end, a summarizing tool that we created will also be discussed to test and
improve the current model on text understanding. In order to do this, we will need to understand
the architecture and training process of GPT-3 and related models, thus figuring out its challenges
or limitations of text understanding.We will also analyze the ethical implications of such models
like GPT-3 and provide a roadmap for future research in the field of NLP.

46
The ability to summarize a lengthy text is crucial for anyone who works in the media, business, or
education since we live in a fast-paced world where we must constantly ingest vast amounts of
information. GPT-3 (Generative Pre-trained Transformer 3) is one such model that has the
potential to replace humans in this time-consuming and challenging work. Fortunately, the
development of natural language processing (NLP) models has given the ability to automate text
summarization. One of the tasks that GPT-3 has shown particular proficiency in is text
summarization. GPT-3 can summarize lengthy texts by identifying the most necessary
information while presenting it concisely and coherently.
The study will involve comparing the performance of summaries generated by GPT-3 to those
performed by human summarizers using a sample of long documents. The main target of this
research is the IELTS Reading Test, one of the four sections of the International English
Language Testing System (IELTS) exam that assesses non-native speakers’ ability to understand
and analyze written texts.
The reason IELTS is an ideal research target is that GPT-3 has been shown to perform
remarkably well on question answering and text completion tasks, which are equivalent to
multiple choice, yes/no/not given, and filling in the blank tasks in the IELTS Reading test.
Furthermore, IELTS Reading tests cover a wide range of topics, including natural science, social
studies, business, and global issues. Lastly, IELTS is a widely recognized English language
proficiency test used for several purposes, including applying to study abroad or finding
employment.
Analyzing the strengths and limitations of GPT-3 in this context can help us to gain valuable
insight into the potential implications of using NLP models like GPT-3 to automate language
proficiency testing and assessment. If GPT-3 proves to be an effective and reliable tool for
summarizing and comprehending lengthy texts, it could significantly make the collecting
information process more efficient and accessible. Moreover, this research could also inform the
development of more sophisticated language testing tools that incorporate NLP models and other
advanced technologies.
2. RELATED WORKS
In comparison to GPT-3, which is the main focus of this research, there have been other great
LLM models from Google, and Meta: BERT and Roberta respectively. It’s also important to
understand transformer and attention, the foundation behind GPT model.
2.1. Large Language Model
Large Language Model (LLM) (0) is a type of artificial intelligence that uses deep learning
techniques to analyze and understand natural language. LLMs are designed to learn from massive
amounts of text data and use that knowledge to generate human-like responses to questions or
statements. These models have been trained on vast corpus of text data to store world knowledge
within their neural network weights. LLMs have many practical applications in natural language
processing, such as machine translation, speech recognition, and text generation. Despite their
potential, there are also concerns about their ethical and social implications, including issues
related to bias, privacy, and misinformation.
Top LLMs known at this moment are: ChatGPT (OpenAI) (?), and BARD (Google), and
LLaMA(0) (Meta).

47
2.2. Transformer
In 2017, Vaswani et al. introduced the Transformer (0), a potent deep-learning model. Natural
language processing (NLP) has undergone a revolution as a result of its ability to create models to
analyze collections of inputs or outputs (text, speech, etc.). The application of the Transformer (0)
was seen in many stateof-the-art NLP tasks, including language translation, language modeling,
and question answering, with unprecedented levels of accuracy.
What sets the Transformer (0) apart from previous NLP models is its use of attention
mechanisms. The Transformer (0) model uses self-attention to enable the processing of sequences
of variable length without relying on sequential processing like traditional recurrent neural
networks. This allows the model to be much more efficient and effective at processing long
sequences of text or speech. Since then, The Transformer (0) has become a foundational model in
the field of NLP and impacts several popular deep learning frameworks, including TensorFlow
(0), PyTorch (0), and Keras(0). The Transformer (0) is undeniably an active area of research and
development because it has created a wide range of opportunities for NLP.
2.3. Attention
The attention mechanism is a powerful concept in deep learning that has gained widespread use
in various natural language processing (NLP) applications. It enables the model to focus on
important features in the input sequence while processing it, which can significantly improve the
model’s performance (0). The fundamental principle of attention mechanisms is to emphasize the
pertinent portions of the input sequence while dismissing the unimportant ones. It enables the
model to develop the ability to selectively attend to various input components, which is very
advantageous when processing lengthy sequences.
Many NLP tasks, including machine translation, text categorization, question answering, and text
summarization, have proven to perform much better when attention mechanisms are applied.
More specifically, compared to conventional models like recurrent neural networks, attention
mechanisms have made it possible for models to handle lengthy input sequences more
successfully (RNNs (0)). Attention mechanisms have evolved with different types and variations,
including self-attention and multi-head attention. These processes have allowed for the creation
of cutting-edge models and have grown to be a necessary component of the NLP toolkit in the
field.
2.4. BERT
Bidirectional Encoder Representations from Transformers (BERT (0)) is an effective deep
learning model for natural language processing (NLP) that was introduced by Google in 2018.
This pre-trained language model can be fine-tuned for several NLP tasks, such as text
summarization and text comprehension. BERT’s capacity to comprehend context by taking into
account the words that surround a word in a phrase in both directions sets it apart from other
language models. This is accomplished via a novel pre-training method that trains the model on a
large corpus of text data using a masked language model and a task for predicting the following
phrase. For a variety of NLP tasks, such as question answering, named entity identification, and
sentiment analysis, BERT has produced state-of-the-art results. While it can properly capture the
meaning of a sentence or document, it has also demonstrated significant potential for text
summarization and text interpretation. BERT (0) can be fine-tuned for text summarizing to
provide accurate and informative summaries of lengthy articles or documents. It can accomplish
this by learning to identify and summarize the text’s key points. BERT may be used in text
comprehension to examine and extract the meaning from a sentence or document, allowing the

48
model to more accurately respond to queries, categorize text, or carry out other NLP activities.
Overall, BERT (0) is a formidable NLP tool that might revolutionize how we interpret and
analyze textual data. Text summarization, text interpretation, and other NLP tasks now have more
options thanks to their capacity to capture the meaning and context of language.
2.5. RoBERTa (0)
RoBERTa (0) (Robustly Optimized BERT Approach) is a state-of-the-art natural language
processing (NLP) model that was introduced by Facebook AI in 2019. RoBERTa (0) is based on
the same architecture as BERT, but it has been trained with improved pre-training techniques and
larger amounts of data, resulting in a significant improvement in performance on a range of NLP
tasks. One of the key differences between RoBERT (0) and BERT (0) is the training data.
RoBERTa (0)(0)(0) was trained on a much larger corpus of text, with up to 160 GB of training
data, compared to 16 GB for BERT. This larger corpus of text allowed RoBERTa (0)(0) to
develop a more nuanced understanding of language and improve its performance on a range of
NLP tasks. In addition to the larger corpus, RoBERTa (0) also uses a more advanced pre-training
approach. It removes the next sentence prediction task used in BERT and instead focuses on
optimizing the masked language modeling objective. It also uses dynamic masking, which
randomly masks different tokens during training rather than masking the same tokens in every
epoch. These improvements lead to better generalization and improved performance on a range of
NLP tasks. RoBERTa (0) has achieved state-of-the-art results on a range of NLP benchmarks,
including the GLUE benchmark, which evaluates a model’s performance on a range of natural
language understanding tasks. It has also shown impressive results in text classification, named
entity recognition, and question-answering tasks. Overall, RoBERTa (0) represents a significant
improvement in the performance of NLP models, and its success has led to the development of
many other models based on its architecture and pre-training techniques. It has opened up new
possibilities for NLP tasks, and its impact can be seen in many popular NLP applications today.
3. RESEARCH
3.1. How GPT-3 was Created
GPT-3 (Generative Pre-trained Transformer 3) is a language model created by OpenAI that uses
deep learning techniques to generate human-like text. GPT-3 bases on 3 main stages: Pre-
training: GPT-3 first goes through unsupervised training with its massive, internet-harvested set
of text data (such as Wikipedia, websites, articles, etc.) using a transformer-based neural network
architecture. Transformer-based (0) is a type of neural network architecture used in natural
language processing (NLP) applications (language translation, sentiment analysis, and text
creation). The concept of self-attention is the foundation of this model where each word in a
sequence is compared to all the other words in the same sequence to calculate a weighted
representation of the input sequence. This makes NLP be able to capture context and word
relationships more effectively than typical recurrent neural network (RNN) architectures.
Moreover, the Transformer (0) model consists of an encoder and a decoder, each includes several
layers of multi-head self-attention and feedforward neural networks. During training, the model is
optimized to predict the correct output given an input sequence. Fine-tuning: After being pre-
trained, GPT-3 may be fine-tuned for certain tasks, including text completion, translation,
summarization, and question-answering. Fine-tuning involves adapting the pre-trained model to a
specific task by providing it with labeled examples and adjusting the model’s parameters to
optimize performance. ChatGPT, for example, was fine-tuned with 175B weights.

49
Furthermore, ChatGPT used Transformers (0) and attention layer with tokens made up from a
piece of text.
Inference: After fine-tuning, the model is ready to generate text. Given a prompt (such as a
sentence or paragraph), GPT-3 uses its knowledge of language patterns and context to generate a
plausible continuation of the prompt. GPT-3 can generate text in various styles and tones, from
technical writing to creative writing, and can even mimic the writing style of a particular author.
3.2. Building a Summarizing Tool
In this project, we built a summarizing tool based on GPT. This tool was created with Streamlit,
open-source app framework for Machine Learning and Data Science, and Python programming
language. Firstly, we accessed the OpenAI website to obtain the API key for API calls and
training. We also gather 100 different prompts from sample research paper, Youtube video
transcript, and article with the responses of their summaries respectively.
After training and obtaining the finetuned model, the model was used as the core of our back end
to do the prediction. In our tool, there are 4 types of documents used for summarizing: article,
YouTube videos, research paper or PDF, and pure text. With article, we used "newspaper" library
to read infomation from URL. The tool used scrapping technique to fetch article’s content as well
as their thumbnail picture, author(s). The article’s content and thumbnail picture were kept for the
output of this model. Likewise, with Youtube videos, a tool called "youtube_transcript_api" was
used to fetch the content of the video. In this research, 100 videos with various length from 5 to
40 minutes were used to to be summarized by the tool. Next, with PDF/Research paper, we used
"fitz" module to extract text from PDF.
In this research, we tested the model’s ability to do IELTS reading test. Our IELTS (0) Reading
exam dataset came from British Council with answer key. We have collected over 25 exams with
more than 75 reading tasks. In the exam, there are multiple types of questions: Yes/No/Not Given
question, Fill in the blank, one choice, multiple choice, select correct corresponding paragraphs.
Figure 1: How the summarizing tool works

50
Figure 2: Example of question and answer process
3.3. Improving GPT’s Text Understanding on Long Text
In other to summarize a long chunk of text, the token limit of ChatGPT is 4096. We divided the
long text into multiple chunks of shorter texts of a maximum of 4096 tokens. Each short chunk of
text went through summarizing to retrieve the key idea from the text. The key ideas were then
connected to gain the big picture of the big idea. We also stored these short chunks of key ideas
in a database, which then were used to answer questions from the text.
Figure 3: Summarizing database
In terms of answering questions based on the text, we used get embedding module from OpenAI
to encode the question we wanted to ask from the text, each sentence, and each summarized main
idea into word embeddings. Then, a comparison of the similarity of the embeddings was done. By
doing the similarity comparison, it was able to answer the questions, which is similar to the way
human do IELTS Reading Test by scan and skimming techniques.
Figure 4: Similarity comparison
For visual inputs such as diagrams, flowcharts with questions such as fill in the blank. It was
challenging to understand the context of the blank’s location.

51
Figure 5: Example of visual fill-in-the-blank question
4. RESULTS
After testing GPT-3 on 50 IELTS Readings tests, we found that the average total accuracy of
GPT-3 on one test is 65.38%, and the total accuracy of all 75 tasks is 58.70% based on the actual
answers. This is equivalent to band 5.0-6.0 in IELTS. Our analysis of four main types of IELTS
reading tests includes: Yes/ No/ Not Given, Fill in the blank, Matching paragraph, and Multiple
choice. The result also suggested that GPT-3 performs best on Multiple choice questions.
Table 1: IELTS (0) Reading Result from GPT-3
Type of question Accuracy
Yes/No/Not Given 75.85%
Fill in the blank 90.65%
Correct letter 6.18%
One choice 85.03%
Multiple choice 18.84%
5. FUTURE WORK
For further research, we would love to test the capability of the GPT-4 model from OpenAI. By
taking visual inputs, some IELTS questions such as fill in the blanks in diagrams, flowcharts, and
visuals would be possible. One way we thought of was to use OCR technology and imagine a
captioning technique to describe the characteristics of a chart. In terms of imagine captioning, the
AI model usually employs a neural network architecture, specifically a combination of
convolutional neural networks (CNNs) and recurrent neural networks (RNNs). The CNN extracts
visual features from the image, which are then fed into an RNN that generates the corresponding
textual description. We are looking forward to training the model on the IELTS Reading Chart
images.
REFERENCES
[1] Radford, K. Narasimhan, T. Salimans, I. Sutskever, et al., “Improving language understanding by
generative pre-training,” 2018.
[2] Council, “Ielts,” 2016.
[3] T. Brants, A. C. Popat, P. Xu, F. J. Och, and J. Dean, “Large language models in machine
translation,” 2007.

52
[4] H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E.
Hambro, F. Azhar, et al., “Llama: Open and efficient foundation language models,” arXiv preprint
arXiv:2302.13971, 2023.
[5] Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin,
“Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017.
[6] M. Abadi, “Tensorflow: learning functions at scale,” in Proceedings of the 21st ACM SIGPLAN
International Conference on Functional Programming, pp. 1–1, 2016.
[7] S. Imambi, K. B. Prakash, and G. Kanagachidambaresan, “Pytorch,” Programming with
TensorFlow: Solution for Edge Computing Applications, pp. 87–104, 2021.
[8] N. Ketkar and N. Ketkar, “Introduction to keras,” Deep learning with python: a hands-on
introduction, pp. 97–111, 2017.
[9] L. R. Medsker and L. Jain, “Recurrent neural networks,” Design and Applications, vol. 5, pp. 64–
67, 2001.
[10] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional
transformers for language understanding,” arXiv preprint arXiv:1810.04805, 2018.
[11] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V.
Stoyanov, “Roberta: A robustly optimized bert pretraining approach,” arXiv preprint
arXiv:1907.11692, 2019.

A Comparative Study of Text Comprehension in IELTS Reading Exam using GPT-3

More Related Content

Similar to A Comparative Study of Text Comprehension in IELTS Reading Exam using GPT-3 (20)

More from AIRCC Publishing Corporation (20)

Recently uploaded (20)

A Comparative Study of Text Comprehension in IELTS Reading Exam using GPT-3