Urdu Poetry Generated by Using Deep Learning

Techniques
Muhammad Shoaib Farooq, Ali Abbas
Department of Computer Science, School of System and Technology, University of Management and Technology, Lahore,
54000
Corresponding author: Muhammad Shoaib Farooq ([email protected])

Abstract
This study provides Urdu poetry generated using different deep-learning techniques and algorithms. The
data was collected through the Rekhta website, containing 1341 text files with several couplets. The data
on poetry was not from any specific genre or poet. Instead, it was a collection of mixed Urdu poems and
Ghazals. Different deep learning techniques, such as the model applied Long Short-term Memory Networks
(LSTM) and Gated Recurrent Unit (GRU), have been used. Natural Language Processing (NLP) may be
used in machine learning to understand, analyze, and generate a language humans may use and
understand. Much work has been done on generating poetry for different languages using different
techniques. The collection and use of data were also different for different researchers. The primary
purpose of this project is to provide a model that generates Urdu poems by using data completely, not by
sampling data. Also, this may generate poems in pure Urdu, not Roman Urdu, as in the base paper. The
results have shown good accuracy in the poems generated by the model.
Keywords: Urdu Poetry, Poetry Generation, LSTM, GRU, Deep Learning.

Introduction section will provide discussion on different

methods along with their technical definitions and
Artificial Intelligence (AI) has helped computers
equations that are going to be used in this model.
change significantly over the past few decades,
from simple tools to smart machines. The fourth
industrial revolution witnessed the pervasive
application of artificial intelligence across Literature review
numerous industries, with observable effects on Applying deep learning techniques to generate poetry
self-driving vehicles, business analysis, self- in various languages has gained significant attention
learning machines, medical report analysis, recently. This review explores the landscape of Urdu
market exposure, and complex problems. Deep poetry generation using deep learning, focusing on the
learning NLP techniques have enabled AI to methodologies employed and the uniqueness of the
analyze human language and recognize images. dataset used. The literature review analyzes the work
Recent years have seen a notable increase in of other researchers in computational creativity,
research into language processing, particularly in explicitly focusing on poetry generation using deep
Urdu, including sentiment analysis, writing styles, learning techniques. The selected literature spans the
and poetry generation. However, the field of Urdu last 10 years and is sourced from reputable platforms
poetry creation still requires investigation. The such as JSTOR, IEEE Xplore, ScienceDirect, and
purpose of this paper is to produce Urdu poetry in SpringerLink. Various papers have discussed
Sales Urdu. The paper falls under the generic AI implementing BiLSTM, GPT, GPT-2, LDA, LSI,
domain and the sub-domain of Natural Language VSM, RNN, and CNN for poetry generation using
Processing. The literature review reveals a different datasets. Additionally, the review highlights
substantial gap in the generation of Urdu text, research on multilingual poetry generation,
which is the primary objective of this paper. encompassing Urdu, Arabic, English, Roman, and
Urdu's distinctive writing style is another factor, Hindi. The base paper's primary focus is on deep
as it is essential for the model's training. The learning-based Urdu poetry generation while
other essential but challenging aspect is data acknowledging the limitation of limited work on pure
acquisition and cleansing. The methodology Urdu poetry instead of Roman Urdu.

1|Page
A paper by [1] analyzes machines' historical creativity to classify Arabic poetry based on different emotions.
and the advancements that have improved their They have Arabic poetry data from online resources
computational methods. The paper focuses on the and implemented SVM, Naive Bayes, voting feature
enhanced computational creativity of machines, intervals, and hyper pipes.
exemplified by machine learning techniques. These [7] A paper worked on a neural network-based
machines can now generate and evaluate new artifacts technique for producing Urdu and Hindi poetry. The
based on given data. The study covers various levels model has been trained on an Urdu and Hindi poetry
of creativity, including autonomous evaluation, corpus, enabling it to generate compositions with the
change, and non-randomness. It explores learning to correct meter and rhyme scheme. [8] Another paper
evaluate using regression or classification techniques presents a model for creating poetry using deep
and the algorithms used in these models. learning techniques and enhanced phonetic and
semantic embeddings.
[2] The paper delves into machine creativity, Similar study done by [9] which analyzes the data
particularly in generating poetry through machine collection of Urdu poetry. This analysis concentrates
learning algorithms. They sourced data from online on the distribution of words, meters, and rhythms
repositories, such as Gutenberg, containing over 50 throughout the corpus. This paper [10] presents a
thousand eBooks. Emotion poems were classified technique for writing Urdu poetry using RNNs and
using the GPT-2 architecture, evoking emotions like LSTMs models. The model has been instructed in an
sadness, anger, anticipation, joy, trust, and dreams. Urdu poetry collection. In addition, the paper
Five showed high scores after training eight emotion examines the challenges of writing poetry in Indic
words and comparing them with subsets of the corpus. languages, such as their complex morphology and
The top 20 poems, out of 1000, were chosen based on syntax. [11] The model has been trained in a
GPT scores and reviewed by ten native English compilation of Urdu poetry and can write poems with
speakers. the correct meter and rhyme scheme.
[3] researchers have recently turned their focus to [12] The study examines the significance of social
emotion classification from web information. While media as a platform for sharing content and the
much of the existing research has centered on difficulty of comprehending how individuals
categorizing emotions from informal text like chat, communicate and express their opinions about
text messages, and social media content, less attention particular products. The study focuses on Urdu
has been paid to emotion classification from formal sentiment analysis due to the language's widespread
literature, such as poetry. The paper introduces a deep use, presenting the system architecture and
neural network model, specifically a BiLSTM model, experimental approaches and achieving a verified
for emotion classification from poetry text. They accuracy of 66 percent in sentiment analysis of Urdu
evaluate the model's performance using a benchmark comments on multiple websites.
poetry dataset and successfully categorize poems into [13] Presents a research study that utilizes a text
emotional types: love, anger, loneliness, suicide, and classification algorithm to identify poetry styles.
surprise. The proposed model's effectiveness is Knowledge representation algorithms are employed
compared against various baseline methods, including for text classification, enhancing the intelligent
machine learning and deep learning models. [4] analysis of poetry styles. The study combines various
utilizes neural memory to generate creative Chinese intelligent algorithms to develop an Urdu poetry style
poetry, acknowledging that statistical methods alone analysis system, including preprocessing the poetry
are insufficient for poetry generation. The paper documents in the corpus and mapping them into vector
incorporates machine learning, employing memory forms accessible by a computer. Simulation tests are
augmentation alongside RNN implementation. [5] conducted to verify the effectiveness of the poetry
Deep learning was applied for sentiment analysis style analysis system based on the text classification
using poetry from various online sources. The author method.
employed a bidirectional Long Short-Term Memory [14] Explores the use of natural language processing
(LSTM) model with CNN. The model achieved a peak and computational creativity to generate poetry in the
accuracy of 88%. [6] The paper uses machine learning Ottoman Turkish language through machine learning.

2|Page
The data was collected from OTAP, and a vocabulary used Markov chains and LSTM RNN models to
was created. The paper addresses three main generate poetry in the Romanian language. The
limitations. To address these challenges, the author authors [19] have focused on generating poetry rather
used a Finite State Transducer (FST) before training than based on emotions, as all did in the previous
the RNN recurrent neural network model, aiming to papers discussed above.
connect different syllables in the poem. However, this This paper by [20] has worked on classifying Urdu
approach has some limitations in generating variations Ghazals using 3000 different Ghazals datasets. Urdu
with a single word, for which the FSA model may be poetry has a vast and complex poetry structure because
used to score the words. it has a connection back to Arabs. In this paper, after
[15] The paper focuses on interactive poetry the data preprocessing, the authors have applied
generation using a model similar to the previous study. different machine learning techniques such as support
The author integrates recurrent neural networks with vector machine, random forest, decision tree, Naive
the Finite State Acceptor. They discuss the Bayes, and KNN models to identify the Urdu ghazals.
development of an app for real-time poetry generation Similar work has been done by [21] using the same
using machine learning models, where users input a machine-learning models. Different genres of poems
topic, and the app generates a poem accordingly. The were taken in the Arabic language. The author has
real-time nature of the process poses challenges, discussed that preprocessing was an essential step for
including slower generation capability due to on-the- this machine learning model as it has increased
spot topic input, limited evaluation of poem quality, classification accuracy. The highest average accuracy
and the model's inability to adjust the generated poem. obtained from the three models was 51 % from the
However, the app offers configurable tools to edit linear support vector classification. Here is another
poems, providing an interactive experience. paper by [22], which uses 450 Hindi poem data to do
The research was done by [16] on the generation of classification using supervised machine learning using
Urdu Poetry using the NLP technique for research in R language. They have used random forest and naive
neurocognitive poetry and styling of poetry on Bayes models to compare the misclassification error in
computers. Data for this study is collected from the corpus dataset. They have 83% accuracy in their
Rekhta, a famous Urdu poetry platform. The study model.
uses the data of 48,761 words of ghazals produced by The research done by [23] uses a Handwritten Urdu
4,754 poets in 800 years. The study uses Character Dataset; this dataset is found from the
Multidimensional Scaling to numerically analyze the Nastaliq Urdu script. The information was acquired
similarities and differences between different poetic from 750 Kashmir Valley residents and included
works done by different well-known poets. solitary and positioned characters and digits.
The authors of this paper [17] have analyzed the These authors [24] have presented a paper in which
emotions in the Roman Urdu text from the data of they have performed different unsupervised machine
approximately 10,000 sentences. The authors have learning models on different datasets to compare their
created a corpus of data. They have used a rule-based, accuracy in results. They have used three datasets
recurrent convolutional neural network model and N- collected from news articles: Allam Iqbal's and other
gram to evaluate the sentiment analysis on the corpus poets' poetry. They have done tokenization,
dataset. The results from the RCNN model have been lemmatization, removing stop words, and stemming
compared using Rule-based and N-gram models. on the corpus dataset.
[18] have also discussed similar models and generated [25] His paper has combined statistical analysis with
English and Romanian poetry through machine the NLP to learn about the architecture of poetry
learning experiments. The author has used a generation. This paper uses the following statistical
quantitative approach and provided a statistical models: LSI, LDA, and VSM. This has shown
experiment to compare two poetical texts using promising results from the lantern pattern recognition.
information entropy and N-gram informational Research done by [26] shows the extraction of
energy. After this statistical analysis, the author semantic and logical topics from any text collection;
jumped into a deep neural network model that RNN machine learning-based approaches proved helpful.
has used to generate Byron's poetry. Another paper has Different topic modeling approaches are used for Urdu

3|Page
poetry to show that these approaches are equally Significance of the study:
helpful for text generation.
These are just a few of the papers that have been The current research uses creative computing
published on Urdu poetry generation using ML methods to generate poetry in different
models. The field is still in its early stages, but there languages. NLP models have significantly
has been significant progress in recent years. As the simplified the analysis of business customer
models become more sophisticated, they can generate comments and reviews. Data Science has
more creative and semantically meaningful poetry. accelerated data analysis and pattern
Overall, the literature review highlights the progress recognition, making it more efficient than
and potential of using NLP and deep learning traditional methods. Intelligent machines are now
techniques for creative applications in various being taught languages and generating new
languages, providing valuable insights into content, including poetry. The paper aims to
computational creativity. generate Urdu poetry based on renowned Urdu
poets' styles and contribute to the field of Artificial
Intelligence. The objective is to replicate the
Statement of the problem and writing styles of legendary poets with high
significance of the study precision and enable computers to create poetry
for everyone. This research will revolutionize
Problem statement Urdu poetry and make it easier to write poetry in
the manner of renowned authors such as Mirza
This paper will focus on the Urdu Poetry Ghalib.
generation with the help of Deep Neural Networks
and Natural Language processing models such Methodology
as N-Gram, Bi-Gram models, Gated Recurrent This study employs a Recurrent Neural Network
Units, and Long Short-term Memory Networks. (RNN), a standard machine learning model for
language and sequence analysis, in its
Scope of the study investigation. Before delving into RNN, the
Forward Feed Neural Network is discussed,
Scope: which consists of unprocessed input, hidden, and
output layers. However, this model's requirement
 The study focuses on generating poetry in to consider previous inputs renders it
Urdu Salees, distinct from Roman Urdu, a inappropriate for detecting sequences in
relatively less explored area in poetry language processing. RNN overcomes this
generation research. limitation by preprocessing and storing the
 The research encompasses multiple genres previous input state, enabling it to evaluate both
of poetry, creating a diverse dataset for previous and current language states when
analysis and generation. predicting the next word following a given word.

Limitations:

 The data used in the study is anonymous and

not associated with any specific poet, which
might limit the ability to analyze poetry from
individual authors.
 The study utilizes the complete dataset,
forgoing sampling methods, which might
impact the generalization of results.
The RNN model has an initial concealed state
(h0) and receives input at every time step. The

4|Page
input and the previous hidden state (h(t-1)) are forward propagation graph is sequential,
processed by the hidden layer to produce the parallelization cannot reduce the runtime, which
current hidden state (h(t)). The current concealed remains O(τ). The memory cost is also O(τ) because
state is output and passed as the prior hidden the states computed during the forward pass must be
state to the following time phase. This procedure stored until used during the rearward run.
is repeated for each input at various time Backpropagation through time is accomplished by
intervals. The model employs shared weights (U, applying backpropagation to an unrolled graph at a
V, W) to process all concealed layers efficiently. cost of O(τ).

Forward Pass Computing Gradients

We presume a hyperbolic tangent activation function We have been provided with a loss function denoted
for the hidden layer in the illustration. Since RNN is by L; we must compute the gradients for our three-
used to predict words and characters, we regard the weight metrics, U, V, W, and the bias terms B and C.
output as unnormalized log probabilities for each Then, we update them with an L-based learning rate.
possible value of the discrete variable. As a post- Like conventional backpropagation, the gradient gives
processing phase, the SoftMax operation can express us an idea of how the loss varies for each weight
discrete variables as a vector of normalized parameter.
probabilities over the output.
We revise the weights W using the following equation
RNN forward pass is represented by the below to minimize loss:
equation:

Nodes in our computational graph (t) include the

parameters U, V, W, b, and c, as well as the sequence
of nodes indexed by t for x (t), h(t), o(t), and L. We
must recursively calculate the gradient nL for each
node n in the graph based on the gradient obtained at
nodes that follow it. The output gradient is determined
using o(t) as the SoftMax function argument to
The input sequence's duration corresponds to the generate the vector y of probabilities over the output.
output signal's length in this example. The entire loss
of all x values coupled with a set of y values equals the
sum of all losses across all time increments. The
Softmax function acquires the vector of probabilities
over the output by receiving the outputs as an operand.
Given the current input, the loss L is defined as the
negative log probability of the correct target. As shown in the diagram below, the hidden state h(t)
has a gradient that flows from both the present output
Backward Pass and the future concealed state at time t.

The forward propagation pass traverses the graph from

left to right, while the backward propagation pass
computes the gradient from right to left. Because the

5|Page
 Using the chain rule, the gradient calculations for each
parameter are as follows:

Issues

The RNN is a strong and straightforward model in

theory but challenging to train effectively. The
vanishing gradient and exploding gradient issues are
vital reasons for its difficulties. Backpropagation
Through Time (BPTT) training requires gradient
We commence after the sequence and proceed
propagation from the terminal to the initial cell. The
backward. At the final time step, h(τ) has only O(τ) as
sum of these gradients can either become zero or
a descendant; therefore, its gradient is straightforward.
develop exponentially, resulting in the problem of
expanding gradients, in which the gradient norm
substantially increases during training.

LSTM Model

Our research uses the RNN subtype, LSTM (Long et

We can then repeat backward in time, from t=1 to t=1, al.), as we are working on Urdu poetry generation. We
to back-propagate gradients across time, noticing that need a model that can process the sequence of the cell,
h(t) for t has both o(t) and h(t+1) as descendants. so LSTM is a perfect fit for our model. It detects which
words in Urdu poetry will come next to maintain the
The gradient is given below.
beauty of poetry.

The LSTM model consists of three gates. The first

component is the Forget Gate, followed by the Input
Gate and the Output Gate. In the neglect gate,
information is processed, such as which information in
a sentence must be essential and which is unnecessary.
After obtaining the gradients on the interior nodes of The irrelevant information is wiped out, and the
the computational graph, we can obtain the gradients excellent and essential information is passed to the
on the parameter nodes. input layer. Then, in the input, the data is processed for
the following upcoming output. Therefore, this project
is beneficial in language processing tasks.

6|Page
An LSTM network can learn this pattern, which occurs know about the stock market when asked whether it
every 12 time periods. It does not merely utilize the will rise or fall tomorrow.
previous forecast; it also maintains a longer-term
context, which helps it avoid the issue of other models' GRU Model
long-term reliance.
GRUs are a superior form of recurrent neural
Working Analysis networks. GRU utilizes the update and reset gates to
circumvent the RNN's vanishing gradient problem.
As mentioned, LSTM employs a series of gates to Two vectors determine the data to be transmitted to the
process information effectively. A typical LSTM output. They are exceptional because they can be
comprises three gates: neglect, input, and output. taught to retain knowledge from the past without
losing it over time or eliminating irrelevant data.
Forget Gate
Update Gate
The disregard gate is the beginning of the procedure.
We will determine which portions of the cell state are The update gate helps the model determine how much
advantageous using the previous concealed state and information from previous time steps should be carried
new input data. Consider each element of this vector forward to subsequent time steps. This is extremely
as a filter that allows more data to pass as its value useful because the model can replicate all historical
approaches 1. The transferred values are then data, eradicating the possibility of gradients
multiplied point-by-point with the previous cell state. disappearing. We will later examine how to utilize the
update gate.
Input Gate
Reset Gate
This stage attempts to determine what new data should
be recorded in the network's long-term memory based The model uses this gate to determine how much
on the previously concealed state and new input data. historical knowledge should be forgotten.
The new memory network and the input gate are
neural networks that receive the same inputs, the This is identical to the formula for the update gate. The
previous concealed state, and the new input data. The distinction lies in the weights and usage of the gate,
new memory network is an activated neural network which will be discussed subsequently. The restored
that has learned how to build a new memory update gate is depicted in the following diagram.
vector by combining the existing concealed state with
new input data. Given the prior concealed state
context, this vector comprises information from the
new input data. Given the new data, this vector
indicates how much each component of the network's
long-term memory should be updated. The input gate
is a sigmoid-activated network that determines which
portions of the new memory vector will be retained.

Output Gate

The output gate is responsible for determining the new

concealed state. This decision will be based on the
recently updated cell state, the prior concealed state,
and the new input data. One might assume that we
could output the updated cell state; however, this Working
would be akin to someone sharing everything they

7|Page
The initial step is to establish a candidate concealed GRU
state.
GRU is also known as a gated neural network. It
solves vanishing gradient problems that arise in
standard recurrent neural networks. GRU is an
improved version of LSTM because both are
constructed identically and, in some instances,
It multiplies the input and hidden state from the produce identical results. GRU is a more
previous timestamp, t-1, by the output of the reset gate sophisticated variant of Recurrent Neural
rt. The concealed state of the candidate is determined Network. It employs a reset gate and an update
by passing all of this information to the tanh function. gate. These two vectors determine which data is
transmitted to the output. They are helpful
because they can be programmed to store data
for an extended period, eliminating the need to
delete unnecessary data.

The most important aspect of this equation is how we

Update Gate:
use the value of the reset gate to determine how much
the preceding concealed state can influence the
The update gate helps the model determine how
candidate state. much historical data from previous time stages
should be transmitted to the future. This is
Hidden State especially effective because the model can
replicate all historical data and eliminate the
The candidate state, once obtained, generates possibility of gradient issues with vanishing
the concealed state Ht. Here, the Update values.
checkpoint enters into action. Unlike LSTM, GRU
utilizes a single update gate to govern historical Reset Gate:
information (Ht-1) and the most recent
information. The model utilizes this gate to determine how
much historical data should be forgotten. We
insert the h (t-1) and x t lines, multiply them by
their respective weights, add the outcomes, and
then flatten the curves with the sigmoid function.

If it is close to zero, the first term in the equation

Final Memory in the Present State
will vanish, indicating that the new hidden state
will not contain much information from the
If, during the sentiment analysis of a movie
previous one. The second portion merges into
review, the most important information is
one, indicating that the current concealed state
contained solely in the first line and all other
will only contain information from the candidate
information is irrelevant, then our model can
state.
extract the sentiment from the first line and
disregard the remaining text. It accepts the
current state's input, which triggers the Update
gate. Therefore, the Update gate is required to
complete this concluding stage.

Similarly, if the value of it is on the second term, Difference between GRU and LSTM:
the current concealed state will depend solely on
the first term.

8|Page
Long short-term memory refers to the RNN 𝑂𝑢𝑡𝑝𝑢𝑡:
architecture utilized in deep learning. LSTM
networks are ideally adapted for processing, ‫ملتا نہیں کامل یا نقص کچھ گیا‬
categorizing, and forming forecasts based on ‫ہے اتنا فرق میں نئی اور میں روشنی پرانی‬
time string data, as there may be indeterminate
time intervals between essential events in a time ‫نہیں انداز کا ساحل اسے ملتی نہیں کشتی اسے‬
string. In 2014, a solution was developed for the
‫ہوں میں کہ ہے کیا نے میں کے نام کی اس کہ ہے کیا‬
typical recurrent neural network issue of declining
gradient. GRU and LSTM share many similarities. ‫دی ہے آنکھیں مری تو کیا کیا مجھے یہ‬
Both algorithms regulate the process of
𝑂𝑢𝑡𝑝𝑢𝑡:
memorization using a gating technique.
‫ہو کیوں آسماں کا اس دشمن کے جس‬
Interestingly, GRU is less complicated and
computes much faster than LSTM. The ‫ہیں کہتے کو کس ستانا تو آزمانا ہے یہی‬
fundamental difference between GRU and LSTM ‫ہے کیا کیا میرا تو تم جب لیے ہو کے عدو‬
is that GRU has three gates—input, output, and
forget—whereas LSTM has only two gates— ‫کی نام کی اس بھی کو بازی مرے ہیں میں‬
reset and update. GRU has fewer gates than
‫میرا ہے کیا کیا تو اب ہے کی اس‬
LSTM, which makes it simpler. GRU is superior
to LSTM since it can be modified easily and does ‫نہیں میری نے میں‬
not require memory units. As a result, it is quicker
𝑂𝑢𝑡𝑝𝑢𝑡:
to train than LSTM and give according to
performance. GRU is used when the data series ‫ہوئے مارے‬
is small; LSTM is used when the data series is
‫نہیں میں مدفن اپنے گرد سوا کے گل جلوۂ‬
more extensive. GRU reveals the veiled layers
and the whole memory but not by LSTM. LSTM ‫کا ناسور نئے ہے ہیولی اک قطرہ قطرہ‬
and GRU are required for complex problem
domains such as machine translation, speech ‫کی بات کی ان کر دیکھ ذوق بھی خوں‬
recognition, speech synthesis, sentiment ‫کی بات کی بات کی بات کی اس‬
analysis, stock price prediction, and machine
comprehension. ‫کی بات کی بات کی بات کی اس‬
‫با کی بات کی بات کی اس‬
Results and Conclusion
This section presents the results of various deep
𝑂𝑢𝑡𝑝𝑢𝑡:
learning models applied to the Urdu poetry ‫ہیں گئے جھلس منظر سارے کہ میں دوپہر ہے برسی آگ وہ‬
generation problem statement. The models
utilized include LSTM and GRU. Specifically, we ‫ہو نہ کہیں اب تازگی وہ تھی تازگی جو سویرے یہاں‬
focus on the outcomes achieved with the GRU ‫بات کی بازی مرے ہے کیا کیا تو بھی کے اس‬
model across different epochs.
‫کی بات کی دور کی بات کی بات کی اس‬
‫میں مرا سودا‬ ‫کیا تو اب ہے کی اس‬
‫صحرا اندازۂ ترا شاید جنوں اے تھا غلط‬ 𝑂𝑢𝑡𝑝𝑢𝑡:
‫ہیں سکتے توڑ کو بو و رنگ طلسم اس سے خودی‬ ‫گزری کیا پہ ان بعد مرے کہ نہیں پتا‬
‫ویں شہر جانے نہ ہے کا طرح غم تعلقف یہی‬ ‫تھا آیا چھوڑ میں زمانے خواب چند میں‬
‫نہ سمجھتی پہ ساز ایک اک ہیں آئے وہ‬ ‫ہے کیا کیا تو بھی سے کوہ سہی ہی ذرے‬

9|Page
 ‫ہے کیا کیا میرا ہے کیا کیا نے میں‬ [6] Joglekar, M., & Mishra, A. (2021). Urdu &
Hindi Poetry Generation using Neural Networks.
‫کیا کیا نے میں‬
Paper:
https://www.semanticscholar.org/paper/Urdu-
The results of the model provide significant results. %26-Hindi-Poetry-Generation-using-Neural-
Further modifications can be made as the era of Mukhtar-
generative AI is started, further approaches can be Joglekar/f42bfb7518656aec622d69561eec0d8f9
used, and results can be refined further. 4dc7661. Semantic Scholar.

References: [7] Shah, S., & Bhardwaj, M. (2022). Poetry

[1] K. Scofield, "ARCHITECTURE FOR Generation Model via Deep Learning
AUTOMATIC POETRY GENERATION incorporating Extended Phonetic and Semantic
THROUGH PATTERN RECOGNITION," core. Embeddings. Paper:
Ac.UK, pp. 1-107, 2017. Available: https://www.researchgate.net/publication/35129
https://core.ac.uk/download/pdf/84871704.pdf. 7685_Poetry_Generation_Model_via_Deep_lear
ning_incorporating_Extended_Phonetic_and_Se
[2]A. Khattak, M. Asghar, H. Khalid, and H. mantic_Embeddings. ResearchGate.
Ahmad, "Emotion classification in poetry text
using deep neural network," Multimedia Tools [8] Ansari, M., & Zafar, A. (2022). Exploratory
and Applications, 2022. Available: Data Analysis of Urdu Poetry. Paper:
https://link.springer.com/article/10.1007/s11042- https://www.mdpi.com/2227-7390/11/2/435.
022-12902-3. MDPI.

[3]J. Zhang, Y. Feng, D. Weng, A. Abel, S. [9] Khan, M. S., & Mehmood, A. (2022).
Zhang, and A. Zhang, "Flexible and Creative Generating Urdu Poetry using RNNs and
Chinese Poetry Generation Using Neural LSTMs. Paper:
Memory," Computer Science Artificial https://www.researchgate.net/publication/35681
Intelligence, 2022. Available: 7609_Exploratory_Data_Analysis_of_Urdu_Poe
https://arxiv.org/abs/1705.03773. try. ResearchGate.

[4]S. Ahmad, M. Asghar, F. Alotaibi and S. [10] Verma, R., & Singh, S. (2022). A Neural
Khan, "Classification of Poetry Text Into the Network Model for Urdu Poetry Generation. In
Emotional States Using Deep Learning Advances in Natural Language Processing (pp.
Technique," IEEE, vol. 8, 2020. Available: 34-43). Springer, Singapore.
https://ieeexplore.ieee.org/abstract/document/90
66983. [11] Z. Rehman and I. Bajwa, "Lexicon-based
sentiment analysis for Urdu language," 2016
[5]O. Alsharif and D. AlShamaa, "Emotion Sixth International Conference on Innovative
Classification in Arabic Poetry Using Machine Computing Technology (INTECH), 2016.
Learning," International Journal of Computer Available:
Applications, vol. 56, no. 16, 2013. Available: https://ieeexplore.ieee.org/abstract/document/78
https://www.researchgate.net/publication/28432 45095?casa_token=B-
6319_Emotion_Classification_in_Arabic_Poetry JJdDAUBFkAAAAA:KHadqbNi8x7mpjAmLa
_Using_Machine_Learning. wHhyrhk99sX37uzHt2uSRroUqg1lXQ8bDJ11d
QMTLpIV0pItG65A-17vWs.

10 | P a g e
[12] C. Wang, "Analysis of Poetry Style Based on https://publications.muet.edu.pk/index.php/muet
Text Classification Algorithm," Scientific rj/article/view/1240.
Programming, vol. 2022, 2022. Available: [20] E. Fusu and M. Pura, "ROetry: First Steps
https://www.hindawi.com/journals/sp/2022/2763 Towards Computer Generated Poetry in
380/. Romanian Language," 2020. Available:
https://ieeexplore.ieee.org/document/9223175.
[13] S. Munir, S. Wasi, and S. Jami*, "A [21] O. Calin, "Statistics and Machine Learning
Comparison of Topic Modelling Approaches for Experiments in English and Romanian Poetry,"
Urdu Text," Indian Journal of Science and 2020. Available: https://www.mdpi.com/2413-
Technology, vol. 12, no. 45, pp. 1-7, 2019. 4155/2/4/92.
Available: https://indjst.org/articles/a- [22] Z. Mahmood et al., "Deep sentiments in
comparison-of-topic-modelling-approaches-for- Roman Urdu text using Recurrent Convolutional
urdu-text. Neural Network model," vol. 57, no. 4, 2020.
Available:
[14] S. Colton, J. Goodwin, and T. Veale, "Full- https://www.sciencedirect.com/science/article/ab
FACE Poetry Generation," 2012. Available: s/pii/S0306457319308441.
http://computationalcreativity.net/iccc2012/wp- [23] F. Mushtaq, M. Misgar, M. Kumar and S.
content/uploads/2012/05/095-Colton.pdf. Khurana, "UrduDeepNet: offline handwritten
Urdu character recognition using deep neural
[15] H. Arif, K. Munir, A. Danyal, A. Salman and
network," Neural Computing and Applications,
M. Fraz, "Sentiment Analysis of Roman
vol. 33, 2022. Available:
Urdu/Hindi using supervised methods," 2022.
https://link.springer.com/article/10.1007/s00521-
Available:
021-06144-x.
http://admin.umt.edu.pk/Media/Site/icic/FileMa
[24] G. Yolcu, "Binârî: A Poetry Generation
nager/Proceedings/Papers/ICIC_2016_paper_12.
System for Ghazals," 2020. Available:
pdf.
https://www.cmpe.boun.edu.tr/~gungort/undergr
[16] S. Rabbani and Z. Qureshi, "Exploratory aduateprojects/Binari%20-
Data Analysis of Urdu Poetry," Arxiv, 2021. %20A%20Poetry%20Generation%20System%2
[Online]. Available: 0for%20Ghazals.pdf.
https://arxiv.org/abs/2112.02145. [25] B. Bena and J. Kalita, "Introducing Aspects
[17] P. Bafna and J. Saini, "Hindi Poetry of Creativity in Automatic Poetry Generation,"
Classification using Eager Supervised Machine 2020. Available:
Learning Algorithms," 2020. Available: https://arxiv.org/abs/2002.02511.
https://ieeexplore.ieee.org/document/9167632. [26] S. Munir, S. Wasi, and S. Jami, "A
[18] M. Ahmed, R. Hasan, A. Ali, and M. Comparison of Topic Modelling Approaches for
Mohammed, "The classification of the modern Urdu Text," Indian Journal of Science and
Arabic poetry using machine learning," vol. 17. Technology, vol. 12, no. 45, 2019. Available:
Available: https://indjst.org/articles/a-comparison-of-topic-
http://journal.uad.ac.id/index.php/TELKOMNIK modelling-approaches-for-urdu-text.
A/article/view/12646.
[19] N. Tariq, I. Ijaz, M. Malik, Z. Malik, and F.
Bukhari, "Identification of Urdu Ghazal Poets
using SVM," vol. 38, 2019. Available:

11 | P a g e