Transactions Template
JOURNAL OF ENGINEERING RESEARCH AND TECHNOLOGY, VOLUME 9, ISSUE 2, OCTOBER 2022
12
Received on (15-03-2022) Accepted on (15-08-2022)
Generating Attractive Advertisement Text Campaigns Using Deep
Neural Networks
Atef Ahmed, Motaz Saad, and Basem Alijla
https://doi.org/10.33976/JERT.9.2/2022/2
Abstract—
Text generation task has drawn an increasing attention in the recent years. Recurrent Neural Networks (RNN)
achieved great results in this task. There are several parameters and factors that may affect the performance of the
recurrent neural networks, that is why text generation is a challenging task, and requires a lot of tuning. This study
investigates the impact of three factors that affect the quality of generated text: 1) data source and domain, 2) RNN
architecture, 3) named Entities normalization. We conduct several experiments using different RNN architectures
(LSTM and GRU), and different datasets (Hulu and booking). Evaluating generated texts is a challenging task.
There is no perfect metric judge the quality and the correctness of the generated texts. We use different evaluation
metrics to evaluate the performance of the generation models. These metrics include the training loss, the
perplexity, the readability, and the relevance of the generated texts. Most of the related works do not consider all
these evaluation metrics to evaluate text generation. The results suggest that GRU outperforms LSTM network,
and models trained on booking set is better than the ones that trained on Hulu dataset.
Index Terms— Deep learning, Recurrent Neural network, Advertisements campaigns, text generation.
I. INTRODUCTION
Online adverting is the process of marketing and advertis-
ing services and products over the internet Motaz . It has at-
tracted the interest of investors and business owners. For in-
stance, 77 % of EU businesses have a website and 26% of
them use internet to advertise. In addition, 86 % of EU enter-
prises used at least one type of social media to build their im-
age and to market their products [2]. The revenue of digital
ads was worth $126 billion [3]. A successful advertising cam-
paign is the one that has attractive ads, which are delivered to
relevant and interested consumers (audience) with the pre-
cise, meaningful, and relevant contents. Generating attractive
and successful ad campaign is beneficial and worthwhile, and
it is subject to reach target customers at the right time [4].
Institutional advertisers use targeted advertisement method to
generate attractive campaigns based on the requirements of
advertising exchange system [5]. The very old methods of
creating attractive contents of advertisement campaigns are
either by hand of content writer or automatically base on fil-
in-the blank" templates [6]. However, generating successful
advertising campaigns that meet the customer's needs is very
challenging, time-consuming and an expensive task. Signifi-
cant advertising knowledge and good understanding of cus-
tomers needs is required.
Machine learning and deep learning are successfully used
in various applications, including machine translation [7, 8],
text summarization [9, 10], text generation [11-15], speech-
to-text and text-to-speech [16]. Deep learning has evolved
many network architectures such as Recurrent Neural Net-
works (RNNs) [17], Long Short-Term Memory networks
(LSTM) [18], and Gated recurrent Unit networks (GRU)
[14]. Recent research showed impressive results of using
deep learning techniques in NLP applications such as
text generation and text summarization [11].
The work of [12] proposes a novel end-to-end model
named to generate the AD post. The authors split the AD post
generation task into two subprocesses: (1) select a set of prod-
ucts via the SelectNet (Selection Network). (2) generate a post
including selected products via the MGenNet (Multi-Genera-
tor Network). Concretely, SelectNet first captures the post
topic and the relationship among the products to output the
representative products. Then, MGenNet generates the de-
scription copywriting of each product. Experiments con-
ducted on a large-scale real-world AD post dataset demon-
strate that their proposed model achieves impressive perfor-
mance in terms of both automatic metrics as well as human
evaluations.
The work of [19] proposed explore the possibility of col-
laboratively learning ad creative refinement via A/B tests of
multiple advertisers. For generating new ad text, the authors
used an encoder-decoder architecture with copy mechanism,
which allows some words from the (inferior) input text to be
https://doi.org/10.33976/JERT.9.2/2022/2
Atef Ahmed, Motaz Saad, and Basem Alijla / Generating Attractive Advertisement Text Campaigns Using Deep Neural Networks (2022)
13
copied to the output while incorporating new words associ-
ated with higher click-through-rate.
In[20], the authors proposed a query-variant advertisement
text generation method that aims to generate candidate adver-
tisement texts for different web search queries with various
needs based on queries and item keywords. To solve the prob-
lem of ignoring low-frequency needs, they proposed a dy-
namic association mechanism to expand the receptive field
based on external knowledge, which can obtain associated
words to be added to the input. These associated words can
serve as bridges to transfer the ability of the model from the
familiar high-frequency words to the unfamiliar low-fre-
quency words. With association, the model can make use of
various personalized needs in queries and generate query-var-
iant advertisement texts.
This paper proposes a method of using deep learning mod-
els (LSTM and GRU networks) to generate attractive text ad-
vertising campaigns that meet customer needs using pre-de-
fined keywords. We investigate the generation of advertise-
ments text campaigns mainly in two domains: hotel Booking
and TV streaming (Hulu). In addition, two datasets in the do-
main mentioned earlier have been acquired and prepared for
this research, to train the neural networks to generate attrac-
tive ads, based on a given keywords feed as a seed to the neu-
ral network. Besides the automatic evaluation metrics (per-
plexity and readability [21]), human annotators subjectively
evaluated the readability and relevance of the generated ads.
The rest of this manuscript is organized as follows. Section
II describes the methodology of advertisement text generation
including: data acquisition, data Integration, data Pre-pro-
cessing, Ads generation, and the evaluation. Experimental
Studies and evaluation methods are presented in Section III.
The discussion and experiments results are presented in Sec-
tion IV, Finally, Section V presents Summary and Conclu-
sions.
II. DEEP NEURAL NETWORKS TO GENERATE
ADVERTISEMENT CAMPAIGNS
-shows the used meth خطأ! لم يتم العثور على مصدر المرجع.
odology in this work. The methodology consists of five main
steps: data acquisition, data integration, data pre-processing,
text generation and evaluation for generating Advertisement
text campaign using recurrent neural networks. These steps
are described in detains in the following sub-sections. Alt-
hough we use deep learning techniques but data prepro-
cessing is needed because the data is noisy as it is collected
from internet.
a. DATA ACQUISITION
The data is collected using SEMrush toolkit [22], it pro-
vides marketing information such as top ads, keyword
analytics, and search tracking, etc.., for a particular website.
The SEMrush retrieve and rank the Ads campaigns for the
top-ranked website using Google and Bing search engines.
Table 1 describes the main charachteristics of the
collected datasets. The collected data is limited to
adversisment campagin for hotel and flights reservation,
which is collected from Expedia.com and booking.com
websites, and TV and movies streaming collected from
Hulu.com websites. The data includes 42K text lines
(campaigns) from Booking and 13K text lines campaigns
from Hulu. The average campaign length is 67.53 and 227.07
for Booking and Hulu datasets respectively. The average
number of words per line is 11.13 and 39.15 for Booking and
Hulu datasets respectively. It is remarkable that Hulu
campaigns length is shorter than Booking campaigns as
shown in the table.
Table 1 : The main properties of collected dataset
Datasets Size
Max
Length
Min
Length
Average
Lengths
Average #
of words
Booking 42k 85 19 67.53 11.13
Hulu 13k 368 19 227.07 39.15
b. DATASET INTEGRATION AND PRE-
PROCESSING
Datasets were collected from two different sources. So, in-
tegrating data in a single and consistent representation is per-
formed. Then the dataset is pre-processed to be suitable to be
feeder to the neural networks. خطأ! لم يتم العثور على مصدر
-depicts the main steps of data pre-processing, includ المرجع.
ing data clearing and data normalization, and Name Entity
(NE) normalization.
Data cleaning involves the processes of removing and cor-
recting corrupt, unnecessary, or inaccurate records. So unnec-
essary HTML tags like , , and
are removed.
Moreover, all duplicated records in the data are deleted.
Data normalization involves the process of converting text
to lower case and removing special characters and punctua-
tions.
Named Entity refers objects name such as person's name, lo-
cation's name, and product's names [23]. To further normalize
the text, name entities are replaced with tag name using Geo-
Text [24] library and using static-NE list. Geotext [25] is A
Python Library used to extract country and city from given
text, and it is trained on data taken from geonames.org to rec-
ognize cities and countries names for another dataset or text.
All cities and countries are replaced by i-city and i-coun la-
bels respectively.
geotext may fail to recognize the names of some cities and
countries because geotext depends on the training of data. So,
static-NE list for cities and countries is proposed to overcome
Data
Accusition
Data
Integration
Data Pre-
processing
Text
Generation
using RNN
Evalution
Data cleaning Data normalization NE normalization
Figure 1: General Five Steps Metodology for Ads Generation
Figure 2: Pre-processing steps
Atef Ahmed, Motaz Saad, and Basem Alijla / Generating Attractive Advertisement Text Campaigns Using Deep Neural Networks (2022)
14
the limitation of geotext. Two lists of 4144 city names and
206 countries are collected from geonames.org. Conse-
quently i-city and i-coun labels are proposed to replace city
name and country name respectively.
III. EXPERIMENTAL STUDIES AND
EVALUATION METHODS
This section presents the proposed methods for advertise-
ment text generation. Two implementations denoted as shake-
spear TensorFlow (TF)1 and RNN TF Char2 and Word3 Lev-
els are adopted to implement the Recurrent Neural Network
(RNN) for GRU and LSTM encoding respectively. Both are
sequence-to-sequence model that take keywords (i.e., seed
text) as input to generate relevant text. For instance, Hotel,
Reservations, Flights, Booking, and travel are general key-
words that could be used for generating Ads related to Book-
ing domain. keywords such as Series, TV, Movies, Channels,
Episode, and Season could be supplied to the model for gen-
erating Ads related to Movie domain. The shake-spear TF
only support the character level, while the RNN TF support
both character level and word level encoding.
The following factors are considered in the application of
series of experiments to investigates their impact on the qual-
ity of generated advertisements campaigns.
• Dataset domain: Datasets in Booking/Reservations and
Movies (Hulu) domains are considered to train the NNs.
• Neural network architectures: LSTM and GRU neural
networks are investigated to generate the text.
• Name entity replacement: the impact of replacing
named entities with tags using GeoText and Static lists
are used to investigate the impact on the quality of gen-
erated texts.
• Input / output encoding level: character level and word
level encoding sequence are also explored.
The subsections present the experimental settings, and the
evaluation metrics.
a. PARAMETERS SETTINGS
Table 2 describes the parameters settings of LSTM and GRU
neural networks that are used in the experiments. The param-
eters settings of character-level GRU and both character-
level and word-level LSTM are presented. The parameters
values are the most recommended values, which are tunned
after a series of experiments.
Table 2: Parameters setting values for LSTM and GRU NN
Parameter
Char-level
LSTM
Word-level
LSTM
GRU
RNN size 128 256 512
Hidden layers 2 2 3
Sequence length 50 25 30
Number of epochs 2000 2000 10
Learning Rate 0.002 0.002 0.001
Optimizer Adam Adam Adam
1 https://github.com/martin-gorner/tensorflow-rnn-shakespeare
2 https://github.com/sherjilozair/char-rnn-tensorflow
The datasets that are used in our experiments are described
in Table 1. The datasets are split into three subsets, 70% for
Training subset, 15% for validation and 15% for testing. Ex-
perimental studies focus on character level encoding over the
word level encoding, because character encoding does not
suffer from out-of-vocabulary issues, and being able to model
different and rare morphological variants of a word, and do
not require segmentation [7].
b. EVALUATION METRICS
The neural networks are trained on the forementioned da-
tasets, and the evaluation metrics are the loss error and per-
plexity (PPL) criterion in order to judge the performance of
learning models [26]. Moreover, readability and relevance of
the generated text are subjectively assessed by human anno-
tators, and also readability is objectively evaluated with sta-
tistical propertied using a python tool called textStat [21].
Text relevance refers to the match between the information
inferred from the text and the reader’s goal [27]. In other
words, text relevance means the match between the gendered
text and the keywords/domains/campaigns used to for gener-
ation. The more match between reader's goals and inferred in-
formation the more relevant to the supplied keyworks. In this
study, A total of 54 Human annotators are hired from Amazon
Mechanical Turk to evaluate the generated texts. The annota-
tors are English native speaker and eligible to do “Human In-
telligence Tasks” (HITs)4. They are distributed into 18 groups
of three participant in a every group. Each group is provided
with the generated campaigns and the same keywords, which
are used in the generation process and asked to assess the rel-
evance of the text by answering to two points: rating scale (R)
for relevance and (I) for irrelevance. The majority answer of
the three answers is consider as output of evaluation results.
Readability refers to the rate of easy of understanding the
intended meaning of text. The Less complex, difficult, gram-
matical and linguistical errors text is the more readable text
[28].
The groups of annotators are also asked to evaluate the
generated campaign, and to assess readability in four-point
rating scale (easy, normal, difficult, and confusing). Three
annotators are asked to assess a given generated text and the
final readability label is determined by voting their annotation
as shown in table
Table 3.
Table 3 Votes to determine Readability
Evaluation Result Vote Result
Easy
Easy Easy
Standard
Difficult
Confused Vote Easy
Confusing
3 https://github.com/hunkim/word-rnn-tensorflow
4 Selecting eligible workers - Amazon Mechanical Turk
https://github.com/martin-gorner/tensorflow-rnn-shakespeare
https://github.com/sherjilozair/char-rnn-tensorflow
https://github.com/hunkim/word-rnn-tensorflow
https://docs.aws.amazon.com/AWSMechTurk/latest/AWSMechanicalTurkRequester/SelectingEligibleWorkers.html
Atef Ahmed, Motaz Saad, and Basem Alijla / Generating Attractive Advertisement Text Campaigns Using Deep Neural Networks (2022)
15
The Textstat tool uses the Flesch Reading Ease Score
(FRES) test to assess the overall readability of text
based on the Flesch Reading Ease Formula [29]. FRES
is a seven points diffculty scale, and human anotator in
this study evaluate readability in four point scal as
shown in Table 4. We make this mapping to convert
FRES measure to four point scale to be consistent with
human anotators evalautions.
Table 4: Normalize FRES to corresponding four-point scale
Score Difficulty Normalized 4-point Scale
90-100 Very Easy
Easy 80-89 Easy
70-79 Fairly Easy
60-69 Standard Standard
50-59 Fairly Difficult
Difficult
30-49 Difficult
0-29 Very Confusing Confusing
IV. EXPERIMENTS AND EVALUATION RESULTS
The experiments are conducted on a dedicated root server
with a minimal Ubuntu OS version. The server has RAM 64
GB DDR4, Hard drive SSD 500GB, graphics card GeForce®
GTX 1080, CPU Intel® Core i7-6700 Quadcore processor
built and connection speed with 1 GBit/s-Port. Python 3.5 and
TensorFlow with enabled GPU is used to implement the pro-
posed RNN architectures.
A series of experiments are conducted to investigate the
factors mentioned in Section III (dataset domain, NN archi-
tecture, name entity normalization, and input/ output se-
quence level), which affect the quality of generated texts.
Experimental results are presented in the next section.
a. DATASET DOMAIN EXPERIMENTAL STUDY
To investigate the influence of dataset domain on the gen-
erated text, a total of 102 Ads were generated by two Shake-
spear TF character-level models. The first one is trained Hulu
dataset, and the second one in trained on booking datasets.
Table 5 Shows PPL and training loss of TF Shakespeare char-
acter-level models trained on Hulu and Booking datasets.
Table 5 PPL and training loss of TF Shakespeare character-level models
trained on Hulu and Booking datasets
Datasets Loss Error PPL Relevance
Hulu 0.115 80 99%
Booking 0.503 45 99%
The results show that loss Error is 0.115 and 0.503 for
Hulu and Booking datasets respectively. The PPL values are
80 and 45 for Hulu and booking respectively. The results im-
ply that campaigns generated on booking domain fits better
than those that generated on Hulu domain. Human annotators
are totally agreed that 99% of the Ads generated in Hulu and
booking domains are relevant to the provided keywords.
presents the evaluation results of readability of Ads,
which are generated by GRU in Hulu and Booking dataset
domains.
Table 6: results of evaluating Readability for Ads generated by GRU in
Hulu and Booking domains.
Datasets Evaluator Easy Standard Difficult Confused vote
Hulu
Human 89% 8% 2% 1%
Textstat 96% 4% 0% 0%
Booking
Human 98% 0% 0% 2%
Textstat 92% 5% 3% 0%
The results show the percentage of evaluation Readability
as rated by human annotators and TextStat Tool. In general,
it can be noted from the results that the generated texts are
mostly readable. In Hulu domain, 89% and 96% are rated as
easy to read by human and textStat respectively. In the Book-
ing domain, 98% and 92% are rated easy to read by human
and textStat respectively. a very small percentage of Ads, 8%,
2%, and 1% are rated standard, difficult, and confused vote
respectively.
Figure 3 Compares Human evaluation and TextStat tool
evaluation results of readability for the Ads, which are gen-
erated by the GRU neural networks in Hulu and Booking
domains.
The results imply that the two methods of evaluation (i.e.,
Human annotators and TextStat) are very compatible. Also,
the GRU architecture extremely success in generating easy
to read advertisement for both domains.
b. NEURAL NETWORKS EXPERIMENTAL STUDY
This study investigates the performance of the GRU and
LSTM neural networks architectures which are implemented
by shake-spear TF RNN and TF GRU respectively.
Table 7 presents the training loss and the evaluation rele-
vance of Ads, generated by both GRU and LSTM neural net-
works using full character level encoding.
Table 7: Percentage of relevant Ads and training results (i.e., loss error
and PPL) of GRU and SLTM neural network on Booking dataset
NN Loss PPL Relevance
GRU 0.503 45 99%
LSTM 0.505 78 68%
Figure 3: Readability Text from Booking and Hulu dataset.
Atef Ahmed, Motaz Saad, and Basem Alijla / Generating Attractive Advertisement Text Campaigns Using Deep Neural Networks (2022)
16
We limit the dataset in this experiment to booking dataset
(102 ads) because it showed betters results than the Movie
domain dataset as shown in Table 5, and because the main
point in this experiment is to compare LSTM and GRU for
text generation.
The results of this experiments are shown in Table 7. The
results show that the training loss of LSTM and GRU trained
on the Booking dataset are 0.505 and 0.503 respectively. The
loss errors for both are very close to each other’s. On the other
had, the PPL results are significantly different (78 and 45 for
LSTM and GRU respectively). The results imply that Ads
generated by the GRU has fits better than the Ads, which gen-
erated by the LSTM. In addition, the results show that human
annotators rated Ads generated by GRU are more relevant
than the ones by LSTM. It can be observed that there is a sig-
nificant difference between in the performance the LSTM and
GRU in terms of PPL and the relevance.
Table 8 presents the results of evaluating readability (Hu-
man and TextStat) of Ads, which are generated by GRU and
LSTM at the character level encoding in Booking dataset do-
main. The results show that LSTM neural networks generates
43% and 45% easy to read Ads as rated by human and textStat
respectively, while the GRU generates 98% and 92% easy to
reads ads as rated by human and textStat respectively. More
than 55% As generated by LSTM rated as difficult or con-
fused Ads.
Table 8 Readability for Ads generated by GRU and LSTM in Booking
domains
NN Evaluator Easy Standard Difficult
Confused
vote
GRU
Human 98% 0% 0% 2%
Textstat 92% 5% 3% 0%
LSTM
Human 43% 4% 26% 27%
Textstat 45% 25% 30% 0%
The results imply that the Human evaluation is compatible
and supporting the evaluation results of TextStat tool. In gen-
eral, the result suggests that GRU network outperforms the
LSTM network in generating easy to read and more relevance
Ads text.
c. NAME ENTITY EXPERIMENTAL STUDY.
This experiment investigates the impact of NE normaliza-
tion on the quality of generated texts. So, the NE normaliza-
tion is applied on the training dataset (Booking).
Three experimental studies are conducted, NE normaliza-
tion are applied by two different tools, i.e., geotext tool and a
static list. We compare their impact on the DL model in Table
9, which presents the training loss and the PPL and the rele-
vance results using NE normalization by GeoText library,
static list, and without NE normalization for 102 ads. The
texts are generated by GRU model trained on booking data.
Table 9: Percentage of relevant Ads and training results (i.e., loss error
and PPL) of GRU on Booking dataset on three cases of NE removal
NE Library Loss PPL Relevance
Geotext 0.437 43 99%
static-NE list 0.468 40 99%
Without NE Normalization 0.503 45 99%
The results in Table 9 suggest that NE normalization has no
significant impact on the generated texts.
Table 10 presents the results of evaluating readability level of
Ads by generated by GRU NNs trained on Booking dataset.
The table includes the readability level of three cases (Geo-
Text library, static list, and without NE normalization).
Table 10: Results of evaluating readability of Ads generated by GRU on
Booking domains using different normalization techniques
NE Normalization Evaluator Easy Standard Difficult
Confused
vote
Geotext
Human 98% 2% 0% 0%
Textstat 77% 12% 11% 0%
static-list
Human 95% 2% 1% 2%
Textstat 77% 14% 9% 0%
No NE
Normalization
Human 98% 0% 0% 2%
Textstat 92% 05% 03% 0%
The results show that, In the case of applying NE by
geotext tool, Human rated 98% as easy to read and 2%
Ads as standard, while TextStat is evaluated 77%, 12 %
and 11% of Ads as easy to read, standard and difficult
respectively.
In the case of performing NE using static-list, Human
rated 95% of ads easy to read, 2% Ads standard, 1% dif-
ficult and 2% as confused vote. TextStat evaluated 77%,
14%, 9%, and 2% of generated ads as easy to read,
standard, difficult, and confused vote respectively.
The results in Table 10 suggest that the application of
NE normalization does not influence the human evalua-
tion either for readability or relevance, while the textStat
evaluation is negatively affected. The results show that
the percentage of easy-to-read ads is degraded from
98% to 77 % and the percentage of standard and diffi-
cult to read Ads is increased to 12% and 11% respec-
tively.
d. INPUT / OUTPUT TEXT SEQUENCE
EXPERIMENTAL STUDY
This experiment investigates the influence of encoding
level on the performance of LSTM neural networks. The ex-
periment is limited to LSTM because the shake-spear TF only
support character encoding, while RNN TF implementation
supports both character-level and word-level encoding.
Table 11 presents the training loss, the PPL, and the rele-
vance of Ads generated by the LSTM network trained on
Booking dataset on both character-level and word-level en-
coding.
Atef Ahmed, Motaz Saad, and Basem Alijla / Generating Attractive Advertisement Text Campaigns Using Deep Neural Networks (2022)
17
Table 11: Percentage of relevant Ads and training results (i.e., loss error
and PPL) of LSTM on Booking dataset for Character-level and word level
encoding.
Encoding level Loss PPL Relevance
Character 0.505 78 83%
Word 1.232 86 89%
It can be noted from training loss and the PPL results in
the table that the results of LSTM with character-level is bet-
ter than LSTM with word-level encoding level. On the other
hand, word level is more relevant than the character level, and
this is because of the text generated using the character level
scheme has some words that has some typos errors. The re-
sults also suggests that the character level model generates a
text that fits to the target text, while the word level model gen-
erates more relevant texts.
Table 12 presents the results of evaluating readability of
Ads, which are generated by LSTM with the character-level
and word-level encoding. Human and textStat tool evaluation
results is presented.
Table 12: Results of evaluating readability of Ads generated by LSTM
on character and word level encoding in Booking domain.
Encoding
level
Evaluator Easy Standard Difficult
Confused
vote
Character
Human 47% 06% 12% 35%
Textstat 43% 29% 28% 0%
Word
Human 96% 04% 1% 9%
Textstat 67% 14% 19% 0%
The results show that 47%, and 35% of Ads gener-
ated on for character-level are rated by human easy to
read and confused vote. TextStat evaluate that 43% and
29%, 28% are evaluated as easy to read, standard and
difficult respectively. In word-level encoding 96% of
Ads are rated by human as easy to read, and 9% are rated
as confused. TextStat evaluated 67% as easy to read and
14% standard and 19% difficult. The results show that
word-level LSTM performs better than character-level
LSTM in booking domain.
The results in Table 12 also suggest that the word
level scheme is better than the character level because
the texts generated by the character level have some
types / spell errors. In addition, the results in Tables 11
and 12 support this conclusion.
V. CONCLUSION
This research proposed the application of GRU and LSTM
deep neural networks in generating advertisement text cam-
paigns. Two datasets’ domains i.e., hotel Booking, and TV
and movies streaming are included. Presrocessing including
normalization, and Name Entity processing are performed to
reduce the number of strange names and prepare the dataset
for machine learning. The main contribution of this research
is to investigate the influence of four factors including, neural
network architecture, dataset domain, NE normalization, and
input encoding (character / word levels), on generating Ads.
Readability of the generated Ads is subjectively evaluated us-
ing human annotators and objectively assessed using TextStat
tool, whereas the relevancy is only evaluated by human
annotators. The implication is several factors could be tuned
to improve the performance of neural network in generating
attractive Ads. Several experiments have been conducted to
investigate the impact of the factors mentioned earlier. An in-
vestigation has been conducted to determine the influence of
every factor on the quality of generated text. In general, the
results indicate that the GRU networks outperform the LSTM
networking in generating easy to reads ads campaign. In ad-
dition, Training GRU NN on Booking domain has better per-
formance compared to Hulu dataset domain.
It can be concluded from the results that the collected data and
dataset domain and input/output encoding level are the most
common factors influence the performance of the generated
texts
For future work, generating advertisement campaign in Ara-
bic language will be investigated. More experiments on other
dataset domain including brands, shopping product need to be
conducted too. Besides that, investigations pertaining to gen-
erate multiple ads campaigns for every keyword is required in
divers Ads.
REFERENCES
1. Handayani, W., S. Muljaningsih, and H.J.T.S.O.S.J.
Ardyanfitri, Online Marketing Supports Promotion And
Advertising Sales In Communities Dolly Localization.
2018. 2(1): p. 75-87.
2. Eurostat, Internet advertising of businesses-statistics on
usage of ads Statistics
Explained. 2018.
3. IBA, 2019 Internet Ad Revenue Report. 2020.
4. Kong, S., et al., Web advertisement effectiveness
evaluation: Attention and memory. 2019. 25(1): p. 130-
146.
5. Bhatia, V. and V. Hasija. Targeted advertising using
behavioural data and social data mining. in 2016 Eighth
International Conference on Ubiquitous and Future
Networks (ICUFN). 2016. IEEE.
6. Bartz, K., C. Barr, and A. Aijaz. Natural language
generation for sponsored-search advertisements. in
Proceedings of the 9th ACM Conference on Electronic
Commerce. 2008.
7. Lee, J., K. Cho, and T.J.T.o.t.A.f.C.L. Hofmann, Fully
character-level neural machine translation without explicit
segmentation. 2017. 5: p. 365-378.
8. Singh, S.P., et al. Machine translation using deep learning:
An overview. in 2017 international conference on
computer, communications and electronics (comptelix).
2017. IEEE.
9. Allahyari, M., et al., Text summarization techniques: a brief
survey. 2017.
Atef Ahmed, Motaz Saad, and Basem Alijla / Generating Attractive Advertisement Text Campaigns Using Deep Neural Networks (2022)
18
10. El-Kassas, W.S., et al., Automatic text summarization: A
comprehensive survey. 2021. 165: p. 113679.
11. Iqbal, T., S.J.J.o.K.S.U.-C. Qureshi, and I. Sciences, The
survey: Text generation models in deep learning. 2020.
12. Chan, Z., et al. Selection and Generation: Learning
towards Multi-Product Advertisement Post Generation. in
Proceedings of the 2020 Conference on Empirical
Methods in Natural Language Processing (EMNLP). 2020.
13. Hughes, J.W., K.-h. Chang, and R. Zhang. Generating
better search engine text advertisements with deep
reinforcement learning. in Proceedings of the 25th ACM
SIGKDD International Conference on Knowledge
Discovery & Data Mining. 2019.
14. Taneja, P. and K.G. Verma, Text Generation Using
Different Recurrent Neural Networks. 2017.
15. Yang, X., et al. Advertising keyword recommendation
based on supervised link prediction in multi-relational
network. in Proceedings of the 26th International
Conference on World Wide Web Companion. 2017.
16. Hoobyar, T., T. Dotz, and S. Sanders, NLP: The Essential
Guide to Neuro-Linguistic Programming. 2013: William
Morrow.
17. Yang, Z., et al., Review networks for caption generation.
2016. 29: p. 2361-2369.
18. Xiang, L., et al., Novel linguistic steganography based on
character-level text generation. 2020. 8(9): p. 1558.
19. Mishra, S., et al. Learning to create better ads: Generation
and ranking approaches for ad creative refinement. in
Proceedings of the 29th ACM International Conference on
Information & Knowledge Management. 2020.
20. Duan, S., et al. Query-Variant Advertisement Text
Generation with Association Knowledge. in Proceedings of
the 30th ACM International Conference on Information &
Knowledge Management. 2021.
21. Bansal., S. textstat Python Tool. 2018 2018 [cited 2021
8/9/2021]; Available from:
https://github.com/shivam5992/textstat,.
22. Vinayak, S., SEMrush Toolkit For Bloggers: 8 Tools To
Boost Blog Content For More Traffic And Revenue. 2021.
23. Golikova, D.M.J.V.O., Named Entities for Computational
Linguistics. 2018. 15(1): p. 207-215.
24. Hu, Y.J.G.C., Geo‐text data and data‐driven geospatial
semantics. 2018. 12(11): p. e12404.
25. Palenzuela, Y.M. geotext. 2014 [cited 2021 6-9-2021];
Available from:
https://geotext.readthedocs.io/en/latest/readme.html.
26. Klakow, D. and J.J.S.C. Peters, Testing the correlation of
word error rate and perplexity. 2002. 38(1-2): p. 19-28.
27. McCrudden, M.T., G. Schraw, and B. Hoffman, Text
Relevance, in Encyclopedia of the Sciences of Learning,
N.M. Seel, Editor. 2012, Springer US: Boston, MA. p.
3307-3310.
28. Deutsch, T., M. Jasbi, and S.J.a.p.a. Shieber, Linguistic
features for readability assessment. 2020.
29. Farr, J.N., J.J. Jenkins, and D.G.J.J.o.a.p. Paterson,
Simplification of Flesch reading ease formula. 1951. 35(5):
p. 333.
Atef Ahmed. Holds a master’s degrees in data science from
The Islamic University of Gaza, and he is a software engi-
neer.
Motaz Saad is a computer scientist, he holds a Ph.D. degree
in computer science from the University of Lorraine, France.
His research interests include AI, NLP, and machine learn-
ing, and he published several papers in the field. He is cur-
rently working as an assistant professor at The Islamic Uni-
versity of Gaza, Palestine.
Basem O. Alijla received the Ph.D. degree in intelligent
systems from The University Science Malaysia (USM), in
2015. He is currently Assistant Professor in Computer Sci-
ence and deputy Dean Faculty of Information Technology,
Islamic University of Gaza. He published several research
papers in high impact factor journals and international con-
ferences. His research interest includes evolutionary compu-
ting, optimization, machine learning, data mining and fea-
tures extraction and selection.
https://github.com/shivam5992/textstat
https://geotext.readthedocs.io/en/latest/readme.html