GenCrawl: A Generative Multimedia Focused Crawler for Web Pages
Classification
Domenico Benfenati
a
, Antonio Maria Rinaldi
b
, Cristiano Russo
c
and Cristian Tommasino
d
Department of Electrical Engineering and Information Technology (DIETI), University of Naples Federico II, Naples, Italy
Keywords:
Web Crawling, Web Pages Classification, Generative AI, Web Topic Analysis.
Abstract:
The unprecedented expansion of the internet necessitates the development of increasingly efficient techniques
for systematic data categorization and organization. However, contemporary state-of-the-art techniques often
need help with the complex nature of heterogeneous multimedia content within web pages. These challenges,
which are becoming more pressing with the rapid growth of the internet, highlight the urgent need for ad-
vancements in information retrieval methods to improve classification accuracy and relevance in the context
of varied and dynamic web content. In this work, we propose GenCrawl, a generative multimedia-focused
crawler designed to enhance web document classification by integrating textual and visual content analysis.
Our approach combines the most relevant topics extracted from textual and visual content, using innovative
generative techniques to create a visual topic. The reported findings demonstrate significant improvements
and a paradigm shift in classification efficiency and accuracy over traditional methods. GenCrawl represents a
substantial advancement in web page classification, offering a promising solution for systematically organiz-
ing web content. Its practical benefits are immense, paving the way for more efficient and accurate information
retrieval in the era of the expanding internet.
1 INTRODUCTION
The rapid growth of the web has led to an over-
whelming amount of information, with over 5 billion
web pages available online (Kunder, 2018; Bergman,
2001). Effective web page classification is crucial
for various applications, including information re-
trieval, content recommendation, and search engine
optimization. However, web content’s diverse and dy-
namic nature, including textual and multimedia ele-
ments, presents significant challenges for traditional
classification methods (Chakrabarti et al., 1999). Pre-
vious approaches to web page classification have pri-
marily focused on analyzing textual content, often ne-
glecting the valuable information embedded in visual
elements (Rinaldi et al., 2021c; Rinaldi et al., 2021b).
While some methods have attempted to incorporate
multimedia data, they typically treat textual and visual
content separately, failing to leverage the synergis-
tic potential of combining these modalities (Ahmed
a
https://orcid.org/0009-0008-5825-8043
b
https://orcid.org/0000-0001-7003-4781
c
https://orcid.org/0000-0002-8732-1733
d
https://orcid.org/0000-0001-9763-8745
and Singh, 2019). Furthermore, the complexity and
resource-intensive nature of processing large volumes
of multimedia content necessitate the development
of more efficient and scalable solutions (Fern
`
andez-
Ca
˜
nellas et al., 2020). Introducing GenCrawl, a
genuinely innovative, generative, multimedia-focused
crawler, in response to these challenges. This novel
approach, which integrates textual and visual con-
tent analysis, promises to enhance web page clas-
sification accuracy and revolutionize the field. Our
work makes some primary contributions: we intro-
duce a novel interdisciplinary approach that combines
textual and visual content analysis, significantly im-
proving the classification performance of multimedia-
focused web crawlers. This comprehensive approach
ensures that no aspect of web content is overlooked,
enhancing the accuracy of our classification system.
A conventional focused crawler relies on link-based
navigation, which may not prioritize systematic data
acquisition and processing. This problem limits dis-
cerning meaningful patterns, extracting valuable in-
sights, and adapting to dynamic web content evolu-
tion. The data generated may lack refinement, poten-
tially leading to lower data quality and hindering anal-
ysis accuracy (Kumar and Aggarwal, 2023). Ethical
Benfenati, D., Rinaldi, A., Russo, C. and Tommasino, C.
GenCrawl: A Generative Multimedia Focused Crawler for Web Pages Classification.
DOI: 10.5220/0012998900003838
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 16th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2024) - Volume 1: KDIR, pages 91-101
ISBN: 978-989-758-716-0; ISSN: 2184-3228
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
91

and privacy considerations may also pose challenges.
Adopting a more adaptive framework for web crawl-
ing strategies could enhance the effectiveness of con-
ventional focused crawlers. The synergistic integra-
tion of deep learning techniques can enhance crawler
proficiency in analyzing and categorizing web pages,
ensuring seamless incorporation of diverse instances
into the evolving web page classification task (K et al.,
2023). This approach addresses web page classifica-
tion intricacies and advances information representa-
tion and retrieval methodologies in the expansive and
dynamic web-based knowledge dissemination era.
The article is organized as follows: in Section 2
a literature review is presented and discussed, putting
in evidence the novelties of our approach; the system
architecture and the proposed methodology for crawl-
ing strategy and web pages classification methodol-
ogy are discussed in Section 3; a use case of our
crawler together with experimental results are in Sec-
tion 4; eventually, conclusions and future works are
presented in Section 5.
2 RELATED WORKS
General-purpose and special-purpose web agents are
the categories into which web crawlers fall (Bhatt
et al., 2015). Instead of providing a thorough anal-
ysis of general-purpose crawlers, this section high-
lights relevant research on focused crawlers within
the framework of web page classification. Since our
framework is based on ontologies, we carefully eval-
uate works that utilize and conform to this approach.
On the other hand, we present a summary of alterna-
tive approaches, emphasizing studies that show how
well Convolutional Neural Network (CNN) features
operate as general feature extractors for multimedia
content retrieval tasks.
A focused crawler filters millions of pages and
finds relevant resources distant from the initial batch.
Machine learning approaches are used to train fo-
cused crawlers, which can be extracted from on-
line taxonomies or manually classified datasets (Pant
and Srinivasan, 2005). The seminal work on fo-
cused crawlers is (Chakrabarti et al., 1999), where
the authors developed two hypertext mining software:
a classifier for document relevance assessment and
a distiller for identifying hypertext nodes providing
multiple access points to relevant pages. Moreover,
genetic algorithms show intriguing results when it
comes to concentrated crawling.
Ontology-based crawlers employ unsupervised
ontology learning and domain-based ontology with
multi-objective optimization for improved crawling
performance and selection of weighted coefficients
for web pages (Hassan et al., 2017; Liu et al., 2022;
Russo et al., 2020), or for improvement of visu-
alization and document summarization (Rinaldi and
Russo, 2021). Event-based crawlers utilize event
models and temporal intent recognition methods, in-
cluding Google Trends data, to capture and priori-
tize event-related information (Farag et al., 2018; Wu
and Hou, 2023). Phishing detection crawlers use
isomorphic graph techniques to detect phishing con-
tent by identifying subgraph similarities between web
pages (Tchakounte et al., 2022). Machine learning
crawlers implement LSTM and CNN for word em-
beddings and classification, and Attention Enhanced
Siamese LSTM Networks for predicting web page
relevance in specific domains like biomedical infor-
mation (Shrivastava et al., 2023; Mary et al., 2022).
Rule-based and specialized domain crawlers lever-
age rule-based approaches for obfuscating audio file
crawlers in the AIR domain and incremental crawling
systems for the Dark Web (Benfenati et al., 2023; Fu
et al., 2010). Genetic algorithm-based crawlers utilize
modified Genetic Algorithms for web page classifi-
cation based on keyword feature sets (Fatima et al.,
2023). Additionally, an improved genetic algorithm
can increase the focused crawler’s memory and preci-
sion while broadening its search area, focusing on the
direct influence of textual content and topic on user
information retrieval (Yan and Pan, 2018).
The strategy proposed in this article introduces a
multimedia-focused crawler for web page classifica-
tion, which combines textual and visual topics from
text and images as in (Rinaldi et al., 2021a). It uses a
supervised learning algorithm to classify web pages,
including those with text and images. In the latter
case, a generative model extracts and creates images,
improving visual topic extraction and crawling per-
formance. The crawler’s flexibility and adaptabil-
ity to different domains make it more adaptable to
predefined keywords or exemplary documents. Our
study compares a web page classification method us-
ing text or images with existing methods, revealing
superior accuracy, recall, and greater resilience to
noisy or irrelevant content. We also discuss its ben-
efits and drawbacks and suggest future enhancement
directions.
3 PROPOSED FRAMEWORK
This section describes in detail how the proposed
framework is composed, indicating specifically what
the components are and how they function.
Figure 1 shows a sketch of the proposed system
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
92

architecture. Our system involves multiple modules
for crawler tasks. It starts with an online document
repository, generating crawler threads. These threads
retrieve web pages, analyze structures, classify text
and images, and use a model for synthetic image gen-
eration if direct comparison isn’t possible. Extracted
topics from text and images are combined for doc-
ument classification. The crawling process then se-
lects the next page from the repository. The crawler
initialization phase begins by gathering seed URLs
from an online document repository and setting up
a structured workflow for the subsequent stages of
the crawling process. Following this, the web page
retrieval and hyperlink extraction phase constructs a
network of interconnected pages, creating a compre-
hensive dataset that mirrors the web’s interconnected
nature. In the parsing phase, the crawler differenti-
ates between main content (relevant text) and auxil-
iary content (unrelated text), focusing on essential el-
ements. It also identifies visual information, adding
a multimedia dimension to the understanding of web
pages.
In the web page classification phase, machine
learning models categorize textual and visual ele-
ments based on their topics. This classification pro-
cess automates the evaluation of content relevance,
enhancing system efficiency and accuracy. When dis-
crepancies arise between textual and visual classifi-
cations, the generative phase harmonizes these inter-
pretations. Using a Latent Diffusion Model (LDM)
for image generation, it synthesizes images that align
with textual descriptions, ensuring cohesive content
representation. The combined topics of textual con-
tent and generated images are then used to calculate a
priority value for subsequent actions, indicating each
web page’s significance in the crawler’s exploration.
The preprocessing module is crucial for analyzing
textual and visual information on web pages, using
the DMOZ web collection as a data source. Despite
its official closure in 2017, DMOZ remains valuable
for classification tasks due to its heterogeneous na-
ture. A screenshot of this repository is available on
the Kaggle Dataset platform
1
.
The text preprocessing pipeline involves HTML
parsing, normalization, tokenization, stopword re-
moval, lemmatization, removal of special characters
and symbols, spell-checking, feature extraction, and
vectorization. These steps ensure consistent represen-
tation, enhance topic extraction efficiency, and pre-
pare the text for analysis.
For encoding the text of web pages, we used
1
https://www.kaggle.com/datasets/shawon10/
url-classification-dataset-dmoz?select=URL+
Classification.csv
SBERT (Cheng et al., 2023) to create vector repre-
sentations capturing semantic meanings. Clustering
techniques grouped sentences into topics based on se-
mantic similarity, extracting representative keywords
and generating labels for each topic using a rule-based
method. This approach allowed the identification of
main themes or concepts in the web pages.
The focused crawler’s image processing pipeline
involves a sequential process of extracting and ana-
lyzing images from web pages, starting with HTML
parsing, downloading and preprocessing images for
quality enhancement, and extracting relevant features
as feature vectors. The critical phase involves apply-
ing advanced computer vision techniques, explicitly
utilizing the VGG19 model (Simonyan and Zisser-
man, 2014). VGG19 is a mighty Convolutional Neu-
ral Network (CNN) consisting of 19 layers. It has
been trained on a large and diverse dataset featuring
complex image classification tasks, such as ImageNet
(Ridnik et al., 2021), a large image dataset consisting
of 14,197,122 images, each tagged in a single-label
fashion by one of 21,841 possible classes. VGG19’s
adaptability makes it well-suited for the task of visual
topic extraction in our work, as it excels at discerning
patterns and objects in the analyzed images. Figure 2
represents the described textual and visual pipelines.
3.1 Textual Topic Detection
We employ a pipeline to identify representative topics
in text using Sentence-BERT (SBERT) embeddings
and WordNet synsets. Text is preprocessed through
lowercasing, sentence tokenization, and removal of
common linguistic artifacts, such as stopwords, spe-
cial characters, etc, for uniformity. SBERT embed-
dings encode each sentence into a high-dimensional
vector space, capturing contextual relationships. The
SBERT application to the preprocessed text ensure
that the vector created by the model doesn’t effect the
noisy informations in the long text of the web pages
of the dataset. K-means clustering identifies distinct
topics, while WordNet synsets enrich semantic inter-
pretation.
This method’s effectiveness relies on the input
text’s language richness and diversity, facilitating
granular content exploration.
3.2 Visual Topic Detection
The visual topic detection task utilizes multimedia
components, particularly images, to determine a doc-
ument’s principal topic, enhancing the overall frame-
work’s performance. Our research currently focuses
on recognizing multimedia descriptors to measure the
GenCrawl: A Generative Multimedia Focused Crawler for Web Pages Classification
93

Figure 1: System architecture overview.
Figure 2: A detailed pipeline for textual (on the left) and visual (on the right) topic extraction and detection task.
similarity between document images and our multi-
media knowledge base. Diverse descriptors, includ-
ing local, global, and deep features, are evaluated (see
Section 3). We employ the pre-trained VGG19 model
on ImageNet, applied to the complete image, to iden-
tify and visualize relevant image regions associated
with specific predictions, facilitating efficient transfer
learning. This model’s straightforward architecture
promotes ease of interpretability and implementation.
If more than one image are detected from the web
page, we need to select only the images that are more
compliant with the textual topic.
3.3 Text-to-Image Generation
We incorporated a Latent Diffusion Model (LDM)
into the crawling process to address the challenge of
web pages lacking relevant multimedia content. This
model generates high-quality images consistent with
the textual description of the web page, even if no
multimedia data is present initially. The Stable Diffu-
sion (Rombach et al., 2022) serves as the latent model
for translating text into images. Based on a latent dif-
fusion approach, it is tailored for generating and ma-
nipulating images from textual prompts. The model
utilizes the pre-trained CLIP ViT-L/14 text encoder
(Radford et al., 2021), following Imagen’s methodol-
ogy (Saharia et al., 2022). The choice of Stable Dif-
fusion over other generation techniques is justified by
its superior performance, as demonstrated in Section
3.
To show that the visual topic extraction procedure
works in the same way, using generated images, we
chose to show the predictions made by the VGG19
model on the example Figure 3 and extract the prob-
abilities for the first three classes predicted by the
model. As shown in Table 1, the top 3 predicted
classes fully reflect the content of the image, includ-
ing the topics indicated within the textual prompt used
to generate the image.
Table 1: Top 3 Prediction probabilities of generated image
using the prompt ”a parrot that rides a bicycle”.
Top Prediction Probability
macaw 0.919
bicycle-built-for-two 0.027
lorikeet 0.017
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
94

Figure 3: Image generated using the prompt ”a parrot that
rides a bicycle”.
3.4 Combined Topic Extraction
This paper aims to improve classification perfor-
mance by combining two classifiers: one for text-
based topic detection and another for visual-based
topic detection. Existing research suggests that dif-
ferent classifiers may provide complementary infor-
mation models based on the specific patterns requir-
ing classification (Mohandes et al., 2018; Clinchant
et al., 2011).
The textual and visual classifications are com-
bined by normalizing scores from different topic de-
tection methods and scaling them to fit within the
[0,1] interval. The fusion of textual and visual clas-
sifications adopts various schemes, and in line with
(Rinaldi, 2014), this study opts for the SUM op-
erator and the Ordered Weighted Averaging (OWA)
operators proposed by (Yager and Kacprzyk, 2012).
These operators provide a systematic way to aggre-
gate the results, allowing for a more robust and com-
prehensive integration of information from both tex-
tual and visual modalities. The SUM function stands
out as one of the widely adopted techniques for lin-
ear combinations of classifiers, offering various ver-
sions such as weighted sum and average. Its preva-
lence in ensemble methods is attributed to its supe-
rior noise tolerance, contributing to enhanced over-
all performance compared to other elementary func-
tions (Kittler et al., 1998). The versatility of the SUM
function makes it particularly effective in capturing
the collective decision-making power of diverse clas-
sifiers, making it a popular choice in ensemble learn-
ing approaches.
Formally an OWA operator of size n is a function
F : R
n
→ R with a collection of associated weights
W = [w
1
,...,w
n
] whose elements are in the unit range
such that
∑
n
i=1
w
i
= 1. The function is defined as:
F(a
1
,..., a
2
) =
n
∑
j=1
w
j
b
j
(1)
where b
j
represent the jth value of the ⃗a vector or-
dered.
4 EXPERIMENTAL RESULTS
This section details the experiments conducted to
evaluate the performance of key components within
the proposed framework, specifically focusing on tex-
tual, visual, and combined topic detection strategies.
The system outlined in this study exhibits a high de-
gree of generalization owing to the inherent versatility
of the developed modules.
4.1 Dataset Description
This study employs a parsed and pre-elaborated mul-
timedia dataset derived from DMOZ (Sood, 2016), a
renowned and extensive multilingual web directory
recognized for its popularity and open-content rich-
ness. DMOZ was selected as the experimental frame-
work to establish a real-world scenario, offering a
publicly accessible and widely recognized repository
for result comparisons against baseline measures.
The Scraper module compiles URLs for down-
load, using a “text-only” parameter to acquire only
textual components of each document. The com-
plete DMOZ repository subset is used based on web-
scraping policies and link prevalence. It is crucial to
map DMOZ categories on WordNet synset and defi-
nitions and if a corresponding mapping exists for all
categories, because we have to obtain a compliant rep-
resentation of the synsets and the categories that we
have in the dataset. In Table 2, we provide a simple
association between the category extracted and the re-
spective WordNet synset tag and the description of it.
We use English documents for our experiments, au-
tomatically handled with the help of a Python library
porting of the Google algorithm for language detec-
tion and for the scraping of the pages (Danilak, 2017;
Hajba, 2018). Out of a total corpus comprising 12,120
documents, 10,910 were allocated for creating topic
modeling models. The remaining 1,210 documents
were designated as test sets for the comprehensive
evaluation of the entire system. A practical test set
for the system requires documents that undergo tex-
tual and visual analyses. Specifically, efforts were di-
rected towards selecting random documents from the
web directory, ensuring they possess a substantial tex-
tual component and a minimum of three images. A
DOM Parser algorithm was used to identify and re-
tain a “valid” multimedia document aligned with the
system’s objectives.
GenCrawl: A Generative Multimedia Focused Crawler for Web Pages Classification
95

Table 2: Categories with WordNet Synsets and Definitions.
Category Synset Definition Offset
Arts art.n.01 The products of human creativity; works of art collectively 2,743,547
Business commercial enterprise
.n.02
The activity of providing goods and services involving fi-
nancial and commercial and industrial aspects
1,094,725
Computers computer.n.01 A machine for performing calculations automatically 3,082,979
Games game.n.01 A contest with rules to determine a winner 455,599
Health health.n.01 A healthy state of well-being free from disease 14,447,908
News news.n.01 Information about recent and important events 6,642,138
Science science.n.01 A particular branch of scientific knowledge 5,999,797
Shopping shopping.n.01 Searching for or buying goods or services 81,836
Society society.n.01 An extended social group having a distinctive cultural and
economic organization
7,966,140
Sports sport.n.01 An active diversion requiring physical exertion and compe-
tition
523,513
4.2 Textual Topic Detection
The process of annotating the DMOZ category shown
in Table 2 makes the classification using the selected
algorithms easier for comparison of the resulted topic
detection task because they return no information
about the DMOZ category but only about the num-
ber of topics that represent the main topic of the text
corpus of the web page.
To ensure a comprehensive evaluation of our sys-
tem’s performance, we compare two established ref-
erence algorithms widely employed in topic detection
research: Latent Semantic Analysis (LSA) and Latent
Dirichlet Allocation (LDA). Latent Semantic Analy-
sis (LSA) (Landauer et al., 1998) employs a vectorial
representation approach to capture the essence of a
document using the bag-of-words model. Meanwhile,
Latent Dirichlet Allocation (LDA) (Blei et al., 2003)
is a text-mining model rooted in statistical method-
ologies. By benchmarking our system against these
reference algorithms, we aim to provide a robust as-
sessment of its efficacy in comparison to established
techniques in the realm of topic detection.
The three selected strategy performance for tex-
tual topic detection, compared with SBERT, are pre-
sented in Figure 4 and summarized in Table 3. The
Table 3: Accuracy score detail for textual topic detection.
Algorithm Accuracy Num. Correct
LSA 0.1 117
LDA 0.34 407
SBERT 0.53 620
SBERT model yielded the most favorable results re-
garding accuracy for textual topic detection, followed
by LDA, while the LSA algorithm demonstrated com-
paratively lower performance. This discrepancy may
be attributed to the outcomes being contingent on
the feasibility of mapping DMOZ categories onto the
concepts within the proposed ontology, as made in
Table 2. Notably, in the case of LSA, the dimin-
ished accuracy appears linked to challenges associat-
ing specific topics generated by the model with their
corresponding WordNet synsets. It can be postulated
that SBERT exhibits superior generalization in con-
cept recognition and is adept at mitigating noise in-
herent in specific datasets. Algorithm 1 shows the
pseudo-code of the procedure adopted for detecting
the textual topic from the full text of the web page
using SBERT.
Algorithm 1: Detect Topics.
1: function PREPROCESS TEXT(text)
2: return Tokenize, lemmatize, and remove stop
words from text
3: end function
4: function DETECT TOPICS(text, num clusters)
5: processed text ← PREPROCESS TEXT(text)
6: embeddings ← Generate SBERT embeddings for
processed text
7: clusters ← Apply k-means clustering with
num clusters to embeddings
8: for each cluster do
9: synset map ← Map words to WordNet synsets in
processed text
10: top synset ← Identify most common synset in
synset map
11: Associate top synset with the cluster as the repre-
sentative topic
end
12: return (detected topics, associated synsets)
13: end function
4.3 Visual Topic Detection
The task of visual topic detection harnesses multi-
media components, particularly images, to identify a
document’s primary topic, enhancing the overall per-
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
96

formance of our framework. Our methodology takes
a comprehensive approach to visual topic detection,
employing three distinct configurations based on dif-
ferent feature extraction methods. This thorough ex-
ploration allows us to understand the strengths and
limitations of each method.
PHOG (Bosch et al., 2007) is a global feature rep-
resentation method that ensures comprehensive fea-
ture extraction and accuracy through dense grid com-
putation and local contrast normalization with over-
lapping regions, making it suitable for precise visual
analysis tasks.
SIFT (Lowe, 2004) is a robust local feature extrac-
tion technique that identifies key points of interest in
gray-scale images, providing invariant descriptors for
translation, rotation, and scaling, and is widely used
in computer vision tasks.
VGG19 (Simonyan and Zisserman, 2014) is a
deep convolutional neural network (CNN) designed
for image classification tasks, with 19 layers, includ-
ing convolutional and fully connected layers. It ex-
tracts visual content representation from intermediate
layers, particularly the last max pooling layer, yield-
ing deep features suitable for topic detection.
Evaluation of the three configurations based on
utilized features (presented in Figure 4 and summa-
rized in Table 4) reveals VGG19 features exhibit the
highest accuracy, followed by PHOG and SIFT. These
results align with expectations, as VGG19 and PHOG
are promising candidates for precise feature match-
ing, albeit with VGG19’s computational time trade-
off due to its high dimensionality.
PHOG’s superior accuracy over SIFT is attributed
to its global nature. However, it can also be inter-
preted as a local feature due to its processing of im-
ages at various scales and resolutions. Selection of
the optimal feature depends on a comprehensive anal-
ysis of the combined strategy and specific application
requirements.
Table 4: Accuracy score detail for visual topic detection.
Algorithm Accuracy Num. Correct
SIFT 0.29 471
PHOG 0.39 348
VGG19 0.82 1331
4.4 Text-to-Image Generation
In our proposed strategy, an additional step involves
generating multimedia data, prompting us to iden-
tify the most suitable model for this task. We eval-
uate three different models: StackGAN (Zhang et al.,
2017), AttnGAN (Xu et al., 2018), and Stable Diffu-
sion (Rombach et al., 2022). StackGAN employs a hi-
erarchical Generative Adversarial Network for multi-
stage image synthesis, progressively generating high-
resolution images. AttnGAN incorporates attention
mechanisms to enhance fine-grained image genera-
tion by focusing selectively on relevant regions. Sta-
ble Diffusion utilizes stable diffusion processes, con-
trolling the gradual evolution of generated images to
achieve high-quality synthesis. To assess fidelity in
generated images from textual descriptions, we con-
sider several metrics addressing different aspects of
image fidelity. The Fr
´
echet Inception Distance (FID)
quantifies dissimilarity between the distributions of
real and generated images, offering a comprehen-
sive assessment. Other metrics, such as R-precision,
Semantic Object Accuracy (SOA), and CLIP score,
evaluate image-text matching. R-precision measures
visual-semantic similarity by ranking retrieval results
based on image and text features. SOA assesses ob-
ject detection in images, providing class (SOA-C) and
image (SOA-I) averages to gauge model alignment
with textual descriptions. As highlighted in (Hinz
et al., 2020), not all metrics are equally suited for
evaluating generative models. Metrics like FID, SOA,
and CLIP score better align with human visual assess-
ment than IS and R-precision. Thus, we focus on the
FID score, SOA metric, and CLIP score as crucial
metrics for evaluating the generative model’s perfor-
mance. In Table 5, we have reported some metrics
comparisons obtained during the evaluation process.
The results show that Stable Diffusion is the model
that best generates visually correct images that repre-
sent the textual prompt used for generation purposes.
4.5 Combined Topic Detection
The objective of the combined topic detection is
to allocate a singular score based on weights as-
signed to individual textual and visual topic detec-
tion classifiers. The integration employs SUM and
Ordered Weighted Averaging (OWA) operators. Var-
ious weight combinations have been systematically
tested for each proposed scheme, excluding the OWA
schema that employs OWA operators, providing a
fuzzy logic approach. This comprehensive evaluation
aims to identify optimal weight configurations con-
tributing to an effective and nuanced integration of
textual and visual topic detection outputs.
We decided to define four types of combination
between the textual and visual strategy: the A com-
bination pertains to the tests detailed earlier, wherein
visual topic detection demonstrated superior perfor-
mance. The B combination represents an average of
computed scores, while the C combination assigns
GenCrawl: A Generative Multimedia Focused Crawler for Web Pages Classification
97

Figure 4: Accuracy comparison between the selected methods of textual topic detection (on the left) and visual topic detection
(on the right).
Table 5: Metrics value comparison between GAN generative models and Stable Diffusion.
Model FID CLIP SOA-C SOA-I R-precision
StackGAN (Zhang et al., 2017) 12.50 23.18 25.88 39.01 68.40
AttnGAN (Xu et al., 2018) 9.14 28.39 31.70 47.78 83.79
Stable Diffusion (Rombach et al., 2022) 7.31 32.03 33.27 51.81 92.53
greater importance to textual topic detection. Addi-
tionally, a combination was tested using OWA oper-
ators, denoted as the D scheme, with a weight vec-
tor ⃗w = (0.65, 0.35). This diverse set of combina-
tions aims to thoroughly explore the interplay be-
tween textual and visual topic detection and iden-
tify optimal configurations for comprehensive evalu-
ation. The testing procedure encompasses 36 combi-
nations, employing schemes outlined in Table 6. In
Table 6: Weight configuration mapping for combined topic
detection.
Combination Text topic
weight
Image topic
weight
A 0.4 0.6
B 0.5 0.5
C 0.6 0.4
D 0.65 0.35
Figure 5, all results in classification accuracy among
the described combinations are presented. Combina-
tions incorporating visual topic detection and a fuzzy
logic approach (combination D) are the most effec-
tive, particularly with deep feature-based schemas
achieving high accuracy. However, textual topic de-
tection schemes exhibit lower or similar accuracy.
Combinations with equal weights for both classifiers
can sometimes yield suboptimal classification accu-
racy.
Visual classifiers with high accuracy significantly
contribute to topic detection due to their enhanced
discriminatory capabilities, particularly when images
better represent specific concepts. However, terms
with multiple meanings introduce uncertainty in the
task. Despite their high computational demands, the
SBERT model approach and VGG19 network-based
visual topic detection yield the best results. The cate-
gorization process is an offline task prioritizing accu-
racy within the specific domain of interest.
Table 7 only summarizes the most significant ex-
periments. The decision to choose the strategy for ex-
tracting topics from text and images and the method
of combination used for classification was based on
minimizing noise introduced by the image generation
step, ensuring its inclusion as a variable in the evalu-
ation step.
5 CONCLUSIONS AND FUTURE
WORKS
In this work, we have proposed an innovative way to
classify web documents using a generative approach
and combined topic detection algorithm. We have un-
veiled promising directions for advancing the capa-
bilities of an ontology-driven focused crawler. Ex-
ploring its extension to diverse conceptual domains,
accompanied by a comparative assessment across var-
ious ontologies and knowledge bases, offers valuable
insights into its adaptability and performance in cap-
turing domain-specific information. Additionally, the
investigation into alternative deep learning architec-
tures, encompassing attention mechanisms, genera-
tive models, and self-supervised learning, holds the
potential for augmenting the efficacy of image classi-
fication and retrieval tasks.
By evaluating different strategies and weighting
schemes for merging textual and visual classifica-
tions, we have refined the model’s performance. This
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
98

Figure 5: Accuracy performance among all the combinations selected, with all the visual and textual topic detection strategies
descripted.
Table 7: Most relevant test details for combined topic detection.
Combination Accuracy Num.
Correct
Combination Accuracy Num.
Correct
Combination Accuracy Num.
Correct
LSA-SIFT-A 0.29 348 LDA-SIFT-A 0.29 348 SBERT-SIFT-A 0.29 348
LSA-SIFT-B 0.08 100 LDA-SIFT-B 0.12 147 SBERT-SIFT-B 0.09 108
LSA-SIFT-C 0.10 117 LDA-SIFT-C 0.34 408 SBERT-SIFT-C 0.32 389
LSA-SIFT-D 0.33 402 LDA-SIFT-D 0.50 609 SBERT-SIFT-D 0.52 629
LSA-PHOG-A 0.39 471 LDA-PHOG-A 0.39 471 SBERT-PHOG-A 0.39 471
LSA-PHOG-B 0.09 108 LDA-PHOG-B 0.14 171 SBERT-PHOG-B 0.12 149
LSA-PHOG-C 0.10 117 LDA-PHOG-C 0.34 408 SBERT-PHOG-C 0.32 389
LSA-PHOG-D 0.40 480 LDA-PHOG-D 0.58 708 SBERT-PHOG-D 0.59 711
LSA-VGG19-A 0.44 539 LDA-VGG19-A 0.70 846 SBERT-VGG19-A 0.70 846
LSA-VGG19-B 0.09 111 LDA-VGG19-B 0.24 288 SBERT-VGG19-B 0.22 270
LSA-VGG19-C 0.10 117 LDA-VGG19-C 0.34 408 SBERT-VGG19-C 0.32 389
LSA-VGG19-D 0.70 852 LDA-VGG19-D 0.80 966 SBERT-VGG19-D 0.80 965
process has also enhanced our understanding of the
robustness and sensitivity of these strategies under
various conditions. Moreover, the expansion of the
focused crawler framework to include multimedia
content such as audio, video, and 3D models suggests
a comprehensive approach to web page analysis. This
expansion has the potential to significantly enhance
our understanding of multimedia-rich online content,
thereby improving the overall web document classifi-
cation process.
ACKNOWLEDGEMENTS
This research has been partially supported by the
Spoke 9 - Digital Society & Smart Cities - Cen-
tro Nazionale di Ricerca in High-Performance-
Computing, Big Data and Quantum Computing,
funded by European Union - NextGenerationEU -
CUP: E63C22000980007. We also acknowledge
financial support from the PNRR MUR project
PE0000013-FAIR.
GenCrawl: A Generative Multimedia Focused Crawler for Web Pages Classification
99

REFERENCES
Ahmed, Z. and Singh, H. (2019). Text extraction and clus-
tering for multimedia: A review on techniques and
challenges. In 2019 International Conference on Dig-
itization (ICD), pages 38–43. IEEE.
Benfenati, D., Montanaro, M., Rinaldi, A. M., Russo, C.,
and Tommasino, C. (2023). Using focused crawlers
with obfuscation techniques in the audio retrieval do-
main. In International Conference on Management of
Digital, pages 3–17. Springer.
Bergman, M. K. (2001). White paper: the deep web: sur-
facing hidden value. Journal of electronic publishing,
7(1).
Bhatt, D., Vyas, D. A., and Pandya, S. (2015). Focused web
crawler. algorithms, 5:18.
Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent
dirichlet allocation. Journal of machine Learning re-
search, 3(Jan):993–1022.
Bosch, A., Zisserman, A., and Munoz, X. (2007). Repre-
senting shape with a spatial pyramid kernel. In Pro-
ceedings of the 6th ACM international conference on
Image and video retrieval, pages 401–408.
Chakrabarti, S., Van den Berg, M., and Dom, B. (1999).
Focused crawling: a new approach to topic-specific
web resource discovery. Computer networks, 31(11-
16):1623–1640.
Cheng, H., Liu, S., Sun, W., and Sun, Q. (2023). A neural
topic modeling study integrating sbert and data aug-
mentation. Applied Sciences, 13(7).
Clinchant, S., Ah-Pine, J., and Csurka, G. (2011). Semantic
combination of textual and visual information in mul-
timedia retrieval. In Proceedings of the 1st ACM in-
ternational conference on multimedia retrieval, pages
1–8.
Danilak, M. (2017). Langdetect 1.0. 7. Python Package
Index.
Farag, M. M., Lee, S., and Fox, E. A. (2018). Focused
crawler for events. International Journal on Digital
Libraries, 19:3–19.
Fatima, N., Faheem, M., and Dar, M. Z. N. (2023). Op-
timized focused crawling for web page classification.
In 2023 International Conference on Energy, Power,
Environment, Control, and Computing (ICEPECC),
pages 1–6.
Fern
`
andez-Ca
˜
nellas, D., Marco Rimmek, J., Espadaler, J.,
Garolera, B., Barja, A., Codina, M., Sastre, M., Giro-i
Nieto, X., Riveiro, J. C., and Bou-Balust, E. (2020).
Enhancing online knowledge graph population with
semantic knowledge. In The Semantic Web–ISWC
2020: 19th International Semantic Web Conference,
Athens, Greece, November 2–6, 2020, Proceedings,
Part I 19, pages 183–200. Springer.
Fu, T., Abbasi, A., and Chen, H. (2010). A focused crawler
for dark web forums. Journal of the American Society
for Information Science and Technology, 61(6):1213–
1231.
Hajba, G. L. (2018). Website scraping with python. Berke-
ley: Apress.
Hassan, T., Cruz, C., and Bertaux, A. (2017). Ontology-
based approach for unsupervised and adaptive focused
crawling. In Proceedings of The International Work-
shop on Semantic Big Data, pages 1–6.
Hinz, T., Heinrich, S., and Wermter, S. (2020). Semantic
object accuracy for generative text-to-image synthe-
sis. IEEE transactions on pattern analysis and ma-
chine intelligence, 44(3):1552–1565.
K, N. T., S, C., G, B., Dharani, C., and Karishma, M. S.
(2023). Comparative analysis of various web crawler
algorithms.
Kittler, J., Hatef, M., Duin, R. P., and Matas, J. (1998). On
combining classifiers. IEEE transactions on pattern
analysis and machine intelligence, 20(3):226–239.
Kumar, N. and Aggarwal, D. (2023). Learning-based
focused web crawler. IETE Journal of Research,
69(4):2037–2045.
Kunder, M. d. (2018). The size of the world wide web (the
internet). Pobrano z: http://www. world widewebsize.
com/(19.01. 2017).
Landauer, T. K., Foltz, P. W., and Laham, D. (1998). An in-
troduction to latent semantic analysis. Discourse pro-
cesses, 25(2-3):259–284.
Liu, J., Li, X., Zhang, Q., and Zhong, G. (2022). A
novel focused crawler combining web space evolu-
tion and domain ontology. Knowledge-based systems,
243:108495.
Lowe, G. (2004). Sift-the scale invariant feature transform.
Int. J, 2(91-110):2.
Mary, J. D. P. N. R., Balasubramanian, S., and Raj, R.
S. P. (2022). An enhanced focused web crawler for
biomedical topics using attention enhanced siamese
long short term memory networks. Brazilian Archives
of Biology and Technology, 64:e21210163.
Mohandes, M., Deriche, M., and Aliyu, S. O. (2018). Clas-
sifiers combination techniques: A comprehensive re-
view. IEEE Access, 6:19626–19639.
Pant, G. and Srinivasan, P. (2005). Learning to crawl: Com-
paring classification schemes. ACM Transactions on
Information Systems (TOIS), 23(4):430–462.
Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G.,
Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark,
J., et al. (2021). Learning transferable visual models
from natural language supervision. In International
conference on machine learning, pages 8748–8763.
PMLR.
Ridnik, T., Ben-Baruch, E., Noy, A., and Zelnik-Manor,
L. (2021). Imagenet-21k pretraining for the masses.
arXiv preprint arXiv:2104.10972.
Rinaldi, A. M. (2014). Using multimedia ontologies for au-
tomatic image annotation and classification. In 2014
IEEE International Congress on Big Data, pages 242–
249. IEEE.
Rinaldi, A. M. and Russo, C. (2021). Using a multimedia
semantic graph for web document visualization and
summarization. Multimedia Tools and Applications,
80(3):3885–3925.
Rinaldi, A. M., Russo, C., and Tommasino, C. (2021a).
A semantic approach for document classification us-
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
100

ing deep neural networks and multimedia knowledge
graph. Expert Systems with Applications, 169:114320.
Rinaldi, A. M., Russo, C., and Tommasino, C. (2021b). Vi-
sual query posing in multimedia web document re-
trieval. In 2021 IEEE 15th International Confer-
ence on Semantic Computing (ICSC), pages 415–420.
IEEE.
Rinaldi, A. M., Russo, C., and Tommasino, C. (2021c).
Web document categorization using knowledge graph
and semantic textual topic detection. In Computa-
tional Science and Its Applications–ICCSA 2021: 21st
International Conference, Cagliari, Italy, September
13–16, 2021, Proceedings, Part III 21 , pages 40–51.
Springer.
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and
Ommer, B. (2022). High-resolution image synthesis
with latent diffusion models. In Proceedings of the
IEEE/CVF conference on computer vision and pattern
recognition, pages 10684–10695.
Russo, C., Madani, K., and Rinaldi, A. M. (2020). An unsu-
pervised approach for knowledge construction applied
to personal robots. IEEE Transactions on Cognitive
and Developmental Systems, 13(1):6–15.
Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J.,
Denton, E. L., Ghasemipour, K., Gontijo Lopes, R.,
Karagol Ayan, B., Salimans, T., et al. (2022). Photo-
realistic text-to-image diffusion models with deep lan-
guage understanding. Advances in Neural Information
Processing Systems, 35:36479–36494.
Shrivastava, G. K., Pateriya, R. K., and Kaushik, P. (2023).
An efficient focused crawler using lstm-cnn based
deep learning. International Journal of System Assur-
ance Engineering and Management, 14(1):391–407.
Simonyan, K. and Zisserman, A. (2014). Very deep con-
volutional networks for large-scale image recognition.
arXiv preprint arXiv:1409.1556.
Sood, G. (2016). Parsed DMOZ data.
Tchakounte, F., Ngnintedem, J. C. T., Damakoa, I., Ah-
madou, F., and Fotso, F. A. K. (2022). Crawl-
shing: A focused crawler for fetching phishing con-
tents based on graph isomorphism. Journal of King
Saud University-Computer and Information Sciences,
34(10):8888–8898.
Wu, H. and Hou, D. (2023). A focused event crawler with
temporal intent. Applied Sciences, 13(7).
Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang,
X., and He, X. (2018). Attngan: Fine-grained text
to image generation with attentional generative ad-
versarial networks. In Proceedings of the IEEE con-
ference on computer vision and pattern recognition,
pages 1316–1324.
Yager, R. R. and Kacprzyk, J. (2012). The ordered
weighted averaging operators: theory and applica-
tions. Springer Science & Business Media.
Yan, W. and Pan, L. (2018). Designing focused crawler
based on improved genetic algorithm. In 2018
Tenth International Conference on Advanced Compu-
tational Intelligence (ICACI), pages 319–323. IEEE.
Zhang, H., Xu, T., Li, H., Zhang, S., Wang, X., Huang, X.,
and Metaxas, D. N. (2017). Stackgan: Text to photo-
realistic image synthesis with stacked generative ad-
versarial networks. In Proceedings of the IEEE inter-
national conference on computer vision, pages 5907–
5915.
GenCrawl: A Generative Multimedia Focused Crawler for Web Pages Classification
101