SCAN-NF: A CNN-based System for the Classification of Electronic
Invoices through Short-text Product Description
Diego Santos Kieckbusch
a
, Geraldo P. R. Filho
b
, Vinicius Di Oliveira
c
and Li Weigang
d
Departamento de Ci
ˆ
encia da Computac¸
˜
ao, Universidade de Brasilia, Brasilia, Brazil
Keywords:
Convolutional Neural Network, Invoice Classification, Short-text Classification, Tax Auditing, Few-word
Classification.
Abstract:
This research presents a Convolutional Neural Network (CNN) based system, named SCAN-NF, to classify
Consumer Electronic Invoices (NFC-e) based on product description. Due to how individual issuers submit
Consumer Electronic Invoices, processing these invoices is often a challenging task. Information reported is
often incomplete or presents mistakes. Before any meaningful processing over these invoices, it is necessary
to assess the product represented in each document. SCAN-NF is developed to identify correct products codes
in electronic invoices based on short-text product descriptions. Real data from Brazilian NFC-e and NF-e
documents related to B2B and retail transactions are used in experiments. Comparing base single model and
proposed ensemble model approaches, the evaluation results using recall, precision, and accuracy show the
satisfaction of the developed system.
1 INTRODUCTION
Invoices document the transaction of goods and ser-
vices and other business activities. For companies,
they are an important source of financial information
and a fundamental basis for controlling tax funds.
They are also the main source of information on
taxation for regulators. Intelligent processing of in-
voices allows for applications in the context of finan-
cial analysis, fraud detection (He et al., 2020), value
chain analysis, product tracking, and health hazard
alarms (Chang et al., 2020). Since 2010, all Brazil-
ian companies obligatorily report invoices to a cen-
tral financial agency, such as the State Treasury Of-
fice (SEFAZ). Similar measures have also been taken
in Italy (Bardelli et al., 2020) and China (Zhou et al.,
2019)(Yue et al., 2020).
The Brazilian Electronic invoice is a standardized
XML file. While fields are audited for fulfillment and
type, there are breaches for exploits and errors. One
fundamental vulnerability is on the reported product
code, called NCM. NCM is a standardized nomen-
clature for products and services in Mercosur. It is
a
https://orcid.org/0000-0002-9957-0059
b
https://orcid.org/0000-0001-6795-2768
c
https://orcid.org/0000-0002-1295-5221
d
https://orcid.org/0000-0003-1826-1850
used to define the correct taxation and if the product
is eligible for tax exemption. One could miss-classify
products to benefit from lower taxation. Brazil uti-
lizes two types of electronic invoices: Electronic In-
voice (NF-e), which records B2B transactions, and
Consumer Electronic Invoices (NFC-e) that records
retail transactions. Mandatory reports of the NFC-e
begun in 2017 and audition processes performed on
NF-e documents are not performed in NFC-e data.
Manual auditing of these invoices is expensive and
time-consuming, especially for NFC-e data due to a
larger number of issuers and low quality of reported
data. Since tax auditing is a fundamental activity for
the Treasury Office, autonomous or semi-autonomous
tools for processing large invoice datasets are of great
value.
Invoice text data differs in grammar and vocabu-
lary from regular language usage and can be seen as
short-text. Short-text can be defined by the follow-
ing characteristics (Enamoto et al., 2021): individ-
ual author contribution is small; grammar is gener-
ally informal and unstructured, sent and received in
real-time and in large quantity; imbalanced distribu-
tion of classes of interest; large scale data presents a
labelling bottleneck. Even when compared to other
short texts, invoice description is very brief, contain-
ing only a handful of words, often not forming a
complete sentence. This exacerbates the problem of
Kieckbusch, D., Filho, G., Di Oliveira, V. and Weigang, L.
SCAN-NF: A CNN-based System for the Classification of Electronic Invoices through Short-text Product Description.
DOI: 10.5220/0010715200003058
In Proceedings of the 17th International Conference on Web Information Systems and Technologies (WEBIST 2021), pages 501-508
ISBN: 978-989-758-536-4; ISSN: 2184-3252
Copyright
c
 2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
501

domain-specific vocabulary, abbreviations, and typos
as authors use their individual logic.
Related works on invoice classification have fo-
cused on the Chinese case. These solutions have
ranged from using hash trick for dealing with an un-
known number of features (Zhou et al., 2019)(Yue
et al., 2020), semantic expansion trough external
knowledge bases (Yue et al., 2020), classification of
paragraph embedding by k-nearest-neighbors (Tang
et al., 2019) to different artificial neural network ar-
chitectures (Yu et al., 2019)(Zhu et al., 2020). Seman-
tic expansion is prevalent not only on invoice clas-
sification but also on short-text classification (Wang
et al., 2017),(Naseem et al., 2020). These works are
not suited for the Brazilian case either due to language
differences or reliance on knowledge bases only avail-
able in English and Chinese (Grida et al., 2019).
There is a gap in the literature for models suited for
the classification of Brazilian Electronic Invoices.
This work presents a Convolutional Neural Net-
work (CNN) system, named SCAN-NF, for tax audi-
tors to identify suspicious invoices based on textual
product descriptions. We utilize the sentence classi-
fication architecture proposed by Kim (Kim, 2014)
as the basis for our model. Real data from Brazil-
ian Electronic Invoices (NF-e) and Consumer Elec-
tronic Invoices (NFC-e) are used in our experiments.
We compare single, and ensemble model approaches
on recall, precision, and accuracy on both datasets.
We then discuss performance and trade-offs between
approaches. Comparing single model with ensemble
model approaches, the evaluation metrics show the
satisfaction of the developed system. Our ensemble
approach achieved better precision on both datasets.
This article is organized as follows: In section 2,
we present related work on the invoice and short text
classification; in section 3, we describe the SCAN-
NF system and the architecture of the classification
model; in section 4, we present experimental setup on
a study case on real NF-e and NFC-e data. Results
of experiments are presented in section 5, and in the
final section, we present closing remarks and future
works.
2 RELATED WORK
In this section, we highlight other works related to
short text and invoice classification. Taking invoices
as an example of short text, short-text classification is
a broader area, and some solutions may not suit in-
voice classification. In contrast, works aimed at in-
voice classification may not utilize short text process-
ing techniques.
2.1 Traditional Methods
Traditional methods rely on bag-of-words represen-
tation and matrix factorization to create a represen-
tation for text processing. The low word count on
short text documents leads to common co-occurrence
of terms across the document-term matrix, which in-
validate matrix factorization methods.
Early works attempted to address this problem
by expanding available information through auxil-
iary databases. Document expansion seeks to sub-
stitute the representation of short text to represent a
set of related documents. In query-based expansion,
these documents are returned by using short text as
the input on a search engine (Sahami and Heilman,
2006)(Yih and Meek, 2007). Phan (Phan et al., 2008)
proposed a framework for short text classification that
used an external ”universal dataset” to discover a set
of hidden topics through Latent Semantic Analysis.
The problem with document expansion is that it in-
creases the computational cost to search and process
a more significant amount of data. This new data also
introduces noise to the model.
2.2 Neural based Methods
Neural-based methods represent short text as a se-
quence of vectors and utilize convolutional and recur-
rent neural networks to learn a suitable representation
for classification.
The architecture proposed by Kim (Kim, 2014)
serves as the basis for most CNN-based solutions.
Zhang (Zhang and LeCun, 2016) utilized a 12-layer
CNN to learn features from character embeddings.
Character-based representation does not rely on pre-
trained word embeddings and could be used in any
language. Wang (Wang et al., 2017) expanded the
model proposed by Kim (Kim, 2014) by utilizing
concept expansion and character level features. The
model used knowledge bases to return related con-
cepts and included them in the text before the em-
bedding layer. Knowledge bases included: YAGO,
Probase, FreeBase, and DBpedia. A character-based
CNN was used in parallel to the word concept CNN.
Representations learned by both networks were con-
catenated before the final fully connected layer.
Naseem (Naseem et al., 2020) proposed an ex-
panded meta-embedding approach for sentiment anal-
ysis of short-text that combined features provided by
word embeddings, part of speech tagging, and senti-
ment lexicons. The resulting compound vector was
fed to a Bi-LSTM with an attention network. The
rationale behind the choice for an expanded meta-
embedding is that language is a complex system, and
WEBIST 2021 - 17th International Conference on Web Information Systems and Technologies
502

each vector provides only a limited understanding of
the language.
2.3 Invoice Classification
Invoice classification techniques have ranged from
traditional count-based methods to neural-based ar-
chitectures. In 2017, Chinese invoice data was made
public for Chinese researchers, which motivated re-
search in the area. This leads to the prevalence of
works dealing with the Chinese invoice system.
Some works aimed to address the data sparsity
problem by utilizing hash trick for dimensionality re-
duction (Zhou et al., 2019)(Yue et al., 2020). Yue
(Yue et al., 2020) performed semantic expansion of
features through external knowledge bases before us-
ing the hash trick for dimensionality reduction. Tang
(Tang et al., 2019) utilized paragraph embedding to
create a reduced representation and then applied K-
NN classifier. Yu (Yu et al., 2019) utilized a paral-
lel RNN-CNN architecture, with the resulting vectors
being combined in a fully connected layer. Zhu (Zhu
et al., 2020) combined features selected through fil-
tering with representation learned through the LSTM
model.
Unlike most western languages, in which text is
expressed through words with white spaces as separa-
tors, text in Chinese is expressed without separators,
with no clear word boundary. Words are constructed
based on the context. Chinese invoice classification
words leaned towards RNN based architectures in a
way to mitigate errors produced in the word segmen-
tation step.
Chinese works aside, Paalman et al (Paalman
et al., 2019) worked on the reduction of feature space
through 2-step clustering. The first step was to reduce
the number of terms through filtering and then clus-
ter the distributed semantic vector provided by differ-
ent pre-trained word embeddings. This method was
compared to traditional representation schemes and
matrix factorization techniques. In the experiments,
simple term frequency and TF-IDF normalization per-
formed better than LDA and LSA.
2.4 Discussion of Related Work
Term count-based methods mainly address short-text
processing through filtering and knowledge expan-
sion. The problem with filtering is that there is infor-
mation loss in a context where information is already
poor. Semantic expansion is mainly done through
knowledge bases. Communication with knowledge
bases becomes the bottleneck of the system and are
unsuited for invoice processing due to the amount of
invoice data. Furthermore, knowledge bases may not
be available in languages other than English and Chi-
nese (Grida et al., 2019).
The limitation of pre-trained word embeddings
comes down to vocabulary coverage and word sense
(Faruqui et al., 2016). These are significant to invoice
classification. Words in invoices are often misspelled
and abbreviated. Also, taxpayers often mix words of
multiple languages depending on the kind of product
being reported. Finally, invoices possess little to no
context to disambiguate word sense.
Most invoice classification models did not utilize
ANN. Yu (Yu et al., 2019) was the only one to com-
bine both CNN and BiLSTM. However, CNN and
BiLSTM were used in parallel over different fields
of invoice data. Zhu (Zhu et al., 2020) combined
an LSTM network with traditional methods using fil-
tered features. While effective for the Chinese lan-
guage, these architectures do not suit the Brazilian
invoice model. We address these shortcomings by
proposing a CNN-based model that does not rely on
pre-trained word embedding and external knowledge
bases.
3 ARCHITECTURE OF SCAN-NF
In this section, we present an overview of the archi-
tecture of the SCAN-NF system and inner models,
Figure 1. The system works in two phases: a train-
ing phase and a prediction phase. During the training
phase, the system is fed audited data from the tax of-
fice server of SEFAZ to train a supervised model. Two
models are trained, one for the classification of NF-e
Documents and another for NFC-e Documents. After
training, these models are used on new data during the
prediction phase.
The system works as follows: Data is extracted
from the tax office server (label 1 in figure 1). Prod-
uct description and corresponding NCM code for each
product in each invoice are then extracted (label 2 in
figure 1). Text is then cleaned from irregularities (la-
bel 3 in figure 1). A training dataset is constructed
by balancing target classes samples and dropping du-
plicates (label 4 in figure 1). The training set is then
fed to a CNN model that learns to classify product de-
scriptions (label 5 in figure 1). Outputs at the training
phase of the system are used to validate models before
being put into production (label 6 in figure 1). During
the Prediction Phase, trained models are utilized to
classify new data. These datasets may be composed of
invoices issued by a suspected party or a large, broad
dataset used for exploratory analysis (label 7 in figure
1). Models trained in the training phase are then em-
SCAN-NF: A CNN-based System for the Classification of Electronic Invoices through Short-text Product Description
503

Figure 1: Architecture of SCAN-NF.
ployed for the task at hand (label 8 in figure 1). The
final output of the model is the classified set of prod-
ucts inputs (label 9 in figure 1). This set of classified
product transactions are then used in manual auditing
by tax auditors (label 10 in figure 1).
The system is intended to aid tax auditors in
the audition of invoices issued by already suspicious
parties to pinpoint inconsistencies and irregularities.
Currently, NFC-e documents are not audited due to
the amount of data, a large number of issuers, and the
nature of the data. Our solution helps auditors pin-
point inconsistencies in documents reported by an al-
ready suspicious party and allows for the automatic
processing of a larger amount of data. We hope that
this solution will improve the productivity of tax au-
ditors regarding NF-e processing and be the first step
towards NFC-e processing.
There are different possibilities for the classifica-
tion model used in the system. The sentence clas-
sification model proposed by Kim (Kim, 2014) can
be used as a single multi-label classification model.
However, due to the high number of possible NCM
codes and high amount of invoice data, we propose
an ensemble model built from binary classifiers. Bi-
nary classifiers trained on individual classes can be
pre-trained, stored, and then combined in multi-label
classifiers on demand. This allows individual mod-
els to be updated and added without the need of re-
training other models.
Figure 2 presents the flowchart used in single
models. The input layer takes the indexed representa-
tion of text. In the embedding layer, each word index
is replaced by the word vector representation. Input is
then reshaped to be fed to parallel channels of one-
dimensional convolutions layers. Each convolution
layer applies several filters of a given size to the en-
coded sentence. Max pooling is applied to the learned
filters to extract the most useful features. Outputs
of each channel are concatenated in a single vector
flattened and fed to a Fully connected layer that will
output the final classification. The categorical cross-
entropy calculates loss, and the soft-max function acts
as the activation function of the model.
Figure 2: Flowchart of SCAN-NF CNN single model.
Figure 3 present a simplified flowchart of the en-
semble model. The ensemble model is built from bi-
nary classifiers, each trained on a singular target class.
Each binary classifier is built on the flowchart pre-
sented in figure 2. In binary models, the loss is cal-
culated by the binary cross-entropy, and the sigmoid
function is used as the activation function for the last
layer. To offset the imbalance between classes in bi-
nary models, we set class weights to a rate of 180 to 1.
The ensemble model is built using previously trained
binary models. The output of each model is concate-
WEBIST 2021 - 17th International Conference on Web Information Systems and Technologies
504

nated and fed to a single fully connected layer that
performs multi-layer classification. The categorical
cross-entropy calculates loss for the ensemble model,
and the softmax function gives the activation function
of the model.
Figure 3: Flowchart of SCAN-NF CNN ensemble model.
3.1 NF-e and NFC-e
The NF-e is the Brazilian national electronic fis-
cal document, created to substitute physical invoices,
providing judicial validity to the transaction and real-
time tracking for the tax office (SEFAZ, 2015). It
contains detailed information about invoice identifi-
cation, issuer identification, recipient identification,
product, transportation, tax information, and total val-
ues. In our work, we utilize data present in product
transactions, namely product description and NCM
code. Data regarding issuer and recipient is kept hid-
den. NFC-e is a simplified version of the NFC-e
meant to be used in retail services.
There are validations rules for the NCM field in
the NF-e manual (SEFAZ, 2015). According to a
specialist working with tax audition and the schedule
published in the NF-e manual, while validation pro-
cedures for NF-e documents are implemented, these
validation procedures are not planned for NFC-e doc-
uments in the following years. This results in data of
poor quality.
4 PERFORMANCE EVALUATION
We conducted a study case based on real NFC-e and
NF-e documents from SE to validate our model. Data
were separated into training and test sets, and dif-
ferent models were trained. Models were validated
through cross-validation. Hyper-parameter optimiza-
tion was conducted based on the average performance
through all folders of cross-validations.
4.1 The Data
In our experiments, we utilized data provided by the
estate tax office (SEFAZ). Data provided included
both NFC-e and NF-e documents. NF-e data con-
sisted of invoices of cosmetics. NFC-e data consisted
of a larger dataset of products from multiple sectors.
We selected NCM codes present in the NF-e dataset
and created a curated dataset with balanced classes.
Due to disparity in market share, preserving product
frequency would bias the models toward larger issuers
and the most popular products. This could lead mod-
els to better classify invoices from large companies
or learn their representation as the norm. Our design
decision was to drop duplicate product descriptions
for each target class. Table 1 presents detailed infor-
mation on the number of samples used in the experi-
ment. While there is a significant vocabulary overlap
between NF-e and NFC-e documents regarding NF-e
data, NFC-e presents a much more vast vocabulary.
Table 1: Number of samples and datasets used in experi-
ments.
NF-E NFC-E
Number of raw
product samples
198882 99637515
Number of samples
in balanced dataset
36234 49536
Number of balanced
classes
18 18
Vocabulary
Size
3646 15312
Shared Terms 2342
4.2 Metrics
We evaluate models based on the following metrics:
accuracy, precision, recall, and top k Accuracy. Met-
rics are calculated based on the occurrence of True
Positives, True Negatives, False Negatives, and False
Positives.
Accuracy is given by the rate of correct predic-
tions overall predictions: (T P + T N)/(T P + T N +
FP + FN). Top k Accuracy represents how often
the correct answer will be in the top k outputs of the
model. Accuracy is useful for getting an overall idea
of model performance. In unbalanced datasets, recall
and precision can paint a better picture of how the
model behaves.
The recall represents the recovery rate of positive
samples and is given by T P/(T P + FN). Precision
evaluates how correct the set of retrieved samples is
and is given by T P/(T P + FP). We utilize the F1-
SCAN-NF: A CNN-based System for the Classification of Electronic Invoices through Short-text Product Description
505

score, the harmonic mean of precision and recall, to
get a balanced assessment of model performance on
imbalanced classification.
In our experiments, we first set up a CNN archi-
tecture. We defined hyper-parameters through opti-
mization using the hyper-opt library. Table 2 presents
the parameters and values used in optimization, final
parameters are highlighted in bold.
Table 2: Parameters used in optimization.
Parameter Values
Number of Filters on
1D Convolution #1
{50,100,200,300,
400,500,600}
1D Convolution
Kernel Size #1
{3,5,7,9}
Number of Filters on
1D Convolution #2
{50,100,200,300,
400,500,600}
1D Convolution
Kernel Size #2
{3,5,7,9}
Number of Filters on
1D Convolution #3
{50,100,200,300,
400,500,600}
1D Convolution
Kernel Size #3
{3,5,7,9}
Dropout rate [0, 0.29, 0.5]
Optimizer
{Adam, Adagrad,
Adadelta, Nadam}
5 RESULTS
In this section, we present the results of the experi-
ments. The goal of the experiment is to compare sin-
gle model and ensemble model approaches. The sin-
gle model is composed of a single CNN model trained
on multi-label classification. Ensemble model is com-
posed of a set of binary models. Each binary model
is trained on a distinct class in a binary classification
problem. The ensemble model takes the list of binary
models and is then fine-tuned as a multi-label classi-
fication problem. Callbacks are set to stop training
based on validation error loss. Singular models and
binary models were trained through 5 epochs, while
the fine tune of the ensemble model is done through
12 epochs. Each epoch took 4sec/10.000 samples to
be performed. In practice, the ensemble model takes
20 times longer to be trained than the single model
due to the training of binary models and fine-tune of
the ensemble model. Experiments were repeated 10
times.
5.1 NF-e Dataset
Figure 4 presents single and ensemble model perfor-
mance on the NF-e dataset. We present results side by
side. We can see that while model accuracy deviated
sightly, differences in recall and precision were more
evident. There is little spread for all metrics. The sin-
gle model performed slightly better on both accuracy
metrics. Both models presented accuracy above 0.85.
The most notable difference between models comes
from the trade-off between recall and precision. The
single model performed better on recall at the cost of
precision. The single model recall was 15% higher
than the ensemble model at the cost of 5% precision.
Figure 4: Performance of single and ensemble models on
NF-e dataset.
Individual class performance of the ensemble
model is shown in figure 5. Due to the unbalanced
nature of the problem, all classes presented high accu-
racy scores, scoring higher than 96%. Overall, there
was a balance between recall and precision. Of all
models, 15 had an F1-score higher than 0.8, and 7
had an F1-score above 0.9. This signalizes that some
classes are more challenging to predict than others,
and some classification models are less trustworthy.
5.2 NFC-e Dataset
Performance on NFC-e is represented in Figure 6. We
can see that the trade-off between recall and precision
between models also occurs, with the ensemble model
achieving lower recall and higher precision. There is a
drop in accuracy for both models, while top2 accuracy
remained the same. Individual binary model perfor-
mance on NFC-e dataset is shown on Figure 7. Of 18
classifiers, 3 presented F1-Scores lower than 0.6, 12
presented F1-scores above 0.7, and only 1 achieved
an F1-score higher than 0.9.
5.3 Comparison of Approaches
When comparing overall results between datasets,
it becomes clear that NFC-e product classification
WEBIST 2021 - 17th International Conference on Web Information Systems and Technologies
506

Figure 5: Individual Binary Model Performance on NF-e
dataset.
Figure 6: Performance of single and ensemble Models on
NFC-e dataset.
is a more complex problem than NF-e classifica-
tion. In both datasets, the worst-performing and best-
performing classes were the same. This indicates
that identifying certain products is more complex than
others. Performance on NF-e data was higher than
NFC-e data. This is in accord with the NFC-e docu-
ment characteristics.
We can see that there is a trade-off between re-
call and precision, with the ensemble model present-
ing higher precision at the cost of the recall. The
simple model approach is more suited for exploratory
data analysis due to higher recall, while the ensem-
ble model approach is more suited to the audition of
suspicious issuers due to higher precision. There are
also differences in the maintainability of approaches.
Figure 7: Performance of binary models on NFC-e dataset.
The ensemble approach allows individual models to
be updated without the need for all models to be up-
dated. This also impacts the system’s scalability, as
additional classes can be added to the model without
retraining the whole model at each addition.
Models consistently achieved around 95% top2
accuracy on both datasets. This means that models
can be used as recommendation systems for the clas-
sification of product descriptions. This is particularly
valuable for NFC-e documents, in which no homolo-
gation is currently done, and data is more varied. Rec-
ommendations can aid taxpayers in narrowing down
the NCM code given a general text description of the
product. In turn, this could lead to better reported
NFC-e data. Overall, models managed to map prod-
uct descriptions to the corresponding NCM code.
6 CONCLUSION AND FUTURE
WORK
In this work, we showed SCAN-NF, an invoice clas-
sification system based on product description for tax
auditing. We presented related work on short-text and
invoice classification and a set of desired properties
for invoice classification. We then presented SCAN-
NF, a solution for the modeled problem, and the archi-
tecture of the CNN model to power the solution. We
presented two possible configurations for the CNN
models: a single model based on established sentence
classification architecture and our proposed ensem-
SCAN-NF: A CNN-based System for the Classification of Electronic Invoices through Short-text Product Description
507

ble model. Both CNN configurations were validated
on datasets of NFC-e and NF-e documents. Our en-
semble approach presented higher precision on both
datasets. Overall we managed to present an invoice
classification system that can aid tax auditors in au-
diting a larger number of invoices and aid taxpayers
in providing the correct classification of products.
In future work, we will focus on transfer learn-
ing. We hope that the parameters obtained from pre-
training using better represented NF-e documents can
improve performance on the training of NFC-e data.
This would be of great value as manual auditing of
individual invoices is quite expensive. Our main fo-
cus will be using Natural Language Processing (NLP)
techniques such as pre-trained word embeddings and
transformers into our concerning research.
REFERENCES
Bardelli, C., Rondinelli, A., Vecchio, R., and Figini, S.
(2020). Automatic electronic invoice classification us-
ing machine learning models. Machine Learning and
Knowledge Extraction, 2(4):617–629.
Chang, W.-T., Yeh, Y.-P., Wu, H.-Y., Lin, Y.-F., Dinh, T. S.,
and Lian, I. (2020). An automated alarm system for
food safety by using electronic invoices. PLoS ONE,
15(1).
Enamoto, L., Weigang, L., and Filho, G. P. R. (2021).
Generic framework for multilingual short text catego-
rization using convolutional neural network. Multime-
dia Tools and Applications, 80.
Faruqui, M., Tsvetkov, Y., Rastogi, P., and Dyer, C. (2016).
Problems With Evaluation of Word Embeddings Us-
ing Word Similarity Tasks. pages 30–35.
Grida, M., Soliman, H., and Hassan, M. (2019). Short text
mining: State of the art and research opportunities.
Journal of Computer Science, 15(10):1450–1460.
He, Y., Wang, C., Li, N., and Zeng, Z. (2020). At-
tention and Memory-Augmented Networks for Dual-
View Sequential Learning. In Proceedings of the
ACM SIGKDD International Conference on Knowl-
edge Discovery and Data Mining, pages 125–134.
Kim, Y. (2014). Convolutional neural networks for sentence
classification. EMNLP 2014 - 2014 Conference on
Empirical Methods in Natural Language Processing,
Proceedings of the Conference, (2011):1746–1751.
Naseem, U., Razzak, I., Musial, K., and Imran, M. (2020).
Transformer based Deep Intelligent Contextual Em-
bedding for Twitter sentiment analysis. Future Gen-
eration Computer Systems, 113:58–69.
Paalman, J., Mullick, S., Zervanou, K., and Zhang, Y.
(2019). Term based semantic clusters for very short
text classification. In International Conference Recent
Advances in Natural Language Processing, RANLP,
volume 2019-Septe, pages 878–887.
Phan, X. H., Nguyen, L. M., and Horiguchi, S. (2008).
Learning to classify short and sparse text & web with
hidden topics from large-scale data collections. Pro-
ceeding of the 17th International Conference on World
Wide Web 2008, WWW’08, (January):91–99.
Sahami, M. and Heilman, T. D. (2006). A web-based ker-
nel function for measuring the similarity of short text
snippets. Proceedings of the 15th International Con-
ference on World Wide Web, pages 377–386.
SEFAZ (2015). Manual de Orientac¸
˜
ao do Contribuinte -
Padr
˜
oes T
´
ecnicos de Comunicac¸
˜
ao. ENCAT.
Tang, X., Zhu, Y., Hu, X., and Li, P. (2019). An integrated
classification model for massive short texts with few
words. In ACM International Conference Proceeding
Series, pages 14–20.
Wang, J., Wang, Z., Zhang, D., and Yan, J. (2017).
Combining knowledge with deep convolutional neu-
ral networks for short text classification. IJCAI In-
ternational Joint Conference on Artificial Intelligence,
pages 2915–2921.
Yih, W. T. and Meek, C. (2007). Improving similarity
measures for short segments of text. Proceedings
of the National Conference on Artificial Intelligence,
2:1489–1494.
Yu, J., Qiao, Y., Shu, N., Sun, K., Zhou, S., and Yang,
J. (2019). Neural Network Based Transaction Clas-
sification System for Chinese Transaction Behavior
Analysis. In Proceedings - 2019 IEEE International
Congress on Big Data, BigData Congress 2019 - Part
of the 2019 IEEE World Congress on Services, pages
64–71.
Yue, Y., Zhang, Y., Hu, X., and Li, P. (2020). Extremely
Short Chinese Text Classification Method Based on
Bidirectional Semantic Extension. In Journal of
Physics: Conference Series, volume 1437.
Zhang, X. and LeCun, Y. (2016). Text understanding from
scratch.
Zhou, M., Hu, X., Zhu, Y., and Li, P. (2019). A novel clas-
sification method for short texts with few words. In
Proceedings of 2019 IEEE 3rd Information Technol-
ogy, Networking, Electronic and Automation Control
Conference, ITNEC 2019, pages 861–865.
Zhu, Y., Li, Y., Yue, Y., Qiang, J., and Yuan, Y. (2020). A
Hybrid Classification Method via Character Embed-
ding in Chinese Short Text with Few Words. IEEE
Access, 8:92120–92128.
WEBIST 2021 - 17th International Conference on Web Information Systems and Technologies
508