Computer Science and Information Systems 20(4):1459–1481https://doi.org/10.

2298/CSIS230115042T

Deep Learning-based Sentiment Classification in

Amharic using Multi-lingual Datasets⋆

Senait Gebremichael Tesfagergish1 and Robertas Damaševičius1 and

Jurgita Kapočiūtė-Dzikienė2
1
Department of Software Engineering,
Kaunas University of Technology, Kaunas 51368, Lithuania
[email protected]
[email protected]
2
Department of Applied Informatics,
Vytautas Magnus University, Kaunas 44404, Lithuania
[email protected]

Abstract The analysis of emotions expressed in natural language text,

also known as sentiment analysis, is a key application of natural language
processing (NLP). It involves assigning a positive, negative (sometimes
also neutral) value to opinions expressed in various contexts such as social
media, news, blogs, etc. Despite its importance, sentiment analysis for
under-researched languages like Amharic has not received much attention
in NLP yet due to the scarcity of resources required to train such methods.
This paper examines various deep learning methods such as CNN, LSTM,
FFNN, BiLSTM, and transformers, as well as memory-based methods like
cosine similarity, to perform sentiment classification using the word or
sentence embedding techniques. This research includes training and com-
paring mono-lingual or cross-lingual models using social media messages in
Amharic on Twitter. The study concludes that the lack of training data in
the target language is not a significant issue since the training data 1) can
be machine translated from other languages using machine translation as
a data augmentation technique [33], or 2) cross-lingual models can cap-
ture the semantics of the target language, even when trained on another
language (e.g., English). Finally, the FFNN classifier, which combined
the sentence transformer and the cosine similarity method, proved to be
the best option for both 3-class and 2-class sentiment classification tasks,
achieving 62.0% and 82.2% accuracy, respectively.
Keywords: sentiment analysis, monolingual vs. cross-lingual approaches,
deep learning, sentence transformers, Amharic.

1. Introduction
The origin of Sentiment Analysis dates back to the 1950s when it was initially ap-
plied to written paper documents, becoming a vital topic in the NLP field with the
emergence of the Internet and electronic texts (especially non-normative texts).
sentiment analysis is a process of analyzing text to detect its author’s overall posi-
tive, negative, mixed, and neutral sentiment toward the discussed topic. However,
opinions are usually subjective expressions, texts are full of hidden meanings and
sarcasm. Due to all these factors, the sentiment analysis problem is still compli-
cated even for such widely used and resource-rich languages as English.
⋆ An extended version of the paper presented at the ICT Innovations 2022 conference.
1460 Senait Gebremichael Tesfagergish et al.

The need for analyzing text and identifying their sentiments relies on the tech-
nological era we live in today. Everything is shifting online and online comments
and reviews from the end users affect the decision taken by stakeholders in dif-
ferent domains [50]. News with a generally favorable tone has been linked to a
significant price increase. Negative news, on the other hand, may be connected to
a price drop with longer-term consequences. In marketing, the analysis of news
articles can help evaluating online reputation of business companies and brands
[52]. In the entertainment industry, customer reviews and comments are used for
decision-making for other potential buyers of the products [63]. Similarly, pro-
ducers use it to improve their service quality and outline a plan for their coming
products or services. In politics, it helps authorities to make decisions based on
the overall sentiment from the population surveys [16], or manage crisis commu-
nication [7]. A dark side of social networks is that they can be used to criticize
government officials [17], spread hate speech [3], homophoby [25], racism [32], and
conspiracies [23,56,55], aiming to influence events in the real world.
Due to ambiguities in each language and our human understanding, there is
no single solution that could work for all languages. Each language is different
and difficult in its own way, therefore requires adaptation. The identification and
processing of morphological features of a specific language are required for real-life
natural language processing (NLP) tasks [13]. Under-researched languages like
Amharic [21] could not benefit from the application and tools already developed
for the resource-rich languages like English [34]. It is due to its morphological
complexity and unavailability of enough data for solving the sentiment analysis
[39] task. Innovative artificial intelligence (AI) methods such as ensemble learn-
ing [42,29], deep learning models [15] for learning high-dimensional representations
(word embeddings) [35], which can be combined with heuristic optimization meth-
ods [5], are helping under-resourced languages to pass the hardships of collecting
and preprocessing large datasets, instead, they provide a deep insight into the
available data features to make the classification more efficient [24]. Recently,
multi-lingual approaches that can deal with numerous languages at the same time
were proposed to alleviate the problem of scarcity of data for sentiment analysis
in low-resourced languages such as Bengali [57], Serbian [19], Tamil [51], Urdu
[31] and others [48]. However, multilingual models often encounter issues with
highly imbalanced training data across the supported languages. As a consequence,
the effectiveness of these multilingual models for different languages also varies:
e.g., the well-supported English language demonstrates superiority in performance
while resource-scarce languages may suffer from poor or even unacceptable per-
formance.
The aim of this work is to address sentiment analysis for Amharic by benefiting
from 1) datasets that are available for other languages; 2) state-of-the-art multi-
lingual and cross-lingual solutions mainly focused on deep learning and transformer
models [59]. The paper is an extended version of conference paper [54].
The main novelty and contribution of this study is as follows:

– State-of-the-art sentence transformer embedding model (that projects sen-

tences into semantic space) has rarely been used as a sentence vectorization
technique for Amharic sentiment classification (see our previous work [53]).
 Deep Learning-based Sentiment Classification in Amharic... 1461

– We explore multiple approaches: 1) classical machine learning techniques (such

as cosine similarity and K-Nearest Neighbor (KNN)) and 2) traditional deep
learning approaches (such as Feed Forward Neural Network (FFNN), Con-
volutional Neural Network (CNN), Long Short-Term Memory (LSTM), Bidi-
rectional LSTM (BiLSTM)) a applied on the top of Word2Vec as word em-
beddings; 3) hybrid methods connecting the sentence transformer model with
cosine similarity and KNN.
– The lack of Amharic data problem was solved with the help of data machine
translation when translating from English. English is the resource-rich lan-
guage that allowed us to choose datasets in the domain of short texts on
general topics.
– Monolingual (training and testing on the same language) and cross-lingual
(training on one language, testing on another language) solutions compared.
– In the control experiment, we machine-translated the English data into 8 other
languages and performed similar experiments to investigate the impact of the
machine translation quality on the sentiment analysis task.

This paper is structured as follows. Related works are described in Section

2. The dataset used for this experiment is presented in Section 3. Analysis of
vectorization, classification models, and optimization techniques are discussed in
Section 4. Section 5 explains the experiment and its results. Section 6 concludes
with a discussion and conclusion about the overall objectives and achievements of
this research and future works.

2. Related works

Semitic languages like Arabic, Amharic, and Hebrew are widely spoken languages
by over 250 million people in the east, north Africa, and the Middle east. Semitic
languages exhibit unique morphological processes challenging syntactic construc-
tion and various other phenomena that are less prevalent in other natural languages
[64]. Amharic, despite being the second biggest language in the Semitic language
with 27 million native speakers and the official language of Ethiopia 100 million
population), is one of the low-resourced languages and lacks the availability of
resources for electronic data and basic tools for Natural language processing appli-
cations. We choose Amharic intentionally, as a good example of a rather complex,
low-resource language. Hence, our further theoretical research work analysis on
sentiment analysis will also consider these factors.
In this overview, we skip all outdated rule- and dictionary-based approaches,
focusing on the sentiment analysis problem as a supervised text classification prob-
lem by following the current trend in the sentiment analysis community. E.g., the
popular Papers with code portal [37] contains 1047 research papers of authors com-
peting to achieve better sentiment analysis accuracy on 42 benchmark datasets.
The variety of their tested methods covers a huge range of different approaches:
traditional machine learning, traditional deep learning to state-of-the-art trans-
former models. However, these competitions make clear that the transformer
models achieve the highest classification accuracy. Despite the majority of these
1462 Senait Gebremichael Tesfagergish et al.

papers summarizing the research done on the English language, it demonstrates

what to aim for and what might work for other languages.
The SemEval competitions also attract many researchers from all over the world
to compete when solving various sentiment analysis problems: i.e, in SemEval-2019
(311 teams tried to detect emotion classes) [12]; in SemEval-2020 (the 3-class sen-
timent analysis problem with the code-switching for Hinglish and Spanglish was
addressed by 61 and 28 teams, respectively) [44]; in SemEval-2022 (structural
3-class sentiment analysis problem was solved by 32 teams for Norwegian, Cata-
lan, Basque, Spanish, and English languages) [10]. If in 2019 traditional machine
learning approaches Naive Bayes, Logistic Regression, and SVM were still “on the
table”, achieving comparative results to traditional deep learning approaches (as,
e.g., BiLSTM) [43,30], graph convolutional networks [62], attention networks [22]
and 3D-CNN [58], transformer models (e.g., BERT, XLM, Roberta etc.) become
popular in 2020 and dominant in 2022 [26]. Consequently, it motivates us to inves-
tigate transformer models for our sentiment analysis problems. The success of the
pre-trained transformer models (which are later integrated into the classification
framework) highly depends on how well they support the target language (i.e., how
large and comprehensive the training corpora of the target language were used).
Hence, this factor cannot be controlled by us, therefore next to the transformer
models, we are planning to overview and test other (more stable) approaches.
Machine learning (ML) approaches such as Support Vector Machine (SVM),
Logistic Regression (LR), and Naïve Bayes (NB) were used to solve various NLP
tasks for a long time [49]. Amharic sentiment analysis study [45] used NB with
unigram, bigram, and hybrid variants as features. The research was conducted on
600 posts labeled to two classes. The authors managed to get their highest result
at 44% using the bigram feature. Multi-lingual Twitter sentiment analysis in [6]
presented 95% accuracy using the Bag-of-words vector and SVM classifier in En-
glish, Telugu, and Hindi. Naïve Bayes achieves the highest precision performance
in [8] of the Catalan language 2-class sentiment classification of 50,000 tweets,
which is +3% of the Neural network precision. Multi-class sentiment analysis in
the Russian and Kazakh languages presented in [38] proposes their best model for
this classification are Linear Regression, Decision Tree and Random Forest with
74%, 64%, and 70% accuracy respectively on the Russian texts.
Deep learning is a branch of machine learning which aims to model high-level
abstraction in data. This is done using model architectures that have complex
structures or those composed of multiple nonlinear transformations [24]. Many
studies are conducted using deep learning for Amharic sentiment analysis (see
Table 1). Since Arabic shares many similar characteristics with Amharic in terms
of morphology a few research using deep learning methods are also described in
Table 1. Sentiment analysis in Arabic catches the attention of many researchers
as it has a bigger number of speakers with different dialects all over the world and
plenty of datasets are available for conducting such research. In [40], a systematic
review from year January 2000 until June 2020 was conducted to analyze the status
of deep learning for Arabic subjective sentiment analysis tasks. The authors’
findings described that 45% of the selected papers conducted their experiment
using the CNN and RNN (LSTM) methods.
 Deep Learning-based Sentiment Classification in Amharic... 1463

Table 1. Related works using deep learning techniques

Ref. Corpus Language Classification Al- Embedding and Accuracy
gorithm Features
[2] 8,400 tweets (positive, Amharic Flair Graphical embed- 60.51%
negative, and neutral) ding
[3] 1,602 reviews Amharic Deep learning TF-IDF vectoriza- 90.1%
tion
[17] 6,652 samples (positive Amharic BERT Fine-tuned BERT 95%
and negative)
[4] 15,100 (positive and Arabic CNN-LSTM, Fast Text Embed- 90.75%
negative) SVM ding
[5] 2,026, positive (628) Arabic BILSTM Not mentioned 92.61%
and negative (1,398)

Summarizing, the sentiment analysis task for Amharic has been conducted
using different traditional machine learning approaches (SVM, multinomial NB,
Maximum Entropy applied on the top of bag-of-words, Decision Tree) and deep
learning methods. As for all languages, the recent research for Amharic is focused
on deep learning methods because they outperform the traditional machine learn-
ing approaches. However, our goal is to conduct accuracy-oriented comprehensive
comparative research, therefore we will test various Deep Learning methods, from
traditional to transformer models.
Cross-lingual solutions for the sentiment analysis problems are the salvation for
the low-resourced languages [11,1,28,18]. Their aim is to learn a universal classifier
which can be applied to languages with limited labeled data [2], which is exactly
what we have in sentiment analysis problems [6]. The cross-lingual approaches in
sentiment analysis usually vary from the early solutions based on machine transla-
tion to cross-lingual embeddings and multi-BERT pre-trained models [41]. English
– Arabic cross-lingual sentiment analysis presented in [2] concludes that regardless
of the artificial noise added by the machine translation they managed to achieve
the best result of 66.05% in the Electronics domain with the BLUE score of 0.209.
Another study [1] tested the performance of cross-lingual sentiment analysis with-
out good translation from English to Chinese and Spanish language. Authors
explained that in their experiment they observed that sentiment is preserved ac-
curately even if the translation is not accurate, and this inexpensive approach
maintains fine-grained sentiment information between languages.
To our best knowledge, the sentiment analysis problem for Amharic has never
been solved with cross-lingual approaches [4]. In advance, it is difficult to guess
which solution 1) machine-translation-based (not knowing how much the qual-
ity of machine translation can affect the classification result), or 2) cross-lingual
transformers (not knowing how well they support Amharic and their semantic re-
lations with other languages) can be the best. Besides, the machine translation
will help us not only in the cross-lingual settings, but in general when creating the
sentiment analysis dataset we lack for Amharic.
1464 Senait Gebremichael Tesfagergish et al.

3. Datasets

Since we formulate sentiment analysis as the supervised text classification problem,

we need the labeled data, but the selection of Amharic as our research object limits
our choices. To overcome this obstacle, we have decided to use:

1. The Ethiopic Twitter Dataset for Amharic (ETD-AM) dataset [60] which is
probably the only publicly available sentiment analysis dataset for Amharic.
It was introduced by Yimam et al. after being collected from Twitter and
annotated with the Amharic Sentiment Annotator Bot (ASAB) [61]. ETD-
AM stores only tweet ids and their sentiments, therefore for retrieving raw
tweets via the Twitter API, the tweepy python library was used. The re-
trieved original dataset consisted of around 8.6K tweets mapped to 3 (pos-
itive/negative/neutral) classes. Some tweets could not be retrieved via API
calls, resulting in a very small number of samples for the neutral class, this
class was omitted in our experiments. Hence, our sentiment analysis prob-
lem became a 2-class classification problem and the distribution of samples
between these classes can be found in Figure 1.
2. Tweet_Eval [9] the dataset which was borrowed from English. It is an En-
glish dataset containing tweets and adjusted for seven heterogeneous tasks,
namely, irony detection, hate speech detection, offensive language identifica-
tion, stance detection, emoji prediction, emotion recognition, and sentiment
analysis. Thus, we used this dataset for our sentiment analysis problem. Its
original version consisted of around 60K texts from social media, was noisy
(full of spelling mistakes, slang phrases, multi-lingual words, etc.), and needed
pre-processing. This step was utilized to eliminate unnecessary content and
convert it into useful information for the sentiment analysis task. The origi-
nal dataset is a non-normative data resource consisting of a non-Geez script;
therefore – emojis, web links, non-Latin letters, and non-English words were
removed. During the tokenization, the texts were split into tokens with the
Tokenizer from the Python Keras library. The final version of this dataset
used in our sentiment analysis experiments is presented in Figure 1.

Creating a model for a sentiment classification task depends on many factors.

Apart from the parameters of the selected classification model, the quality and
quantity of the dataset used for the training phase have a great impact on the
performance of the trained model. A larger dataset with good quality data will
train a better accurate model. In the case of 2-class sentiment analysis, the dataset
available was small and we needed to augment it with more translated data from
English. The English dataset from Twitter (Sentiment 140) [27] was translated to
Amharic and six other languages, and added to the original dataset. The added
data is balanced where the positive and negative class has 15,000 instances each.
The example of the tweet and its translations is presented in Table 2.
 Deep Learning-based Sentiment Classification in Amharic... 1465

Figure 1. Distribution of sentiments in ETD-AM and Tweet_Eval datasets

Table 2. Example of dataset tweet in seven languages.

Language Sentence
English Its to be expected from electing a Fascist Nazi
Amharic ፋሽስት ናዚን ከመምረጥ የሚጠበቅ ብቃት
Tigrinya ፋሺሽታዊ ናዚ ብምምራጽ ትጽቢት ክግበረሉ ኣለዎ
Lithuanian Jo reikia tikėtis išrinkus fašistinį nacį
Arabic ‫ﺍﻝﻑﺍﺵﻱﺓ ﺍﻝﻥﺍﺯﻱﺓ ﺍﻥﺕﺥﺍﺏ ﻡﻥ ﺍﻝﻡﺕﻭﻕﻉ ﻡﻥ‬
Czech Je třeba se očekávat od zvolení fašisty nacisty
German Es ist zu erwarten, einen faschistischen Nazi zu wählen
French Il faut attendre de l’élection d’un nazi fasciste

4. Methodology

Our methodology is summarized in Figure 2. It includes the following stages:

data cleaning, tokenization, vectorization, and sentiment classification, which are
described in more detail in the following subsections.

Figure 2. Workflow of methodology and Experiment

1466 Senait Gebremichael Tesfagergish et al.

4.1. Vectorization
In Section 2, we have discussed which methods are suitable for sentiment analysis;
this choice is also influenced by the specificity of the datasets (Section 3). However,
the supervised classifiers cannot be trained directly from raw texts. Thus, encoding
of texts into low-dimensional and dense numeric vectors plays an important role
in making these methods applicable. We tested the following embeddings:
– Word2Vec. These types of word embeddings are usually monolingual models
that map each distinct word into its stable fixed-size vector. These embed-
dings (skip-gram and CBOW) are trained to consider the word and its context
in the fixed window. The amount of data used to learn the embeddings have
a huge impact on their quality . The larger the amount of training data used,
the better mapping of the vector space is determined . However, these types
of embeddings suffer from word ambiguity problems: words written in the
same form, but having different semantical meanings will always be vectorized
alike. Unlike other resource-rich languages, the Amharic pre-trained Word2Vec
embeddings are not publicly available. Thus, we trained them using the same
Ethiopic Twitter Dataset for Amharic (ETD-AM) with 300 dimensions, a win-
dow size equal to 5 and with all other parameters with the default values. For
training word embeddings we have used python library.
– Sentence Transformers. These embeddings are state-of-the-art technology that
allows mapping whole sentences into fixed-size vectors [36]. The variety of
sentence transformers is rather large, we are most interested in being capable
to capture the semantics of sentences in relation to similar ones. . Moreover,
the most important requirement is that the model would support Amharic
and preferably be multi-lingual and able to benefit from other languages. The
pre-trained language-agnostic BERT sentence embedding model (LaBSE) [20]
seems the perfect solution to all of it, despite Amharic is not highly supported.

4.2. Classification Methods

We have investigated the following text classification methods:
– CNN was originally developed to image processing and classification but suc-
cessfully adapted to text classification. First, it seeks patterns in the input
by sliding through it with a 1D convolution filter. CNN considers patterns
of sequential words (in our case: word embeddings) thus making it more like
the keyword search-type approach. However, for some problems, CNN demon-
strates rather good performance. In our experiments we use the architecture
presented in Figure 3.
– LSTM, BiLSTM or hybrid. Recurrent Neural Networks (RNNs), LSTMs,
and BiLSTMs are adjusted to work with sequential data (in our case: word
embeddings). These models (Figure 4) use the output of the previous hid-
den state as an input for a current one. However, the RNNs suffer from the
vanishing gradient problem (especially having longer word sequences), it is
highly recommended to choose LSTMs or BiLSTMs instead as they have in-
put/forget/output gates to control this problem. Besides, BiLSTMs are also
 Deep Learning-based Sentiment Classification in Amharic... 1467

adjusted to learn from two directions at the same time (by processing text
from the start to its end and vice versa). Architectures of used LSTM and
BiLSTM approaches are presented in Figure 5.
– The hybrid models that blend different architectures of CNN with LSTM/BiLSTM
sometimes allow to achieve even better performance. We also tested such ar-
chitectures (Figure 6): the CNN model is responsible for the extraction of
features, and BiLSTM or LSTM is used for generalizing them[14].
– Cosine similarity with KNN. This memory-based approach is used with the
LaBSE sentence embeddings. After the LaBSE model projects sentences into
the semantical space, the cosine similarity measure can help determine similar-
ities between these sentences. The calculated value can be in the range [-1,1],
where 0 means that sentences are not similar; 1 - are the same; -1- opposite.
This memory-based method does not have any training phase: it simply stores
all vectorized training samples. Each new testing sample has to be compared
to all training samples and obtains the class of that training sample to which
the cosine similarity value is the largest.
– Feed Forward Neural Network (FFNN) is a simple classifier used when nonlin-
ear mapping is done between inputs and outputs. This method is chosen with
our sentence transformers because other deep neural network model cannot
be applied (LaBSE sentence vectors do not retain any patterns or sequential
characters of the input). The model (Figure 2) is trained to learn the rela-
tionship between sentences from the embeddings. When testing, it returns the
class of the most similar sentence in the training set.
– Bidirectional Encoder Representations from Transformer (BERT) is a transformer-
based technique for NLP pre-training developed by Google. Its generalization
capability is such that it can be easily adapted for various downstream NLP
tasks such as question answering, relation extraction, or sentiment analysis
[46]. Transformers are used to learn the relationship of words in the context.
BERT generates a language model using the encoder. The bidirectional en-
coder reads the sequence in both directions (left-to-right and right-to-left), so
the model is trained from the right and left sides of the target word. Because
the core architecture is trained on a huge text corpus, the parameters of the
architecture’s most internal levels remain fixed. The outermost layers, on the
other hand, adapt to the job and are where fine-tuning takes place. Sentiment
analysis is done by adding a final classification layer on top of the transformer
output for the [CLS] token. Currently, the Amharic pre-trained Bert model is
not available. Therefore, the English model was adapted.

5. Experiment and Results

The dataset described in Section 3 is vectorized by different embeddings explained
in subsection 4.1 and classified using the methods described in subsection 4.2.
Tensorflow, Keras, and PyTorch are used for the implementation of the methods.
Both used datasets ETD-AM (2 classes) and Tweet_Eval (3 classes) were split for
training and testing. Since we formulate our sentiment analysis tasks as the text
classification problem, the usual evaluation metrics such as accuracy, precision,
1468 Senait Gebremichael Tesfagergish et al.

Figure 3. Architecture of CNN model

Figure 4. Architecture of BiLSTM (right) and LSTM (left) model

recall, and f-score were applied. We have also calculated the majority baseline to
see if the accuracies achieved by methods are acceptable (if the achieved accuracy
is above the majority baseline the method is considered appropriate for the solv-
ing problem). Approaches (in which initial parameters are generated randomly
and later adjusted during training) were trained and evaluated several times to
calculate their average result. Table 3 summarizes the results for Amharic with
ETD-AM (2 classes) and Tweet_Eval (3 classes) datasets.

Table 3. Accuracies with ETD-AM (2 classes) and Tweet_Eval (3 classes)

datasets for Amharic.
Model ETD-AM Tweet_Eval
(2-Class) (3-Class)
CNN + Word2Vec 0.46 0.43
LSTM + Word2Vec 0.54 0.32
BILSTM + Word2Vec 0.62 0.39
CNN & BILSTM + Word2Vec 0.41 0.48
CNN & LSTM + Word2Vec 0.39 0.44
Cosine Similarity + Sentence Transformer + KNN 0.82 0.57
FFNN + Sentence Transformer 0.80 0.62
 Deep Learning-based Sentiment Classification in Amharic... 1469

Figure 5. Architecture of hybrid (CNN-BiLSTM & CNN-LSTM) models

Results of different classifiers for binary classification using original and aug-
mented datasets are given in Table 4. Addition of more data translated from
English improves the result of Word2Vec vectorization and deep learning meth-
ods (CNN, BILSTM, CNN-LSTM, CNN-BILSTM), while the best model with the
highest accuracy of 82% that uses the sentence transformer is downgraded by 5%.
A possible reason can be the domain of the texts as sentence transformers use the
semantics of the sentence for embedding.

Table 4. Accuracy of Original data and Accuracy with added translated data
Accuracy Accuracy
Model
(Original dataset) (Augmented dataset)
CNN + Word2Vec 0.46 0.64
LSTM + Word2Vec 0.54 0.49
BILSTM + Word2Vec 0.62 0.68
CNN & BILSTM +Word2Vec 0.41 0.69
CNN & LSTM + Word2Vec 0.39 0.70
Cosine Similarity + Sentence Transformer +KNN 0.82 0.77
FFNN + Sentence Transformer 0.80 0.76
1470 Senait Gebremichael Tesfagergish et al.

The determined best classification model for the 2-class is the Cosine Similarity
with the sentence transformer embedding. To improve the accuracy of this model,
we made a cluster of training sets that have more similarity with the testing
instance then voted for the training instance classes label in that cluster and assign
that class to the testing instance. In other words, we used the KNN classifier on top
of the Cosine Similarity, and in search of the best hyperparameter, we performed
the ablation study and presented the result in Table 5. The best accuracy was
achieved with 157 nearest neighbors.

Table 5. Accuracy of Cosine Similarity with the K-nearest neighborhoods

Hyperparameter value Accuracy of Sentence Transformer
(number of nearest neighbors (NN)) + Cosine Similarity + KNN model
1-NN 0.72
3-NN 0.78
31-NN 0.80
59-NN 0.81
157-NN 0.82

Finally, the Precision, Recall, F1-Score, and Accuracy of all the tested classi-
fiers are summarized in Table 6. The best result was achieved by the hybrid Cosine
Similarity + KNN model and Feed Forward Neural Network for the 2-Class and
3-Class respectively with the state-of-the-art Sentence Transformers embeddings.
The confusion matrix of the best models is also presented in Figure 8.

Table 6. Performance of all tested classification models

Model Classification Precision Recall F1-Score Accuracy
CNN + Word2Vec 2-class 0.65 0.57 0.60 0.64
CNN + Word2Vec 3-class 0.44 0.43 0.42 0.43
LSTM + Word2Vec 2-class 0.27 0.50 0.35 0.54
LSTM + Word2Vec 3-class 0.11 0.32 0.16 0.32
BILSTM + Word2Vec 2-class 0.66 0.60 0.62 0.68
BILSTM + Word2Vec 3-class 0.39 0.39 0.38 0.39
CNN & BILSTM + Word2Vec 2-class 0.72 0.62 0.67 0.69
CNN & BILSTM 3-class 0.48 0.48 0.46 0.48
CNN & LSTM + Word2Vec 2-class 0.69 0.73 0.71 0.70
CNN & LSTM 3-class 0.45 0.44 0.43 0.44
Cos. Similarity + Sentence 2-class 0.822 0.821 0.821 0.821
Transformer + KNN
Cos. Similarity + Sentence 3-class 0.52 0.53 0.52 0.53
Transformer + KNN
FFNN + Sentence Transformer 2-class 0.806 0.799 0.801 0.804
FFNN + Sentence Transformer 3-class 0.61 0.60 0.61 0.62
 Deep Learning-based Sentiment Classification in Amharic... 1471

For the 3-class experiment we used the translated data from English Tweets.
To compare the machine translation quality, we also translated the same data into
six other languages. The result is presented in Figure 6 and in Figure 7.

Figure 6. Different language accuracy for FFNN and Cosine Similarity with
Sentence Transformer embedding

For comparison, we performed an experiment with 3 different training sets and

the same Amharic testing sets. The training sets are:

1. English-language (gold-standard data) training set.

2. Machine-translated Amharic-only training set
3. Machine translated 7 languages + English in no.1 (Tigrinya, Amharic, Arabic,
Czech, German, French, Lithuanian) The result of this monolingual, Cross-
lingual, and all translated training sets are presented in Figure 7. The confu-
sion matrix for the same sets is also presented in Table 7.

In order to investigate if the translation of the dataset has an impact to change

the meaning of the sentence and degrade the quality of the dataset in Amharic we
annotated 100 translated Amharic sentences manually (see Table 8) and tested
using our first and second best methods with the two sets (1. The original senti-
ment from the English dataset and translated Amharic Sentences 2. The manually
annotated sentiment and translated Amharic Sentences). The comparison of some
examples of some English tweets and their translation to Amharic is presented in
9. Note the difference of sentiments between the English and Amharic languages.
1472 Senait Gebremichael Tesfagergish et al.

Figure 7. Accuracy of different training set and Amharic Testing sets for 3-class

Table 7. Confusion matrices of Cosine similarity Vs FFNN for cross-lingual,

mono-lingual and multi-lingual training.
Training-testing mode COS + ST + KNN FFNN + ST

Cross-Lingual
(English-Amharic)

Cross-Lingual
(All languages–Amharic)

Mono-lingual
(Amharic-Amharic)
 Deep Learning-based Sentiment Classification in Amharic... 1473

Figure 8. Confusion matrix of best models using Cosine Similarity and FFNN
with Sentence Transformer for 2-class and 3-class respectively

Table 8. Accuracy with original and human annotated datasets for Amharic.
Original Sentiment Amharic Sentiment
Model (From Original ( Human annotated
English Dataset) when data is translated)
FFNN 0.57 0.55
COS + Sentence Transformer + KNN 0.86 0.63

6. Discussion
We have solved 2-class (positive/negative) and 3-class (positive/negative/neutral)
sentiment classification problems for Amharic. We have investigated a wide range
of classification approaches: traditional Deep Learning (LSTM, BiLSTM, CNN-
LSTM, CNN-BiLSTM applied on top of word vectors); sentence transformer mod-
els with FFNN as the classifier or memory-based learning (Cosine + KNN ) Due
to the scarcity of dataset in Amharic, we added English translated dataset to the
original ETD-AM Amharic dataset for the 2-class classification while we used only
the translated English dataset for the 3-class. The experimental investigation of
different vectorization and classification techniques revealed that the most accurate
approach is the sentence transformers with Cosine Similarity + KNN or FFNN for
the 2-class or 3-class sentiment analysis problems respectively. The used LaBSE
sentence transformer model vectorizes sentences as a whole (without focusing on
separate words or their order) compared to Word2Vec word embeddings. There
are several reasons why the chosen sentence vectorizer outperforms the word-level
vectorizer. Firstly, Amharic has relatively free word order in sentences, therefore
sequences of concatenated word embeddings bring more variety to the training
data due to which the classifiers cannot make robust generalizations. Secondly,
the LaBSE model is the cross-lingual transformer itself as fine-tuned on the paral-
lel corpora of similar sentences for various languages. Despite Amharic is not very
highly supported in the LaBSE model (because of less training data for Amharic),
the cross-linguality mechanisms within LaBSE can compensate for it.
The use of sentence transformers (that accumulate the entire sentence by map-
ping it into the fixed-size vector) limits the options for the classifier. From the
1474 Senait Gebremichael Tesfagergish et al.

Table 9. Difference between sentiment annotations when sentences are

translated to Amharic. 0 - positive, 1 - negative, and 2 - neutral.
English Amharic
English Amharic Sentiment Sentiment
(Original Dataset) (Translated Dataset) (Original (Translated
Dataset) Dataset)
Wow first Hugo Chavez and now Fi- 0 1
del Castro Danny Glover Michael የዎፕስት ሂዩጎ ቻቬዝ እና አሁን
Moore Oliver Stone and Sean Penn ፊደል ካስትሮ ዳኒ ግሎቨር
are running out of heroes ሚካኤል ሙር ጽላት ኦቨር ኦቨር
ሲግናን ፔን ከጀግኖች እየራቁ ነው

The left has really gone Full retard 0 2

haven’t they እነርሱም ዘንድ ተውራት እያለች
በውስጧ ዘውታሪዎች ሲኾኑ የግራ
ጓዶች ናቸው ፡ ፡
The fact that mike pence think 0 2
ማይክ ፖምፒዮ ፣ ግብረ ሰዶማዊ
there’s a cure for being gay is ab-
መሆን ፈውስ ያስገኛል ብለው የሚ-
solutely the most ridiculous state-
ያስቡ ሰዎች መኖራቸው በሕይወቴ
ment I have ever heard in my life
ውስጥ ከሰማሁት ሁሉ እጅግ የሚ-
ያስደስት ነው
it’s free with insurance because of 0 2
Obamacare which Trump wants to በኦባማካሬ ምክንያት የመድን
repeal ዋስትና በማግኘቱ መለከት
ሊደግምለት ይፈልጋል
Thousands flee Raqqa as Turkish 1 0
Kurdish tensions threaten antiIS በሺዎች የሚቆጠሩ ሰዎች ራቃን
campaign ISIS ሸሽተው የቱርክ እንቅስቅሴዎች ፀረ
አይሲስን ዘመቻ አስጊ ሁኔታ ላይ
ሲጥሉ ነበር
You’re also young. 2 1
እርስዎም ወጣት ነዎት
I find myself humming the notes of 2 1
This Is Us sang a few episodes ago የዚህ መጽሐፍ ማስታወሻዎችን
Missing her angelic voice Love the እያዋዛሁ ሳለ ከጥቂት ክፍለ ጊዜ
show በፊት አንድ መላእክታዊ ድምፅዋን
ከፍ አድርጋ ትመለከተዋለች
hi love the tweet got stuff on social 2 1
security tweet ሰላም ትዊቱ በማኅበራዊ ደህንነት
ትዊተር ላይ ብዙ ነገር ተወጥቷል

possible options, we have tested the two most promising, but we could not deter-
mine the best one as the COS + KNN approach was better with ETD-AM, whereas
FFNN with the Tweet_Eval. However, the result is not surprising. The ETD-
AM dataset is the gold dataset that is originally prepared in Amharic; whereas
Tweet_Eval is only machine translated. The translated dataset contains ambigui-
ties and noise due to inaccurate translations of slang, abbreviations, etc., whereas
the original Amharic dataset is clean. However, the COS + KNN method is very
sensitive to noise: since for the testing instance, it can select the label of the most
similar training instance which is not a good representative of the class or even
 Deep Learning-based Sentiment Classification in Amharic... 1475

misleading. On the contrary, FFNN is a less risky option: instances of each class
are generalized therefore some amount of noise makes little impact.
There is a risk that the machine-translated version of the dataset is not suitable
for the solving sentiment analysis problem. To investigate the impact of the ma-
chine translation (both training and testing split) we ran the control experiment on
the original Tweet_Eval English dataset and the same dataset machine-translated
into 7 different languages (see Figure 7). The top line, i.e., the accuracy achieved
with the original English dataset is 72%. The machine translation quality and the
less support in the LaBSE model are the factors that degrade the performance
(with a 3% of accuracy drop for Czech and French; 10% for Amharic, and even
20% for Tigrinya). The results are not surprising, it perfectly correlates with how
well these languages are supported. For the less supported languages, the results
are expected to be lower, but the sentiment analysis task is still solvable.
In additional experiments we eliminated the machine translation step from the
training data preparations by training the model on the original English dataset
and testing on Amharic. Thus, in these cross-lingual experiments, we relied on the
robustness of the LaBSE model and its inner mechanisms to capture the semantics
between languages. It better worked with the FFNN classifier, but the accuracy
of 60% was still 1% lower compared to the monolingual model (trained and tested
only on Amharic). In the second experiment, we used the training data of all 8
languages (including Amharic); the trained model was again tested on Amharic.
This time it achieved 62% which is only 1% higher compared to the monolin-
gual setting. These results allow us to conclude that there is no big difference in
which approach to choose, but it opens more options. The machine translation
of the training dataset is not necessary: similar results can be achieved with the
cross-lingual models. However, if the usage of the machine translation tool is still
considered, it is worth translating the training dataset into better-supported lan-
guages (into which translating we can expect better quality and better support in
the sentence transformer models).

7. Conclusion

Sentiment analysis is a widely recognized NLP task that assigns sentiment labels,
including positive, negative, and neutral (sometimes mixed) to texts. Its successful
implementation can make significant contributions to resolving several societal
issues [47]. However, even for resource-rich languages like English, which possess
extensive data resources and accurate vectorization models, sentiment analysis
remains a relevant and challenging task due to issues such as sarcasm, hidden
meaning, and domain-specific language. In contrast, our study focuses on the
sentiment analysis problem for a resource-scarce language, using Amharic as a
main example.
We formulated the sentiment analysis problem as the supervised 2-class (posi-
tive/negative) and 3-class (positive/negative/neutral) classification problem, there-
fore it requires the training data. We experimented with ETD-AM and Tweet_Eval
datasets originally in Amharic and English, respectively.
1476 Senait Gebremichael Tesfagergish et al.

During our experimentation, we tested a wide range of techniques, including

the latest advances such as sentence transformer models, enabling us to attain
higher levels of accuracy. The best accuracy of 82.2% on the ETD-AM dataset
was achieved using the sentence transformer model in combination with the COS +
KNN classifier, which significantly surpassed the baseline. The sentiment analysis
problem with the ETD-AM dataset was also solved in [60], but due to different
experimental conditions, our results are not directly comparable.
We conducted experiments on the Tweet_Eval dataset under monolingual,
cross-lingual, and multi-lingual settings. For the monolingual experiments, both
the training and testing splits were machine-translated into Amharic. In cross-
lingual experiments, we used English texts for model training and machine-translated
Amharic texts for model testing. In the multi-lingual experiments, we used a mix
of machine-translated texts in eight languages (including Amharic) for the model
training, but only Amharic for the model testing. Across all the monolingual, cross-
lingual, and multi-lingual settings, the FFNN classifier applied on top of sentence
transformers performed the best, achieving the accuracy of 61%, 60%, and 62%,
respectively. However, neither of these settings was significantly superior to the
others. Through the control experiments that involved the machine-translated
Tweet_Eval dataset texts (8 different languages), we observed the correlation be-
tween the language support (machine translation quality, coverage level in the
sentence transformer model) and the sentiment analysis accuracy.
Despite achieving lower accuracy for Amharic compared to English, our results
are still significant and state-of-the-art for Amharic sentiment analysis. Besides,
our research is interesting as it addresses the sentiment analysis problem for a
resource-scarce language and determines the most effective solutions. These find-
ings can also be applied to other low-resource languages facing similar challenges.
We consider this to be an important research direction and we intend to continue
working on this topic in future research.

References

[1] Abdalla, M., Hirst, G.: Cross-lingual sentiment analysis without (good) trans-
lation. In: Eighth International Joint Conference on Natural Language Pro-
cessing (Volume 1). pp. 506–515 (2017)
[2] Al-Shabi, A., Adel, A., Omar, N., Al-Moslmi, T.: Cross-lingual sentiment
classification from english to arabic using machine translation. International
Journal of Advanced Computer Science and Applications 8(12) (2017)
[3] Aldjanabi, W., Dahou, A., Al-Qaness, M.A.A., Elaziz, M.A., Helmi, A.M.,
Damaševičius, R.: Arabic offensive and hate speech detection using a cross-
corpora multi-task learning model. Informatics 8(4) (2021)
[4] Alemu, Y.: Deep learning approach for amharic sentiment analysis (2018)
[5] Alhaj, Y.A., Dahou, A., Al-Qaness, M.A.A., Abualigah, L., Abbasi, A.A.,
Almaweri, N.A.O., Elaziz, M.A., Damaševičius, R.: A novel text classification
technique using improved particle swarm optimization: A case study of arabic
language. Future Internet 14(7) (2022)
 Deep Learning-based Sentiment Classification in Amharic... 1477

[6] Arun, K., Srinagesh, A.: Multilingual twitter sentiment analysis using ma-
chine learning. International Journal of Electrical and Computer Engineering
(IJECE) 10(6), 5992 (Dec 2020)
[7] Babić, K., Petrović, M., Beliga, S., Martinčić-Ipšić, S., Matešić, M., Meštro-
vić, A.: Characterisation of covid-19-related tweets in the croatian language:
Framework based on the cro-cov-csebert model. Applied Sciences 11(21)
(2021)
[8] Balaguer, P., Teixidó, I., Vilaplana, J., Mateo, J., Rius, J., Solsona, F.: Cat-
Sent: a catalan sentiment analysis website. Multimedia Tools and Applica-
tions 78(19), 28137–28155 (Jul 2019)
[9] Barbieri, F., Camacho-Collados, J., Espinosa Anke, L., Neves, L.: TweetEval:
Unified benchmark and comparative evaluation for tweet classification. In:
Findings of the Association for Computational Linguistics: EMNLP 2020. pp.
1644–1650. Association for Computational Linguistics, Online (Nov 2020)
[10] Barnes, J., Oberlaender, L., Troiano, E., Kutuzov, A., Buchmann, J., Agerri,
R., Øvrelid, L., Velldal, E.: SemEval 2022 task 10: Structured sentiment
analysis. In: 16th International Workshop on Semantic Evaluation (SemEval-
2022). pp. 1280–1295. Association for Computational Linguistics (Jul 2022)
[11] Bel, N., Koster, C.H.A., Villegas, M.: Cross-lingual text categorization. In:
Koch, T., Sølvberg, I.T. (eds.) Research and Advanced Technology for Digital
Libraries. pp. 126–139. Springer Berlin Heidelberg, Berlin, Heidelberg (2003)
[12] Chatterjee, A., Narahari, K.N., Joshi, M., Agrawal, P.: SemEval-2019 task
3: EmoContext contextual emotion detection in text. In: 13th International
Workshop on Semantic Evaluation. pp. 39–48 (2019)
[13] Choi, M., Shin, J., Kim, H.: Robust feature extraction method for automatic
sentiment classification of erroneous online customer reviews. International
Information Institute (Tokyo). Information 16(10), 7637 (2013)
[14] Dang, C.N., Moreno-García, M.N., la Prieta, F.D.: Hybrid deep learning
models for sentiment analysis. Complexity 2021, 1–16 (Aug 2021)
[15] Deng, L., Yu, D.: Deep learning: Methods and applications. Found. Trends
Signal Process. 7(3–4), 197–387 (jun 2014)
[16] Dhiman, A., Toshniwal, D.: Ai-based twitter framework for assessing the
involvement of government schemes in electoral campaigns. Expert Systems
with Applications 203 (2022)
[17] Dimova, G.: Who criticizes the government in the media? the symbolic power
model. Observatorio (OBS*) 6(1) (Mar 2012)
[18] Dong, X., de Melo, G.: A robust self-learning framework for cross-lingual
text classification. In: 2019 Conference on Empirical Methods in Natural
Language Processing and the 9th International Joint Conference on Natu-
ral Language Processing (EMNLP-IJCNLP). pp. 6306–6310. Association for
Computational Linguistics (2019)
[19] Draskovic, D., Zecevic, D., Nikolic, B.: Development of a multilingual model
for machine sentiment analysis in the serbian language. Mathematics 10(18)
(2022)
[20] Feng, F., Yang, Y., Cer, D., Arivazhagan, N., Wang, W.: Language-agnostic
BERT sentence embedding. In: 60th Annual Meeting of the Association for
Computational Linguistics (Volume 1). pp. 878–891. Association for Compu-
tational Linguistics (2022)
1478 Senait Gebremichael Tesfagergish et al.

[21] Gereme, F., Zhu, W., Ayall, T., Alemu, D.: Combating fake news in “low-
resource” languages: Amharic fake news detection accompanied by resource
crafting. Information 12(1), 20 (2021)
[22] Gunasekar, M., Thilagamani, S.: Improved feature representation using col-
laborative network for cross-domain sentiment analysis. Information Technol-
ogy and Control 52(1), 100–110 (2023)
[23] Kant, G., Wiebelt, L., Weisser, C., Kis-Katos, K., Luber, M., Säfken, B.: An
iterative topic model filtering framework for short and noisy user-generated
data: analyzing conspiracy theories on twitter. International Journal of Data
Science and Analytics (2022)
[24] Kapočiūtė-Dzikienė, J., Damaševičius, R., Woźniak, M.: Sentiment analysis
of lithuanian texts using traditional and deep learning approaches. Computers
8(1) (2019)
[25] Karayiğit, H., Akdagli, A., Aci, �.�.: Homophobic and hate speech detection
using multilingual-bert model on turkish social media. Information Technol-
ogy and Control 51(2), 356–375 (2022)
[26] Karayiğit, H., Akdagli, A., Acı, �.�.: Bert-based transfer learning model
for covid-19 sentiment analysis on turkish instagram comments. Information
Technology and Control 51(3), 409–428 (2022)
[27] KazAnova, �.�.: Sentiment140 dataset with 1.6 million tweets (Sep 2017),
https://www.kaggle.com/kazanova/sentiment140
[28] Keung, P., Lu, Y., Bhardwaj, V.: Adversarial learning with contextual em-
beddings for zero-resource cross-lingual classification and NER. In: 2019 Con-
ference on Empirical Methods in Natural Language Processing and the 9th
International Joint Conference on Natural Language Processing (EMNLP-
IJCNLP). pp. 1355–1360. Association for Computational Linguistics (Nov
2019)
[29] Khalid, M., Ashraf, I., Mehmood, A., Ullah, S., Ahmad, M., Choi, G.S.:
Gbsvm: Sentiment classification from unstructured reviews using ensemble
classifier. Applied Sciences 10(8) (2020)
[30] Khan, L., Amjad, A., Ashraf, N., Chang, H..: Multi-class sentiment analysis
of urdu text using multilingual bert. Scientific Reports 12(1) (2022)
[31] Khan, L., Amjad, A., Afaq, K.M., Chang, H.T.: Deep sentiment analysis
using CNN-LSTM architecture of english and roman urdu text shared in
social media. Applied Sciences 12(5), 2694 (Mar 2022)
[32] Lee, E., Rustam, F., Washington, P.B., Barakaz, F.E., Aljedaani, W., Ashraf,
I.: Racism detection by analyzing differential opinions through sentiment
analysis of tweets using stacked ensemble gcr-nn model. IEEE Access 10,
9717–9728 (2022)
[33] Liu, X., He, J., Liu, M., Yin, Z., Yin, L., Zheng, W.: A scenario-generic
neural machine translation data augmentation method. Electronics 12(10),
2320 (2023)
[34] Liu, X., Shi, T., Zhou, G., Liu, M., Yin, Z., Yin, L., Zheng, W.: Emotion
classification for short texts: an improved multi-label method. Humanities
and Social Sciences Communications 10(1) (2023)
[35] Ljajić, A., Marovac, U.: Improving sentiment analysis for twitter data by
handling negation rules in the serbian language. Computer Science and Infor-
mation Systems 16(1), 289–311 (2019)
 Deep Learning-based Sentiment Classification in Amharic... 1479

[36] Maas, A.L., Daly, R.E., Pham, P.T., Huang, D., Ng, A.Y., Potts, C.: Learn-
ing word vectors for sentiment analysis. In: 49th Annual Meeting of the As-
sociation for Computational Linguistics: Human Language Technologies. pp.
142–150. Association for Computational Linguistics (Jun 2011)
[37] Meta AI Research: Sentiment analysis, https://paperswithcode.com/
task/sentiment-analysis
[38] Mutanov, G., Karyukin, V., Mamykova, Z.: Multi-class sentiment analysis of
social media data with machine learning algorithms. Computers, Materials
and Continua 69(1), 913–930 (2021)
[39] Nandwani, P., Verma, R.: A review on sentiment analysis and emotion de-
tection from text. Social Network Analysis and Mining 11(1) (Aug 2021)
[40] Nassif, A.B., Elnagar, A., Shahin, I., Henno, S.: Deep learning for arabic
subjective sentiment analysis: Challenges and research opportunities. Applied
Soft Computing 98, 106836 (Jan 2021)
[41] Neshir, G., Atnafu, S., Rauber, A.: Bert fine-tuning for amharic sentiment
classification. In: Workshop RESOURCEFUL Co-Located with the Eighth
Swedish Language Technology Conference (SLTC), University of Gothenburg,
Gothenburg, Sweden. vol. 25 (2020)
[42] Neshir, G., Rauber, A., Atnafu, S.: Meta-learner for amharic sentiment clas-
sification. Applied Sciences 11(18) (2021)
[43] Ombabi, A.H., Ouarda, W., Alimi, A.M.: Deep learning CNN–LSTM frame-
work for arabic sentiment analysis using textual information shared in social
networks. Social Network Analysis and Mining 10(1) (Jul 2020)
[44] Patwa, P., Aguilar, G., Kar, S., Pandey, S., PYKL, S., Gambäck, B.,
Chakraborty, T., Solorio, T., Das, A.: SemEval-2020 task 9: Overview of
sentiment analysis of code-mixed tweets. In: Fourteenth Workshop on Se-
mantic Evaluation. pp. 774–790. International Committee for Computational
Linguistics, Barcelona (online) (Dec 2020)
[45] Philemon, W., Mulugeta, W.: A machine learning approach to multi-scale
sentiment analysis of amharic online posts. HiLCoE Journal of Computer
Science and Technology 2(2), 8 (2014)
[46] Reimers, N., Gurevych, I.: Sentence-BERT: Sentence embeddings using
Siamese BERT-networks. In: 2019 Conference on Empirical Methods in Nat-
ural Language Processing and the 9th International Joint Conference on Nat-
ural Language Processing (EMNLP-IJCNLP). pp. 3982–3992. Association for
Computational Linguistics (Nov 2019)
[47] Roth, S.: The great reset. restratification for lives, livelihoods, and the planet.
Technological Forecasting and Social Change 166, 120636 (May 2021)
[48] Sagnika, S., , Pattanaik, A., Mishra, B.S.P., Meher, S.K.: A review on multi-
lingual sentiment analysis by machine learning methods. Journal of Engineer-
ing Science and Technology Review 13(2), 154–166 (Apr 2020)
[49] Sarker, I.H.: Machine learning: Algorithms, real-world applications and re-
search directions. SN Computer Science 2(3) (Mar 2021)
[50] Shambour, Q.Y., Abu-Shareha, A.A., Abualhaj, M.M.: A hotel recommender
system based on multi-criteria collaborative filtering. Information Technology
and Control 51(2), 390–402 (2022)
1480 Senait Gebremichael Tesfagergish et al.

[51] Shanmugavadivel, K., Sathishkumar, V.E., Raja, S., Lingaiah, T.B., Nee-
lakandan, S., Subramanian, M.: Deep learning based sentiment analysis and
offensive language identification on multilingual code-mixed data. Scientific
Reports 12(1) (2022)
[52] Syllaidopoulos, I., Skraparlis, A., Ntalianis, K.: Evaluating corporate online
reputation through sentiment analysis of news articles: Threats, maliciousness
and real opinions. International Journal of Cultural Heritage 7, 8–22 (2022)
[53] Tesfagergish, S.G., Kapočiūtė-Dzikienė, J., Damaševičius, R.: Zero-shot emo-
tion detection for semi-supervised sentiment analysis using sentence trans-
formers and ensemble learning. Applied Sciences 12(17) (2022)
[54] Tesfagergish, S., Robertas Damaševičius, R., Kapočiūtė-Dzikienė, J.: Deep
learning-based sentiment classification of social network texts in amharic lan-
guage. In: ICT Innovations 2022. Reshaping the Future Towards a New Nor-
mal. Springer International Publishing (2023)
[55] Tuters, M., Willaert, T.: Deep state phobia: Narrative convergence in coro-
navirus conspiracism on instagram. Convergence: The International Journal
of Research into New Media Technologies 28(4), 1214–1238 (Aug 2022)
[56] Vergani, M., Martinez Arranz, A., Scrivens, R., Orellana, L.: Hate speech in
a telegram conspiracy channel during the first year of the covid-19 pandemic.
Social Media and Society 8(4) (2022)
[57] Wadud, M.A.H., Mridha, M.F., Shin, J., Nur, K., Saha, A.K.: Deep-bert:
Transfer learning for classifying multilingual offensive texts on social media.
Computer Systems Science and Engineering 44(2), 1775–1791 (2023)
[58] Xu, X., Zhu, G., Wu, H., Zhang, S., Li, K..: See-3d: Sentiment-driven
emotion-cause pair extraction based on 3d-cnn. Computer Science and In-
formation Systems 29(1), 77–93 (2023)
[59] Xu, Y., Cao, H., Du, W., Wang, W.: A survey of cross-lingual sentiment anal-
ysis: Methodologies, models and evaluations. Data Science and Engineering
7(3), 279–299 (Jun 2022)
[60] Yimam, S.M., Alemayehu, H.M., Ayele, A., Biemann, C.: Exploring Amharic
sentiment analysis from social media texts: Building annotation tools and
classification models. In: 28th International Conference on Computational
Linguistics. pp. 1048–1060. International Committee on Computational Lin-
guistics, Barcelona, Spain (Online) (Dec 2020)
[61] Yimam, S.M., Ayele, A.A., Biemann, C.: Analysis of the ethiopic twitter
dataset for abusive speech in amharic (2019)
[62] Zhang, S., Zhao, T., Wu, H., Zhu, G., Li, K.: Ts-gcn: Aspect-level sentiment
classification model for consumer reviews. Computer Science and Information
Systems 29(1), 117–136 (2023)
[63] Zinko, R., Patrick, A., Furner, C.P., Gaines, S., Kim, M.D., Negri, M., Orel-
lana, E., Torres, S., Villarreal, C.: Responding to negative electronic word
of mouth to improve purchase intention. Journal of Theoretical and Applied
Electronic Commerce Research 16(6), 1945–1959 (2021)
[64] Zitouni, I.: Natural Language Processing of Semitic Languages. Springer
(2014)
 Deep Learning-based Sentiment Classification in Amharic... 1481

Senait Gebremichael Tesfagergish has received hre MSc from Vytautas Mag-
nus University, Kaunas, Lithuania. Currently she is Ph.D. Student at Kaunas
University of Technology, Kaunas, Lithuania. Her topics of interest are Natural
language processing, Deep Learning and Artificial intelligence solutions. She is an
author of 7 research papers.

Robertas Damaševičius received the Ph.D. degree in informatics engineering

from the Kaunas University of Technology, Lithuania, in 2005. He is currently a
Professor with the Department of Applied Informatics, Vytautas Magnus Univer-
sity, Lithuania, and Department of Software Engineering, Kaunas University of
Technology. He is the author of more than 500 articles and a monograph pub-
lished by Springer. His research interests include sustainable software engineering,
human–computer interfaces, assisted living, and AI explainability. He is also the
Editor-in-Chief of the Information Technology and Control journal.

Jurgita Kapočiūtė-Dzikienė is a full professor at Vytautas Magnus University

and a language technology specialist at JSC Tilde IT. She has been working in
the field of AI since 2005 and her main research focuses on language technolo-
gies. Despite her hobby being methodologies/solutions for her native Lithuanian
language, she also enjoys working with other languages, especially if it is related
to multilingualism/cross-linguality problem-solving. Jurgita has worked on 20
projects. She is an editorial board/program committee member of several inter-
national journals/conferences and a co-author of 60 scientific publications.

Received: January 15, 2023; Accepted: June 25, 2023.