Natarajan Meghanathan et al. (Eds) : DMAP, CSITA, ARIN, ISPR, AISO - 2019
pp. 55-68, 2019. © CS & IT-CSCP 2019 DOI: 10.5121/csit.2019.91606
AMHARIC-ARABIC NEURAL MACHINE
TRANSLATION
Ibrahim Gashaw and H L Shashirekha
Mangalore University, Department of Computer Science, Mangalagangotri,
Mangalore-574199
ABSTRACT
Many automatic translation works have been addressed between major European language
pairs, by taking advantage of large scale parallel corpora, but very few research works are
conducted on the Amharic-Arabic language pair due to its parallel data scarcity. Two Long
Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) based Neural Machine
Translation (NMT) models are developed using Attention-based Encoder-Decoder architecture
which is adapted from the open-source OpenNMT system. In order to perform the experiment, a
small parallel Quranic text corpus is constructed by modifying the existing monolingual Arabic
text and its equivalent translation of Amharic language text corpora available on Tanzile. LSTM
and GRU based NMT models and Google Translation system are compared and found that
LSTM based OpenNMT outperforms GRU based OpenNMT and Google Translation system,
with a BLEU score of 12%, 11%, and 6% respectively.
KEYWORDS
Amharic, Arabic, Neural Machine Translation, OpenNMT
1. INTRODUCTION
"Computational linguistics from a computational perspective is concerned with understanding
written and spoken language, and building artifacts that usually process and produce language,
either in bulk or in a dialogue setting." [1].
Machine Translation (MT), the task of translating texts from one natural language to another
natural language automatically, is an important application of Computational Linguistics (CL)
and Natural Language Processing (NLP). The overall process of invention, innovation, and
diffusion of technology related to language translation drive the increasing rate of the MT
industry rapidly [2]. The number of Language Service Provider (LSP) companies offering
varying degrees of translation, interpretation, localization, language, and social coaching
solutions are rising in accordance with the MT industry [2]. Today many applications such as
Google Translate and Microsoft Translator are available online for language translations.
There are different types of MT approaches, and many researchers have classified them in
different ways. Oladosu et al. [3] have classified MT into two main categories: single and hybrid
approaches. The single approach uses only one method to translate from one natural language to
another, and it includes rule-based, direct, corpus-based, and knowledge-based approaches to MT.
A hybrid approach is a combination of the statistical method and the rule-based approach, which
includes a word-based model, a phrase-based model, a syntax-based model, and forest-based
model. Single approaches achieve low performance because they are inconsistent and inflexible
for large scale applications and give a shallower representation of knowledge resulting in lower
56 Computer Science & Information Technology (CS & IT)
quality and fluency of the output. The hybrid approach combines the strength of two or more
approaches to improve the overall quality of MT [3].
Deep learning NMT approach is a recent approach of MT that produces high-quality translation
results based on a massive amount of aligned parallel text corpora in both the source and target
languages. Deep learning is part of a broader family of ML methods based on artificial neural
networks [4]. It allows computational models that are composed of multiple processing layers to
learn representations of data with various levels of abstraction. These methods have improved the
state-of-the-art research in language translation [5]. NMT is one of the deep learning end-to-end
learning approaches to MT that uses a large artificial neural network to predict the likelihood of a
sequence of words, typically modeling entire sentences in a single integrated model. The
advantage of this approach is that a single system can be trained directly on the source and target
text no longer requiring the pipeline of specialized systems used in statistical MT. Many
companies, such as Google, Facebook, and Microsoft, are already using NMT technology [6].
NMT has recently shown promising results on multiple language pairs. Nowadays, it is widely
used to solve translation problems for many language pairs. However, much of the research on
this area has focused on European languages despite these languages being very rich in resources.
MT is a very ambitious research task in NLP, and the demand for it is growing. Several MT
systems have been developed all over the world, particularly from English to other natural
languages, such as Arabic, Germany, Chinese, French, Hindi, Japanese, Spanish, and Urdu [7].
Though a limited amount of work has been done in different Ethiopian languages in the field of
NLP, the MT system for Amharic-Arabic language pair is still in its infancy due to lack of
parallel corpora. Therefore, it is essential to construct Amharic-Arabic parallel text corpora,
which is very much required to develop the Amharic to Arabic NMT system.
Amharic language is the national language of Ethiopia spoken by 26.9% of Ethiopia’s population
as mother tongue and spoken by many people in Israel, Egypt, and Sweden. Arabic is a Semitic
language spoken by 250 million people in 21 countries as the first language and serving as a
second language in some Islamic countries. Ethiopia is one of the nations, which have more than
33.3% of the population who follow Islam, and they use the Arabic language to teach religion and
for communication purposes. Both of these languages belong to the Semitic family of languages,
where the words are formed by modifying the root itself internally and not merely by the
concatenation of affixes to word roots [8].
NMT has many challenges, such as; domain mismatch, size of training data, rare words, long
sentences, word alignment, and beam search [9] depending on the language pairs. Some of these
challenges are addressed in this paper.
2. RELATED WORK
Different attempts have been made in the past to construct parallel text corpus from Amharic and
Arabic languages pair with the English language. Some of these works are as follows;
G.A Mekonnen, A. Nurnberger, and Bat [10], describe the acquisition, preprocessing,
segmentation, and alignment of Amharic-English parallel corpus that consists of 1,45,820
Amharic-English parallel sentences (segments) from various sources. This corpus is larger in size
than previously compiled Amharic-English Bilingual Corpus (AEBC), which is hosted by
European Language Resource Association with the size of 13,379 aligned segments (sentences)
and the Low Resource Languages for Emergent Incidents (LORELEI) developed by Strassel and
Tracey [11], that contains 60,884 segments.
Sakre et al., [12], presents a technique that aimed to construct an Arabic-English corpus
automatically through web mining. The system crawled the host using GNU Wget in order to
Computer Science & Information Technology (CS & IT) 57
obtain English and Arabic web pages and then created candidate parallel pairs of documents by
filtering them according to their similarity in the path, file name, creation date, and length.
Finally, the technique measured the parallelism similarity between the candidate pairs according
to the number of transnational tokens found between an English paragraph and its three Arabic
neighbor paragraphs. However, in this work, they did not test or compare different models of
statistical translation using the constructed parallel corpus.
Ahmad et al. [13] reported the construction of one million words English-Arabic Political Parallel
Corpus (EAPPC) that consists of 351 English and Arabic original documents and their
translations. The texts were meta-annotated, segmented, tokenized, English-Arabic aligned,
stemmed, and POS-tagged. Arabic and English data were extracted from the official website of
'His Majesty King Abdullah II' and from 'His book' and then reprocessed from metadata
annotation to alignment. They built a parallel concordancer, which consisted of two parts: the
application through which the end-user interacts with the corpus and the database which stores
the parallel corpus. Experiments carried out examined the translation strategies used in rendering
a culture-specific term, and results demonstrated the ease with which knowledge about translation
strategies can be gained from this parallel corpus.
Inoue et al. [14], describe the creation of Arabic-Japanese portion of a parallel corpus of
translated news articles, which is collected at Tokyo University of Foreign Studies (TUFS). Part
of the corpus is manually aligned at the sentence level for development and testing. The first
results of Arabic-Japanese phrase-based MT trained on this corpus reported a BLEU score of
11.48.
Alotaibi [15], described an ongoing project at the College of Languages and Translation, King
Saudi University, to compile a 10-million-word Arabic-English parallel corpus to be used as a
resource for translation training and language teaching. The corpus has been manually verified at
different stages, including translations, text segmentation, alignment, and file preparation, to
enhance its quality. The corpus is available in XML format and through a user-friendly web
interface, which includes a concordance that supports bilingual search queries and several
filtering options.
All the above-related works use different lexical resources like machine-readable bilingual
dictionaries or parallel corpora, for probability assessment or translation. However, when there is
a lack of such a lexical resource, an alternative approach should be available [16]. Nowadays,
NMT Models are widely used to solve various translation problems. Learning Phrase
Representations using Recurrent Neural Network (RNN) Encoder-Decoder for statistical MT [17]
benefits more natural language-related applications as it can capture the linguistic regularities in
multiple word level as well as phrase level. But it is limited to target phrases, instead of using a
phrase table.
Dzmitry et al. [18], extended NMT Encoder-Decoder that encodes a source sentence into a fixedlength
vector, which is used by a decoder to generate a translation. It automatically search for
relevant parts of a source sentence to predict a target word without having to form these parts like
a hard segment explicitly. Their method yielded good results on longer sentences, and the
alignment mechanisms are jointly trained towards a better log-probability of producing correct
translations that need high computational cost.
A. Almahairi et al. [19], proposed NMT for the task of Arabic translation in both directions
(Arabic-English and English-Arabic) and compared a Vanilla Attention-based NMT system
against a Vanilla Phrase-based system. Preprocessing Arabic texts can increase the performance
of the system, especially normalization, but the model consumes much time for training.
58 Computer Science & Information Technology (CS & IT)
3. TRANSLATION CHALLENGES OF AMHARIC AND ARABIC LANGUAGES
Amharic and Arabic Languages are characterized by complex, productive morphology, with a
basic word-formation mechanism, root-and-pattern. The root is a sequence of consonants, and the
pattern is a sequence of Vowels (V) and Consonants (C) with open slots in it. It combines with
the pattern through a process called interdigitating (intercalation): each letter of the root (radical)
fills a slot in the pattern. For example, the Amharic root s.b.r (sabr) denoting a notion of breaking,
combines with the pattern CVCC (the slots and vowels are denoted by C and V respectively)
[20].
In addition to the unique root-and-pattern morphology, they are characterized by a productive
system of more standard affixation processes. These include prefixes, suffixes, infixes, and
circumfixes, which are involved in both inflectional and derivational processes. Consider the
Arabic word " وسوف یكتوبنھ " (wasawf yaktwubunahu) and its English translation "and they will
write it". A possible analysis of these complex words defines the stem as "aktub" (write), with an
inflectional circumfix, "y-uwna", denoting third person masculine plural, an inflectional suffix,
"ha" (it), and two prefixes, "sa" (will) and "wa" (and).
Morphological analysis of words in a text is the first stage of most natural language applications.
Morphological processes define the shape of words. They are usually classified into two types of
processes [21];
1. A derivational process that deals with word-formation; such methods can create new
words from existing ones, potentially changing the category of the original word. For
example, from the Arabic root " كتب " (wrote), the following words are derived; " "الكآتیب
(the writer), " الكیتآب " (the book), “ المكتبة ” (the library), “ مكتبھ ” (library). The same is true in
Amharic also. For example, from Amharic root “ፃፈ” (he wrote), the following words are
derived; "ፀሀፊ" (writer), "መፅሀፍ" (the book), "ቤተመፅሃፍ" (the library).
2. Inflectional processes are usually highly productive, applying to most members of a
particular word class. For example, Amharic nouns inflect for number, so most nouns
occur in two forms, singular (which is considered in the citation form) and plural,
regularly obtained by adding the suffix "ዎች" to the base form. This process makes the
translation ambiguous. The considerable number of potential types of words and the
difficulty of handling out-of-lexicon items (in particular, proper names) combined with
prefix or suffix makes the computation very challenging. For example in the word
"aysäbramm" the prefix "ay" and the suffix "amm" (he doesn't break) are out-of-lexicon
items.
The main lexical challenge in building NLP systems for Amharic and Arabic languages is the
lack of machine-readable lexicons, which are vital resources. The absence of capitalization in
Arabic and Amharic languages makes it hard to identify proper nouns, titles, acronyms, and
abbreviations. Sentences in the Arabic language are usually long, and punctuation has no or little
effect on the interpretation of the text.
Standard preprocessing techniques such as capitalization, annotation, and normalization cannot be
performed on Amharic and Arabic languages due to issues of orthography. A single token in
these languages can be a sequence of more than one lexical item, and hence be associated with a
sequence of tags. For example, the Amharic word "አስፈረደችብኝ" ("asferedachibegn"), where
translated to English will be "a case she initiated against me was decided in her favor". The word
is built from the causative prefix "as" (causes), a perfect stem "ferede" (judged), a subject maker
clitics "achi" (she), a benefactive marker "b" (against) and the object pronoun "egn" (I).
Computer Science & Information Technology (CS & IT) 59
Contextual analysis is essential in both languages to understand the exact meaning of some
words. For example, in Amharic, the word "ገና" can have the meaning of "Christmas holiday" or
"waiting for something until it happens." Diacritics (vowels) are most of the time, omitted from
the Arabic text, which makes it hard to infer the word meaning and complex morphological rules
should be used to tokenize and parse the text.
The corpus of the Arabic language has a bias towards religious terminology as a relatively high
frequency of religious terms and phrases are found. Characters are sometimes stretched for
justified text, which hinders the exact match for the same word. Synonyms are very common in
Arabic. For example, “year” has three synonyms عام, حول, سنھ , and all are widely used in everyday
communication.
Discretization is defined as a symbol over and underscored letters, which are used to indicate the
proper pronunciations as well as for disambiguation purposes. Its absence in Arabic texts poses a
real challenge for Arabic NLP, as well as for translation, leading to high ambiguity. Though the
use of discretization is significant for readability and understanding, they don’t appear in most
printed media in Arabic regions nor on Arabic Internet web sites. They are visible in the Quran,
which is fully discretized to prevent misinterpretation [8].
4. CONSTRUCTION OF AMHARIC-ARABIC PARALLEL TEXT CORPUS
Construction of a parallel corpus is very challenging and needs a high cost of human expertise.
MT can produce high-quality translation results based on a massive amount of aligned parallel
text corpora in both the source and target languages [7]. MT systems need resources that can
provide an interpretation/suggestion of the source text and a translation hypothesis. Parallel
corpus consists of parallel text that can promptly locate all the occurrences of one language
expression to another language expression and is one of the significant resources that could be
utilized for MT tasks. [22].
As Amharic-Arabic parallel text corpora are not available for MT tasks, we have constructed a
small parallel text corpus by modifying the existing monolingual Arabic and its equivalent
translation of Amharic language text corpora available on Tanzile [23]. Quran text corpus
consists of 114 chapters, 6236 verses, and 157935 words. The organization of the Quran text is
categorized into verses (sequence of sentences and phrases). A sample verse in Arabic and its
equivalent translation in Amharic and English is shown in Table 1. A total number of 13,501
Amharic-Arabic parallel sentences corpora have been constructed to train the Amharic to Arabic
NMT system by splitting the verses manually into separate sentences of Amharic language as a
source sentence and Arabic language as a target sentence as shown in Table 2. The total size of
the corpus is 3.2MB, and it is split into training (80%) and test (20%).
The chapters in the Quran start with " ب سِْمِ ال ل َّھِ الرَّحْمَٰنِ الرَّحِیم ◌ِ" (in the name of Allah the most
gracious and the most merciful) in the Arabic text corpus. As it is merged with the first verses of
each chapter, we split it into a separate line. In the case of Amharic corpus, the equivalent
translation of this sentence is placed only in the first chapter. Therefore, it is added before the first
line of the equivalent translated Amharic verses.
60 Computer Science & Information Technology (CS & IT)
Table 1. Sample Verse of Quran in Arabic and its equivalent translation in Amharic and English
Chapte
r No:
Verse
No
Original Arabic
Verses
Equivalent translation of Amharic
Verses
Equivalent translation of
English Verses
یاَ أ یَُّھاَ ال ذَِّینَ آمَنوُاإ ذَا تَدَاینَتمُ 2:282
ب دَِیْنٍ إ لِىَٰ أ جََلٍ مُّسَمًّى فاَكْتبُُوه وَل یْكَْتبُ بَّیْنَكُمْ كَاتبٌِ ب اِل عَْدْلِ
وَلَا یأَ بَْ كَاتبٌِ أ نَ یكَْتبَُ كَمَا
عَل مََّھُ الل َّھُ فلَ یْكَْتبُْ وَل یُْمْللِِ
ال ذَِّي عَلیَْھ ال حَْقُّ وَل یْتَقَّ الل َّھ رَبَّھُ وَلَاب یخََسْ مِن ھُْ شَیْئ اً فإَ نِ
كَانَ ال ذَِّي عَلیَْھ ال حَْقُّ سَف یِھًا
أ وَْ ضَعِیف اً أ وَْ لَا یسَْتَطِیعُ أ نَ
یُمِلَّ ھُوَ فلَ یُْمْللِْ وَل یُِّھُ ب اِل عَْدْلِ
وَاسْتَشْھِدُوا شَھِیدَیْنِ مِن
رِّجَالكُِمْ فإَ نِ ل مَّْ یكَُونَا رَجُلیَْنِ
فرََجُلٌ وَامْرَأ تََانِ مِمَّن
تَرْضَوْنَ مِنَ الشُّھدََاءِ أ نَ
تَضِلَّ إ حِْدَاھُمَا فتَذَُك رَِّ إ حِْدَاھُمَا
الْأُخْرَىٰ وَلَا یأَ بَْ الشُّھدََاءُ إ ذَِا
مَا دُعُوا وَلَا تَسْأ مَُوا أ نَ
تَكْتبُُوه صَغِیرًا أ وَْ كَب یِرًا إ لِىَٰ
أ جََل ھِِ ذٰلَ كُِمْ أ قَ سَْط عِندَ الل َّھِ
وَأ قَ مَُْو ل لِشَّھاَدَةِ وَأ دَْنَىٰ أ لََّا
تَرْتَابُوا إ لَِّا أ نَ تَكُونَ تجَِارَة حَاضِرَة تدُِیرُونَھاَ بیَْنَكُمْ فلَیَْسَ
عَلیَْكُمْ جُنَاحٌ أ لََّا تَكْتبُُوھَا
وَأ شَْھِدُوا إ ذَِا تَباَیعَْتمُْ وَلَا
یُضَارَّ كَاتبٌِ وَلَا شَھِیدٌ وَإ نِ
تَف عَْل وُا فإَ نِھَّفُ سُوقٌ ب كُِمْ وَاتقَّ وُا
الل َّھ وَیُعَل مُِّك مُُ الل َّھُ وَالل َّھُ ب كُِلِّ
شَيْءٍ عَل یِم ◌ٌ
እናንተ ያመናችሁ ሆይ! እስከ ተወሰነ ጊዜ ድረስ
በዕዳ በተዋዋላችሁ ጊዜ ጻፉት፡፡ ጸሐፊም
በመካከላችሁ በትክክል ይጻፍ፡፡ ጸሐፊም አላህ
እንደ አሳወቀው መጻፍን እንቢ አይበል፡፡
ይጻፍም፡፡ ያም በርሱ ላይ ዕዳው ያለበት ሰው
በቃሉ ያስጽፍ፡፡ አላህንም ጌታውን ይፍራ፡፡
ከእርሱም (ካለበት ዕዳ) ምንንም አያጉድል፡፡
ያም በርሱ ላይ ዕዳው ያለበት ቂል፣ ወይም
ደካማ፣ ወይም በቃሉ ማስጻፍን የማይችል
ቢኾን ዋቢው በትክክል ያስጽፍለት፡፡
ከወንዶቻችሁም ሁለትን ምስክሮች
አስመስክሩ፡፡ ሁለትም ወንዶች ባይኾኑ
ከምስክሮች ሲኾኑ ከምትወዱዋቸው የኾኑን
አንድ ወንድና አንደኛዋ ስትረሳ አንደኛይቱ
ሌላዋን ታስታውሳት ዘንድ ሁለት ሴቶች
(ይመስክሩ)፡፡ ምስክሮችም በተጠሩ ጊዜ እንቢ
አይበሉ፡፡ (ዕዳው) ትንሽ ወይም ትልቅ ቢኾንም
እስከ ጊዜው ድረስ የምትጽፉት ከመኾን
አትሰልቹ፡፡ እንዲህ ማድረጋችሁ አላህ ዘንድ
በጣም ትክክል ለምስክርነትም አረጋጋጭ
ላለመጠራጠራችሁም በጣም ቅርብ ነው፡፡ ግን
በመካከላችሁ እጅ በጅ የምትቀባበሏት ንግድ
ብትኾን ባትጽፉዋት በናንተ ላይ ኃጢኣት
የለባችሁም፡፡ በተሻሻጣችሁም ጊዜ አስመስክሩ፡፡
ጸሐፊም ምስክርም (ባለ ጉዳዩ ጋር) አይጎዳዱ፡፡
(ይህንን) ብትሠሩም እርሱ በእናንተ (የሚጠጋ)
አመጽ ነው፡፡ አላህንም ፍሩ፡፡ አላህም
ያሳውቃችኋል፡፡ አላህም ነገሩን ሁሉ ዐዋቂ
ነው፡፡
"O believers, when you negotiate a
debt for a fixed term, draw up an
agreement in writing, though better
it would be to have a scribe write it
faithfully down; and no scribe
should refuse to write as God has
taught him, and write what the
borrower dictates, and have a fear
of God, his Lord, and not leave out
a thing. If the borrower is deficient
of mind or infirm, or unable to
explain, let the guardian explain
judiciously; and have two of your
men to act as witnesses; but if two
men are not available, then a man
and two women you approve, so
that in case one of them is
confused the other may prompt
her. When the witnesses are
summoned, they should not refuse
(to come). But do not neglect to
draw up a contract, big or small,
with the time fixed for paying back
the debt. This is more equitable in
the eyes of God, and better as
evidence and best for avoiding
doubt. But if it is a deal about
some merchandise requiring
transaction face to face, there is no
harm if no (contract is drawn up)
in writing. Have witnesses to the
deal (and make sure) that the
scribe or the witness is not harmed.
If he is, it would surely be sinful
on your part. And have a fear of
God, for God gives you the
knowledge, and God is aware of
everything."
Computer Science & Information Technology (CS & IT) 61
Table 2. Split Sentences of Quran Chapter 2 Verse number 282
Original Arabic
Sentences
Equivalent translation of
Amharic sentences
Equivalent translation of English
sentences
ی ا أ یَُّھاَ ال ذَِّینَ آمَنوُا እ ና ን ተ ያ መና ች ሁ ሆይ O believers
إ ذَِا تَدَاینَ تْمُْ ب دَِیْنٍ إ لِىَ أ جََلٍ مُسَمًّى
فاَكْتبُُوه ◌ُ
እ ስ ከ ተ ወ ሰ ነ ጊ ዜ ድ ረ ስ በ ዕ ዳ
በ ተ ዋ ዋ ላ ች ሁ ጊ ዜ ጻ ፉ ት
when you negotiate a debt for a fixed
term, draw up an agreement in writing
ول یْكَْتبُْ بیَْنَكُمْ كَاتبٌِب ال عَْدْل ◌َِ ጸሐፊም በመካከላችሁ በትክክል ይጻፍ though better it would be to have a
scribe write it faithfully down
وَلَا یأَ بَْ كَاتبٌِ أ نَْ یكَْتبَُ كَمَا عَل مََّھُ
للهَّ ◌ُ
ጸ ሐ ፊ ም አ ላ ህ እ ን ደ አ ሳ ወ ቀ ው
መጻ ፍ ን እ ን ቢ አ ይ በ ል
and no scribe should refuse to write as
God has taught him
ل یْكَْتبُ ◌َْف ይጻፍም and write
یُْمْللِِ ال ذَِّي عَلیَْھِ ال حَْق ◌وَلُ ◌ّ ያ ም በ ር ሱ ላ ይ ዕ ዳ ው ያ ለ በ ት ሰ ው
በ ቃ ሉ ያ ስ ጽ ፍ
what the borrower dictates
ول یْتَقَِّ للهَّ رَبَّھ ◌َُ አላህንም ጌታውን ይፍራ and have fear of God, his Lord
وَلَا یبَْخَسْ مِن ھُْ شَیْئاً ከ እ ር ሱም ካ ለ በ ት ዕ ዳ ምን ን ም
አ ያ ጉ ድ ል
and not leave out a thing
فإَ نِْ كَانَ ال ذَِّي عَلیَْھ ال حَْقُّ سَف یِھًا أ وَْ
ضَعِیف اً أ وَْ لَا یسَْتَطِیعُ أ نَْ یُمِلَّ ھُوَ
یُمْللِْ وَل یُِّھُ ب اِل عَْدْل ◌ِْ ف ل
ያ ም በ ር ሱ ላ ይ ዕ ዳ ው ያ ለ በ ት ቂ ል
ወ ይ ም ደ ካ ማ ወ ይ ም በ ቃ ሉ ማስ ጻ ፍ ን
የ ማይ ች ል ቢ ኾ ን ዋ ቢው በ ት ክ ክ ል
ያ ስ ጽ ፍ ለ ት
If the borrower is deficient of mind or
infirm, or unable to explain, let the
guardian explain judiciously
وَاسْتَشْھِدُوا شَھِیدَیْنِ مِنْ رِجَالكِ مُ ◌ْ ከ ወን ዶ ቻች ሁም ሁለ ት ን ምስ ክ ሮ ች
አ ስ መስ ክ ሩ
and have two of your men to act as
witnesses
فإَ نِْ لمَْ یكَُونَا رَجُلیَْنِ فرََجُلٌ وَامْرَأ تََانِ
مِمَّنْ تَرْضَوْنَ مِنَ الشُّھدََاءِ أ نَْ تَضِلَّ
إ حْدَاھُمَا فتَذُ كَِّرَ إ حِْدَاھُمَا الْأُخْرَى
ሁ ለ ት ም ወ ን ዶ ች ባ ይ ኾ ኑ ከ ምስ ክ ሮ ች
ሲ ኾ ኑ ከ ምት ወ ዱዋ ቸ ው የ ኾ ኑ ን አ ን ድ
ወ ን ድ ና አ ን ደ ኛ ዋ ስ ት ረ ሳ
አ ን ደ ኛ ይ ቱ ሌ ላ ዋ ን ታ ስ ታውሳ ት
ዘ ን ድ ሁ ለ ት ሴ ቶ ች ይ መስ ክ ሩ
but if two men are not available, then a
man and two women you approve, so
that in case one of them is confused the
other may prompt her
وَ لَا یأَ بَْ الشُّھدََاءُ إ ذِ اَ مَا دُعُوا ምስ ክ ሮ ች ም በ ተ ጠሩ ጊ ዜ እ ን ቢ
አ ይ በ ሉ
When the witnesses are summoned they
should not refuse (to come).
وَلَا تَسْأ مَُوا أ نَْ تَكْتبُُوه صَ غِیرًا أ وَْ
یرًا إ لِىَ أ جََلھِ ◌ِِ كب
ዕ ዳ ው ት ን ሽ ወ ይ ም ት ል ቅ ቢ ኾ ን ም
እ ስ ከ ጊ ዜ ው ድ ረ ስ የ ምት ጽ ፉ ት
ከ መኾ ን አ ት ሰ ል ቹ
But do not neglect to draw up a contract,
big or small, with the time fixed for
paying back the debt
ذَل كُِمْ أ قَ سَْط عِن دَْ للهَّ وَأ قَ وَْمُ ل لِشَّھاَدَة وَأ دَْنَىأ لََّا تَرْتَابُوا
እ ን ዲ ህ ማድ ረ ጋ ች ሁ አ ላ ህ ዘ ን ድ
በ ጣም ት ክ ክ ል ለ ምስ ክ ር ነ ት ም
አ ረ ጋ ጋ ጭ ላ ለ መጠራ ጠራ ች ሁ ም በ ጣም
ቅ ር ብ ነ ው
This is more equitable in the eyes of
God, and better as evidence and best for
avoiding doubt
إ لَِّا أ نَْ تَكُونَ تجَِارَة حَاضِرَة تدُِیرُونَھاَ بیَْنَكُمْ فلَیَْسَ عَلیَْكُمْ جُنَاحٌ أ لََّا
تَكْتبُُوھَا
ግ ን በ መካ ከ ላ ች ሁ እ ጅ በ ጅ
የ ምት ቀ ባ በ ሏ ት ን ግ ድ ብ ት ኾ ን
ባ ት ጽ ፉ ዋ ት በ ና ን ተ ላ ይ ኃ ጢኣ ት
የ ለ ባ ች ሁ ም
But if it is a deal about some
merchandise requiring transaction face
to face, there is no harm if no (contract is
drawn up) in writing.
وَأ شَْھِدُوا إ اَِذ تَباَیعَْتمُ ◌ْ በ ተ ሻ ሻ ጣች ሁም ጊ ዜ አ ስ መስ ክ ሩ Have witnesses to the deal
وَلَا یُضَارَّ كَاتبٌِ وَلَا شَھِید ◌ٌ ጸ ሐፊ ም ምስ ክ ር ም ባ ለ ጉ ዳ ዩ ጋ ር
አ ይ ጎ ዳ ዱ
(and make sure) that the scribe or the
witness is not harmed
ن تَف عَْل وُا فإَ نِھَُّ ف سُُوقٌ ب كُِم ◌ِْْ وَ إ ይ ህ ን ን ብት ሠሩ ም እ ር ሱ በ እ ና ን ተ
የ ሚጠጋ አ መጽ ነ ው
If he is, it would surely be sinful on your
part
وَاتقَّ وُا للهَّ ◌َ አ ላ ህ ን ም ፍ ሩ And have fear of God
وَیُعَل مُِّكُمُ للهَّ ◌ُ አ ላ ህ ም ያ ሳ ውቃች ኋ ል for God gives you knowledge
وَللهَُّ ب كُِلِّ شَيْءٍ عَل یِم ◌ٌ አ ላ ህ ም ነ ገ ሩ ን ሁሉ ዐ ዋ ቂ ነ ው and God is aware of everything
5. AMHARIC-ARABIC NMT
In this work, we adopted openNMT Attention-based Encoder-Decoder architecture, because
attention mechanisms are being progressively used to enhance the performance of NMT by
selectively focusing on sub-parts of the sentence during translation [24]. As described in [25],
”NMT takes a conditional language modeling view of translation by modeling the probability of a
target sentence w1:T given a source sentence x1:S as( : | ) = Π
( |:, ) .
This distribution is estimated using an Attention-based Encoder-Decoder architecture”.
62 Computer Science & Information Technology (CS & IT)
Two special kinds of Recurrent Neural Network (RNN) LSTM and GRU which are capable of
learning long-term dependencies are used in this research work. RNN is a type of neural network
for sequential data that can remember its inputs due to an internal memory which is more suited
for machine learning problems. It can produce predictive results in sequential data that the
information cycles through a loop when it makes a decision. It takes into consideration the current
inputs and also previously received inputs, which is learned earlier [26].
LSTM was first introduced by S. Hochreiter and J. Schmidhuber [27], to avoid the long-term
dependency problem. LSTM inherit the exact architecture from standard RNNs, with the
exception of the hidden state. The memory in LSTMs (called cells) takes as input the previous
state and the current input. Internally, these cells decide what to keep in and what to eliminate
from the memory. Then, they combine the previous state, the current memory, and the input.
LSTM calculates a hidden state ht as;
where t, i, f, o, W, U are called the time step, input gate, forget gate, output gate, the recurrent
connection at the previous and current hidden layer, and the weight matrix connecting the inputs
to the current hidden layers respectively. ~ is a "candidate" hidden state that is computed based
on the current input and the previous hidden state. C is the internal memory of the unit.
GRU extends LSTM with a gating network generating signals that act to control how the present
input and previous memory work to update the current activation, and thereby the current
network state. Gates are themselves weighted and are selectively updated [28]. For GRU, the
hidden state ht is computed as;
where, ℎ~ is activation r is a reset gate, and z is an update gate.
Computer Science & Information Technology (CS & IT) 63
Both LSTM and GRU are designed to resolve the vanishing gradient problem which prevents
standard RNNs from learning long-term dependencies through gating mechanism. The general
architecture of LSTM and GRU adopted from [29] [30] is shown in Figure 1.
(a) Long Short-Term memory (b) Gated Recurrent Unit
Figure 1: LSTM and GRU Architecture
A basic form of NMT consists of two components; an encoder which computes a representation
of source sentence S and a decoder which generates one target word at a time and hence
decomposes the conditional probability [24]. The Attention-based Encoder-Decoder architecture
used for Amharic-Arabic NMT is shown in Figure 2.
In general, the proposed model works as follows:
1. Reads the input words one by one to obtain a vector representation from every encoder
time step using LSTM/GRU based encoder
2. Provide the encoder representation to the decoder
3. Extract the output words one by one using another LSTM/GRU based decoder that is
conditioned on the selected inputs from the encoder hidden state of each time step
With this setting, the model is able to selectively focus on useful parts of the input sequence and
hence, learn the alignment (matching segments of original text with their corresponding segments
of the translation).
Figure 2: Attention based Encoder-Decoder architecture for Amharic-Arabic NMT
64 Computer Science & Information Technology (CS & IT)
As shown in Figure 2, Using LSTM Attention based Encoder-Decoder, each word from the
source sentence is associated with a vector wϵRd and will be transformed into [w0, w1, w0, w2,
w3, w4] ϵRdx5 by the encoder, and then an LSTM over this sequence of vectors is computed.
This will be the encoder representation (attention weights) e = [e0, e1, e0, e2, e3, e4]. The
attention weights and word vectors of each time step is fed to another LSTM cell to compute the
context vector which is computed as:
ht = LSTM(ht−1,[wi−1,Ct])
st = g(ht)
pt = softmax(st)
it = argmax(pt)
where, g is a trans-formative function that outputs a vocabulary size vector. A soft-max is then
applied to st to maximize it to a vector of probability pt ϵRV. Each entry of pt will measure how
likely is each word in the vocabulary and the highest probability pt is taken as it = argmax(pt ) ,
corresponding vector of =
The vector Ct is attention or context vector which is computed at each decoding step first with the
function f(ht -1, et ′) → αt ′ϵR then a score for each hidden state et ′ of the encoder is computed. The
sequence of αt ′ is normalized using a soft-max and Ct ′ is computed as the weighted average of
et ′. The same procedure is also applied for GRU based NMT
6. EXPERIMENTS AND RESULTS
An OpenNMT system which is an open-source toolkit for NMT is used to construct the NMT
model and translate the text in Amharic to the Arabic language. Preprocessing of both Amharic
and Arabic scripts have a great impact on the performance of the NMT system. Sentences are
split and aligned manually and then all punctuation marks are removed from texts. After
extensive experiments, the maximum source and target sequence length are set to 44, maximum
batch size for training and validation is set to 80 and 40 respectively and learning rate to 0.001
with Adam optimization for both LSTM and GRU RNN type. The remaining parameters are used
as default. The system saves the model for each of 10,000 training samples and then computes
accuracy and perplexity of each model 10 times. Perplexity is a measure of how easily a
probability distribution (the model) predicts the next sequence. A low perplexity indicates that the
translation model is good at predicting/translating the test set.
Computer Science & Information Technology (CS & IT) 65
Table 3 and Table 4 shows LSTM-based and GRU-based NMT evaluation, where, ”Val ppl”,
”Val Acc” ”Av. pred score” and ”Pred ppl” represents Validation Perplexity, Validation
Accuracy, Average Prediction Score and Prediction Perplexity respectively. The result indicate
that LSTM-based NMT outperforms GRU-based NMT. Since this is the first experiment done on
Amharic and Arabic parallel text corpus we consider it as a good performance with small size
corpus.
Table 3. LSTM-based NMT Evaluation
Epoc
hs
BLEUScore
Val.
PPL
Val.
Accuracy
Av. Pred.
Score
Pred.
ppl
1 0.11 12725 33.21 -0.49 1.64
2 0.11 41181.5 33.50 -0.40 1.49
3 0.11 100996 33.64 -0.35 1.41
4 0.12 100417 34.34 -0.34 1.40
5 0.12 99881.1 34.32 -0.34 1.40
6 0.12 99876.1 34.33 -0.34 1.40
7 0.12 99876 34.33 -0.34 1.40
8 0.12 99876 34.33 -0.34 1.40
9 0.12 99876 34.33 -0.34 1.40
10 0.12 99876 34.33 -0.34 1.4
Table 4. GRU-based NMT Evaluation
Epoc
hs
BLEUScore
Val.
PPL
Val.
Accuracy
Av. Pred.
Score
Pred.
ppl
1 0.108 13647 32.65 -0.51 1.66
2 0.101 65598.4 32.68 -0.39 1.48
3 0.098 172950 32.38 -0.35 1.43
4 0.105 173231 33.10 -0.34 1.40
5 0.105 175635 33.12 -0.34 1.40
6 0.105 175701 33.11 -0.34 1.40
7 0.105 175702 33.11 -0.34 1.40
8 0.105 175702 33.11 -0.34 1.40
9 0.105 175702 33.11 -0.34 1.40
10 0.105 175702 33.11 -0.34 1.40
The models are evaluated using Bilingual Evaluation Understudy (BLEU). BLEU is a score for
comparing a candidate translation of the text with reference translations. The primary
programming task for a BLEU implementer is to compare n-grams of the candidate with the ngrams
of the reference translation and count the number of matches. These matches are positionindependent.
More the matches, better the candidate translation is. BLEU is inexpensive to
calculate and it is quick to use. It is expressed as the following equation [31];
where pn is an n-gram precision that uses n-grams up to length N and positive weights wn that
sum to one.We also compared the two recurrent units LSTM and GRU based OpenNMT
translation algorithm with Google Translation System which is a free multilingual translation
system developed by Google to translate multilingual text [6] and the results are shown in Figure
3. LSTM based OpenNMT outperforms over GRU based OpenNMT and Google Translation
system, which is BLEU score of 12%, 11%, and 6% respectively.
66 Computer Science & Information Technology (CS & IT)
Figure 3: Best BLEU-Scores of LSTM and GRU based OpenNMT translation and Google
Translation System
7. CONCLUSION
Many researchers have investigated to solve translation problems for many language pairs and
NMT has recently shown promising results on multiple language pairs. However, much of the
research on this area has focused on European languages despite these languages being very rich
in resources. Since Amharic and Arabic languages lack parallel corpus for the purpose of
developing NMT, small size Amharic-Arabic parallel text corpora have been constructed to train
the Amharic to Arabic NMT system by splitting the verses manually into separate sentences of
Amharic language as a source sentence and Arabic language as a target sentence. Using the
constructed corpus LSTM-based and GRU-based NMT models are developed and evaluated
using BLEU. The results are also compared with Google Translation system. Since this is the first
experiment done on Amharic and Arabic parallel text corpus, we consider it as a good
performance for small size corpus. Extensive experiments with a large amount of training data
could be implemented for better performance.
REFERENCES
[1] R. Mitkov, The Oxford handbook of computational linguistics. Oxford University Press, 2004.
[2] B. Maylath, “Current trends in translation,” Commun. Lang. Work, vol. 2, no. 2, pp. 41–50, 2013.
[3] J. Oladosu, A. Esan, I. Adeyanju, B. Adegoke, O. Olaniyan, and B. Omodunbi, “Approaches to
Machine Translation: A Review,” FUOYE J. Eng. Technol., vol. 1, pp. 120–126, 2016.
[4] MemoQ, “5 Translation Technology Trends to Watch Out for in 2019,” 2019. [Online]. Available:
https://slator.com/press-releases/5-translation-technology-trends-to-watch-out-in-2019/.
[5] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, p. 436, 2015.
[6] Y. Wu et al., “Google’s neural machine translation system: Bridging the gap between human and
machine translation,” arXiv Prepr. arXiv1609.08144, 2016.
[7] S. Islam, A. Paul, B. S. Purkayastha, and I. Hussain, “CONSTRUCTION OF ENGLISH-BODO
PARALLEL TEXT CORPUS FOR STATISTICAL MACHINE TRANSLATION,” Int. J. Nat. Lang.
Comput. Vol, vol. 7, 2018.
[8] H. L. Shashirekha and I. Gashaw, “DICTIONARY BASED AMHARIC-ARABIC CROSS
LANGUAGE INFORMATION RETRIEVAL,” Comput. Sci. \& Inf. Technol., vol. 6, pp. 49–60,
2016.
[9] P. Koehn and R. Knowles, “Six challenges for neural machine translation,” arXiv Prepr.
arXiv1706.03872, 2017.
Computer Science & Information Technology (CS & IT) 67
[10] Gezmu Andargachew Mekonnen Andreas Nurnberger and T. B. Bati, “A Parallel Corpus for
Amharic-English Machine Translation,” 2018.
[11] S. M. Strassel and J. Tracey, “LORELEI Language Packs: Data, Tools, and Resources for
Technology Development in Low Resource Languages.,” in LREC, 2016.
[12] M. M. Sakre, M. M. Kouta, and A. M. N. Allam, “automated construction of Arabic-English parallel
corpus,” J. Adv. Comput. Sci., vol. 3, 2009.
[13] A. A.-S. Ahmad, B. Hammo, and S. Yagi, “ENGLISH-ARABIC POLITICAL PARALLEL
CORPUS: CONSTRUCTION, ANALYSIS AND A CASE STUDY IN TRANSLATION
STRATEGIES,” Jordanian J. Comput. Inf. Technol., vol. 3, no. 3, 2017.
[14] G. Inoue, N. Habash, Y. Matsumoto, and H. Aoyama, “A Parallel Corpus of Arabic-Japanese News
Articles,” in Proceedings of the Eleventh International Conference on Language Resources and
Evaluation (LREC-2018), 2018.
[15] H. M. Alotaibi, “Arabic-English Parallel Corpus: A New Resource for Translation Training and
Language Teaching,” Arab World English J. Vol., vol. 8, 2017.
[16] P. Resnik and N. A. Smith, “The web as a parallel corpus,” Comput. Linguist., vol. 29, no. 3, pp. 349–
380, 2003.
[17] K. Cho et al., “Learning phrase representations using RNN encoder-decoder for statistical machine
translation,” arXiv Prepr. arXiv1406.1078, vol. 3, 2014.
[18] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by jointly learning to align and
translate,” arXiv Prepr. arXiv1409.0473, vol. 7, 2014.
[19] A. Almahairi, K. Cho, N. Habash, and A. Courville, “First result on Arabic neural machine
translation,” arXiv Prepr. arXiv1606.02680, vol. 1, 2016.
[20] M. Wordofa, “Semantic Indexing and Document Clustering for Amharic Information Retrieval,”
AAU, 2013.
[21] M. El-Haj, “Multi-document arabic text summarisation,” University of Essex, 2012.
[22] L. Rura, W. Vandeweghe, and M. Montero Perez, “Designing a parallel corpus as a multifunctional
translator’s aid,” in XVIII FIT World Congress= XVIIIe Congrès mondial de la FIT, 2008.
[23] J. Tiedemann, “Parallel Data, Tools and Interfaces in OPUS.,” in Lrec, 2012, vol. 2012, pp. 2214–
2218.
[24] M.-T. Luong, H. Pham, and C. D. Manning, “Effective approaches to attention-based neural machine
translation,” arXiv Prepr. arXiv1508.04025, 2015.
[25] G. Klein, Y. Kim, Y. Deng, J. Senellart, and A. M. Rush, “Opennmt: Open-source toolkit for neural
machine translation,” arXiv Prepr. arXiv1701.02810, 2017.
[26] N. Donges, “Recurrent Neural Networks and LSTM,” 2018. [Online]. Available:
https://towardsdatascience.com/recurrent-neural-networks-and-lstm-4b601dd822a5.
[27] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Comput., vol. 9, no. 8, pp.
1735–1780, 1997.
[28] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical evaluation of gated recurrent neural
networks on sequence modeling,” arXiv Prepr. arXiv1412.3555, 2014.
68 Computer Science & Information Technology (CS & IT)
[29] A. Graves, “Generating sequences with recurrent neural networks,” arXiv Prepr. arXiv1308.0850,
2013.
[30] S. Du, T. Li, X. Gong, Z. Yu, Y. Huang, and S.-J. Horng, “A hybrid method for traffic flow
forecasting using multimodal deep learning,” arXiv Prepr. arXiv1803.02099, 2018.
[31] K. Wolk and K. Marasek, “Neural-based machine translation for medical text domain. Based on
European Medicines Agency leaflet texts,” Procedia Comput. Sci., vol. 64, pp. 2–9, 2015.
AUTHORS
Ibrahim Gashaw Kassa is a Ph.D. candidate at Mangalore University
Karnataka State, India, since 2016. He graduated in 2006 in Information System
from Addis Ababa University, Ethiopia. In 2014, he obtained his master’s
degree in Information Technology from the University of Gondar, Ethiopia., and
he serves as a lecturer at the University of Gondar from 2009 to May 2016. His
research interest is in Cross-Language Information Retrieval, Machine
translation Artificial Intelligence Natural Language Processing
Dr. H L Shashirekha is a Professor in the Department of Computer Science,
Mangalore University, Mangalore, Karnataka State, India. She completed her
M.Sc. in Computer Science in 1992 and Ph.D. in 2010 from University of
Mysore. She is a member of Board of Studies and Board of Examiners (PG) in
Computer Science, Mangalore University. She has several papers in
International Conferences and published several papers in International Journals
and Conference Proceedings. Her area of research includes Text Mining and
Natural Language Processing.
No comments:
Post a Comment