00051000948 english – vietnamese translation of deep learning terms in the book “deep learning” based on peter newmark’s framework

00051000948 english – vietnamese translation of deep learning terms in the book “deep learning” based on peter newmark’s framework

VIETNAM NATIONAL UNIVERSITY

UNIVERSITY OF LANGUAGES AND INTERNATIONAL STUDIES

FACULTY OF POSTGRADUATE STUDIES

NGUYỄN DANH TRUNG NGHĨA

ENGLISH – VIETNAMESE TRANSLATION OF DEEP LEARNING TERMS IN THE BOOK “DEEP LEARNING” BASED

ON PETER NEWMARK’S FRAMEWORK

(Nghiên cứu chiến lược dịch Anh – Việt các thuật ngữ Học sâu sử dụng trong

sách “Học sâu” dựa trên khung lý thuyết của Peter Newmark)

MA THESIS

Field: English Linguistics Code: 8220201.01

Ha Noi –2025

VIETNAM NATIONAL UNIVERSITY

UNIVERSITY OF LANGUAGES AND INTERNATIONAL STUDIES

FACULTY OF POSTGRADUATE STUDIES

NGUYỄN DANH TRUNG NGHĨA

ENGLISH – VIETNAMESE TRANSLATION OF DEEP LEARNING TERMS IN THE BOOK “DEEP LEARNING” BASED

ON PETER NEWMARK’S FRAMEWORK

(Nghiên cứu chiến lược dịch Anh – Việt các thuật ngữ Học sâu sử dụng trong

sách “Học sâu” dựa trên khung lý thuyết của Peter Newmark)

DECLARATION

I declare that the thesis, entitled “English – Vietnamese translation of deep learning terms in the book “Deep Learning” based on Peter Newmark’s framework.”

(Nghiên cứu chiến lược dịch Anh – Việt các thuật ngữ Học sâu sử dụng trong sách

“Học sâu” dựa trên khung lý thuyết của Peter Newmark), has been composed

solely by myself and that it has not been submitted, in whole or in part, in any previous application for a degree Except where states otherwise by reference or acknowledgment, the work presented is entirely my own

Hanoi, 2025

Approved by

SUPERVISOR

(Signature and full name)

Supervisor: Assoc.Prof Dr Lê Hùng Tiến

ACKNOWLEDGEMENTS

I like to express my profound gratitude to my respected Supervisor, Assoc Prof Dr

Le Hung Tien, who has consistently provided invaluable insights into my thesis, guided me in the correct route, and offered unwavering encouragement and patience throughout the process

I express sincere gratitude to my wife and family, who have always supported me throughout the thesis process, providing immense care, encouragement and support

I would like to express my profound gratitude to my friends, my wife’s colleagues for their close and effective interaction and collaboration, which significantly alleviated the challenges of completing my thesis and provided me with a complete insight of the Deep learning field

ABSTRACT

This research aims to explore translation procedures used in the Vietnamese translated version of “Học sâu” conducted by a group of engineers named dlbooksvn, the original version named “Deep learning” by Ian Goodfellow, Yoshua Bengio, and Aaron Courville, published by The MIT Press, 2016 His study utilises the approach proposed by P Newmark (1988) to systematically categorise terminologies associated with Deep Learning and to compare them with contemporary translation methods in the industry The research findings indicate that deep learning terminology consists of complex multi-layer concepts that may hinder the understanding of both professionals and learners in the domain The shift and couplet procedures by Peter Newmark are the most common ones, along with others which are: literal translation, transference, naturalisation, descriptive equivalence, recognised translation, addition, omission, and functional equivalence The study suggests the application of translation procedures when dealing with deep learning terms in Vietnam

2.1.2 Peter Newmark’s theory of translation procedure 6

3.4 Quantitative and qualitative data analysis 24

CHAPTER 4: FINDINGS AND DISCUSSION 26 4.1 Translation procedure for technical terms 26

5.1.1 Answers to question 1: What translation procedures are used to translate deep

5.1.2 Answers to question 2: What translation procedures are most commonly used

5.2 Significance of the findings 50 5.2.1 Contribution to Deep learning development in Vietnam 50

CHAPTER 1: INTRODUCTION

1.1 Rationale

Deep learning is an advanced branch of machine learning Beginning as just a mathematical model of a biological neuron introduced by Walter Pitts and Warren McCulloch in 1943, deep learning has developed rapidly, becoming one of the most promising branches of modern machine learning Deep learning's rapid evolution is reflective of the accelerating pace of technological advancement, transforming from

a simple mathematical model of a neuron to a powerful tool that can discern patterns and relationships within massive data sets that defy human comprehension

In their seminal 1943 work, Pitts and McCulloch developed a model for how neurons in the brain might work This marked the conception of the artificial neural network - a network designed to mimic the way the human brain learns from experiences In the following decades, the model underwent refinements and expansions as computing power and available data increased However, it was not until the dawn of the 21st century that deep learning truly began to gain traction, riding on the wave of exponentially growing digital data and ever-improving computational power

However, the translation of these types of documents presents problems As a means to transfer intellectual knowledge, these documents contain new words derived in the process of developing a theory in the source language (SL), which are not available in the target language (TL) In other cases, familiar words in common situations may present new sets of features that require different words in the TL to carry the message

This real situation has inspired the researcher to conduct an investigation on the translation procedure of deep learning terminologies from English to Vietnamese, as applied by a group of Vietnamese specialists named “dlbooksvn”

1.2 Scope of the study

Not many machine learning materials, particularly deep learning materials or documents, are found in Vietnam’s bookstores This research focuses on studying English Deep Learning terms and their Vietnamese equivalents, mainly from the book “Học sâu” (2022) translated by dlbooksvn and its original version, which is the book “Deep Learning” by Ian Goodfellow, Yoshua Bengio, and Aaron Courville (2016)

1.3 Aims of the study

This research aims to explore the translation procedures used in the Vietnamese-translated version of “Học sâu” conducted by a group of engineers named dlbooksvn, based on the model proposed by P Newmark (1988)

● Firstly, the researcher aims to indicate the types of translation procedures applied in the process of translating deep learning terminologies

● Secondly, the study suggests possible translation procedure could be used when translating deep learning terminologies

For the purpose of providing the most relevant and practical approach to this field, the sample text used to collect data is “Deep Learning” by Ian Goodfellow, Yoshua Bengio, and Aaron Courville (2016) The Vietnamese version, namely “Học sâu” (2022), translated by dlbooksvn, is chosen for the study

1.5 Significance of the study

This research clarifies the Vietnamese translation procedure used in technical-related texts performed by specialists in the field Therefore, this work is expected to become a good source of reference for students, language practitioners

in the translation and interpretation field, and anyone interested in translating

specialised terminologies

1.6 Research method

There ar two primary research methodologies: qualitative and quantitative With the objective of examining the terminologies used in the source material and the translation utilised by dlbooksvn, it was decided that a combination of quantitative and qualitative to form a descriptive analysis would be the primary method of research

1.7 Organization of the thesis

The paper is divided into five chapters, as follows:

Chapter 1: Introduction

This chapter provides readers with an overall summary of the research project, including an explanation of its purpose, goals, and objectives, as well as its organisational structure

Chapter 2: Theoretical background

This chapter will discuss the theoretical context in which the research should be interpreted in light of the topics covered

Chapter 3: Methodology

This chapter provides the main methodology and procedures for data collection

Chapter 4: Findings and Discussion

This chapter presents, analyses, and discusses the results of what the researcher found out from the collected data It also gives answers to the research questions

Chapter 5: Conclusion

This chapter summarises the overall study and suggests some forms of further study

in the field

CHAPTER 2: THEORETICAL BACKGROUND

2.1 An overview of translation

2.1.1 Translation procedure

Translation, as a means to transfer knowledge and understanding between different

languages, is needed to catch up with the overall requirements

The study of translation procedures is central to the field of translation studies Different scholars have proposed various taxonomies and approaches to describe how translators go about the complex task of transferring meaning from one

language and culture to another

2.1.1.1 Russian approaches (Retsker and Shveitser)

Retsker (1974) and Shveitser (1973) undertook one of the first systematic efforts to categorise the relationships between source and target languages in translation Their paradigm categorised interactions into three primary types: equivalence, variant/contextual correspondence, and other translational transformations This taxonomy was designed to address how translation techniques manage linguistic discrepancies while preserving semantic equivalence Retsker and Shveitser defined equivalence as existing when a source and target item have the same meaning, regardless of differences in linguistic form This rigid similarity is observed at the level of individual lexical items or phrases For instance, the English term 'carbon monoxide' is consistently translated in the same manner, irrespective of the intended language Because of differences in how meanings are structured in different languages, variant or contextual correspondence keeps the same meaning while requiring the use of different words For example, the English word 'see' can be translated into German as 'sehen' or 'sehen machen' The third category covered scenarios in which neither strict meaning matches nor contextual correspondence is possible between source and target The translation needs to go through additional divergence through 'other translational transformations' This approach categorises translation techniques in accordance with the level of formal deviation necessary to achieve semantic equivalency

2.1.1.2 French approaches (Vinay and Darbelnet)

Vinay and Darbelnet (1995) provided an approach to translation which can be characterised by their focus on the practical aspects of the translation process and their emphasis on achieving equivalence between the source text and the target text They introduced the concept of direct and oblique translation, which is central to their theory These concepts are mentioned as follows: “"Translation may operate either by matching textual material (words, terms, idioms) that have analogous functions (designata) in both languages, or by resorting to transposing (oblique processes) We call the former direct translation and the latter oblique translation”

In other words, direct translation involves transposing the exact meaning through analogous structures in the target language Oblique translation involves modifying the syntax From these two translation methods, Vinay and Darbelnet proposed seven translation procedures:

● Borrowing - introducing a foreign term into the target language

● Calque - translating a term literally via a calque or loan translation

● Literal translation - word-for-word translation

● Transposition - changing grammar elements like active/passive voice

● Modulation - altering the viewpoint, framework, or category of expression

● Equivalence - translating the idea rather than the literal meaning with a fully different stylistic construction

● Adaptation - altering the message to suit the target cultural context

2.1.1.3 American approach (Malone)

Malone (1988) proposed a taxonomy of translation techniques aimed at capturing the cognitive processes underlying the activity Building on the foundation laid by earlier theorists like Vinay and Darbelnet, Malone categorised translation into overarching “trajectories” or processes: Matching (equation and substitution), Reordering, Recrescence (including amplification and reduction), Repackaging (diffusion and condensation), and Zigzagging (divergence and convergence) Malone's model describes types of translation process without dictating usage It

provides a metalanguage for analysing translation decisions rather than prescriptive rules Its analytic rather than predictive nature means its applicability depends on the degree of structural divergence between source and target Nonetheless, internalising these techniques helps translators think systematically about bridging interlingual gaps

2.1.2 Peter Newmark’s theory of translation procedure

Newmark (1988: 81) proposed translation procedures to translate sentences and other smaller language units He proposed a range of translation procedures that provide flexibility to identify and categorise depending on specific situations

In short, the translation procedure can be fundamentally understood as the choice of appropriate translation methods and procedures employed in certain instances For the purposes of application in my thesis, I have personally chosen to use the theoretical framework about translation techniques that Newmark (1988) proposed

In the book "A textbook of translation" by Peter Newmark (1988:68-91), he

classified the translation procedures into seventeen types, which were as follows:

2.1.2.1 Literal translation

The primary procedure of translation, according to Peter Newmark's theory (1988),

is literal translation Literal translation hinges on directly transferring a text from the source language (SL) into a grammatically correct and meaningful text in the target language (TL) When employing this approach, the translator's primary focus lies in conforming to the grammatical rules of the target language According to Newmark (1995:69), literal translation can range from translating word by word to groups of words, collocations, and even entire sentences As the linguistic unit becomes longer, the occurrence of one-to-one correspondence diminishes Furthermore, he emphasized that common objects often have one-to-one literal translations if there's cultural overlap, such as in French and Italian For example, the English words

"machine," "deep," and "code" each have a single corresponding Vietnamese meaning in the dictionary, namely "máy," "sâu," and "mã," respectively

Consequently, the translator must utilise the existing meanings of these words

in meaning" (p 77) An example of this can be found in the transfer of the French phrase "savoir-faire" to English In this instance, the term is directly transposed, capturing the essence of the skill and tact that underlie the original expression

2.1.2.3 Naturalisation

The third type of procedure is the naturalisation translation Newmark (1988:82) explained this concept as a subsequent step in the transference process While transference involves direct word transfer from SL to TL, naturalisation involves more than just transfer; it entails adapting the SL word to align with the normal phonetic pronunciation and morphological structure of the TL Newmark (1988) articulates this sequence succinctly: "The procedure succeeds transference and

adapts the SL word first to the normal pronunciation, then to the normal morphology of the TL" (p 82) This systematic approach ensures that the transferred word not only retains its core meaning but also seamlessly integrates into the linguistic fabric of the TL

The technique of naturalisation becomes a potent tool in enhancing the efficacy of the transference procedure Pronunciation and morphology, being integral components of language, can significantly influence a term's reception and comprehension in the TL As Newmark underscores, "Words for specific objects, with specialised meanings, sometimes scientific words, can usually be transferred directly without loss or gain in meaning" (p 77) For instance, consider the Vietnamese word "véc-tơ." This term has a similar pronunciation to the English word "vector" in the SL When translating "véc-tơ" into English, the naturalisation procedure would involve adapting the pronunciation to the standard English pronunciation as well as adjusting its morphology to align with the English word structure This way, the term "véc-tơ" is not only recognizable to the target audience due to its shared pronunciation with "vector," but it also adheres to the English language's linguistic norms, resulting in a smoother integration into the target language

2.1.2.4 Cultural equivalent

Cultural equivalent involves the substitution of words or phrases from the source language (SL) with culturally resonant equivalents in the target language (TL) Cultural equivalents go beyond literal translation, acknowledging that certain terms carry connotations, implications, and emotions that are deeply ingrained within specific cultural contexts The foundation of the cultural equivalent procedure rests

on the principle that some concepts and expressions are culturally bound and lack direct equivalents in the TL Newmark (1988) explicates this phenomenon, noting that "culture-bound words often need no alteration in the translation" (p 77)

The application of the cultural equivalent procedure is particularly salient when dealing with idiomatic expressions and concepts that are deeply rooted in cultural

traditions Consider the English idiom "break a leg," used to wish someone good luck before a performance A direct translation of this expression would render it nonsensical or awkward in many languages In the Vietnamese context, an equivalent idiomatic expression might be "chúc may mắn," which captures the sentiment of well-wishing without invoking imagery that is peculiar to a different culture

2.1.2.5 Functional equivalent

This is a standard practice that is used when a culturally free word or a new particular term in the target language (TL) is needed As a result, the SL term is neutralised or generalised This method is the most accurate technique to translate when it is impossible to locate a culturally similar phrase in TL We may extend the process to translate technical terms for which there is no TL equivalent This method may occasionally be employed if the phrase has minimal significance in theatre since it might have an immediate impact Between the language and culture

of SL and the language and culture of TL, this technique resides in the middle Undertranslation results from practising the functional equivalent one-to-one That can be an excessive translation if done one to two

2.1.2.6 Descriptive equivalent

Descriptive equivalent is a translation procedure where the meaning of a source language term or phrase is explained and described in the target language text rather than directly translated The descriptive equivalent procedure is particularly useful when dealing with source text elements that are deeply rooted in the source culture and may not have an immediate counterpart in the target culture or language Instead of attempting to find an exact word or phrase to replace the cultural reference, the translator provides a detailed description or explanation within the target text This description aims to convey the same information, cultural significance, and nuances as the original source text

2.1.2.7 Synonymy

According to Newmark (1995: 84), the term "synonym" is employed to refer to a

target language (TL) word that closely approximates a source language (SL) word within a given context, regardless of whether an exact equivalent exists or not This method is employed when dealing with a second language (SL) term that lacks a distinct one-to-one correspondence, and when the term in question holds no significant relevance within the text, particularly in the case of quality-based adjectives or adverbs The approach is considered suitable only in cases where a literal translation is impossible, as well as when the word in question lacks sufficient significance for a componential analysis In this context, accuracy is subordinate to economy This procedure is observable in numerous instances throughout our daily lives

2.1.2.8 Through-translation

Newmark (1988:84) introduces the concept of through-translation as a technique in which indirect speech or thought structures are transformed into a direct form while still preserving their logical connections This is achieved by removing introductory verbs and retaining the representations in their original form

Through-translation finds application in various scenarios, such as the literal translation of collocations, the names of organisations, compound word components, and phrases Newmark refers to this process as calque or loan translation Notably, when proper nouns, such as the names of organisations, consist

of universally recognizable words, they can be directly transferred In addition,

"international organisations are often known by their acronyms" For instance:

"UNESCO" is translated as "United Nations Educational, Scientific, and Cultural Organization"

2.1.2.9 Shift or transposition

Catford (1965) refers to this as a transposition or shift, and it describes the grammatical change that takes place while translating from SL to TL Newmark's concept of the shift translation procedure encompasses a range of modifications to bridge the linguistic and cultural gaps between the SL and TL These shifts are essential for achieving not only linguistic accuracy but also readability and cultural

relevance in the translation Some key types of shifts, as outlined by Newmark, include:

(i) Change from Singular to Plural: This shift occurs when the SL uses a singular form, but the TL requires a plural form, or vice versa The translator makes this shift to ensure grammatical correctness in the TL

(ii) Change necessary when an SL Structure is absent from the TL: In cases where a specific grammatical or structural feature present in the SL is not found in the TL, the translator must adapt the text by altering the structure or using an alternative approach to convey the same meaning

(iii) Change from an SL Verb to a TL Word: When the SL uses a verb, but the TL requires a different word class, such as a noun or an adjective, the translator performs this shift to maintain syntactic and semantic coherence in the TL

(iv) Change from an SL Noun Group to a TL Noun: This shift involves transforming a noun group (e.g., a noun with its modifiers) in the SL into a single noun in the TL It simplifies the structure while retaining the essential information Newmark's classification of these shifts underscores the dynamic nature of translation, where linguistic, grammatical, and structural adjustments are often necessary to convey the meaning and nuances of the ST effectively in the TL

This involves rephrasing a negative statement positively, or vice versa The term "it" has the potential to be used for any action or quality that is expressed vocally or in

an adjectival manner Newmark provides a reasonable caution that, in theory, a negative statement may possess less persuasive power in comparison to an

affirmative statement However, from a pragmatic standpoint, the significance of emphasis is contingent upon the specific context and manner in which it is conveyed

Newmark emphasises the use of the "negated contrary" as a specific method for reformulating claims The appropriateness of the material is contingent upon its formulation and the context in which it is presented Vinay and Darbelnet (1958) delineate various types of translation, however, their categories have been subject to criticism due to the perceived unpredictability and lack of predictive efficacy for translators In general, Newmark perceives modulation as a descriptive framework rather than a prescriptive one, highlighting its inherent usefulness

2.1.2.11 Paraphrase

This is explained by Newmark (1988) as a clarification or explanation of the meaning of this passage of text When a text is poorly written or contains significant implications and omissions, it is utilised as an "anonymous" text Since this term is frequently used to imply free translation, one should be cautious when categorising paraphrases as a translation technique A statement that is unclear or ambiguous is minimally altered in order to make it clearer For example, the term “intractable” is normal translated into “khó làm, khó chữa, khó bảo”, which means hard to control

or deal with But in the context of deep learning, the translation of the term is “khó tính toán”, which is closer to the meaning hard to calculate

2.1.2.12 Translation label

Newmark (1995: 90) considered translation label as “a provisional translation, usually of a new institutional term, which should be made in inverted commas, which can later be discreetly withdrawn"

2.1.2.13 Recognized Translation

Newmark (1988, p 89) suggests that the translators "should normally use the official or the generally accepted translation of any institutional term" He believes that in translating it is not good to give translators' own titles or a brief explanation and just the accepted term should be used in the translation because changing the

term may cause confusions especially in official or serious informative texts For example, the term “naive Bayes” is translated into “giả luận Bayes” to help readers

understand, rather than translating literally

2.1.2.14 Componential analysis

Componential analysis involves breaking down a word into its core components (or features) to better understand and translate its meaning This method compares a SL word with a TL word which has a similar meaning, but is not an obvious one-to-one equivalent, by demonstrating first their common and then their differing sense components For example, ‘jolly’ in 'jolly good' is mainly pragmatic, a slight, middle-class intensification which can only be over-translated in French (drdlemem)

2.1.2.15 Addition

Addition is the process of adding extra words or explanations in the target language

to make the meaning of the source text clearer, especially when cultural or contextual information needs to be conveyed to the target audience It facilitates the intended readers' understanding of new vocabulary or ideas For example, the term

“biased importance sampling” is transferred into “lấy mẫu theo độ quan trọng có chệch” The translation does not have any additional meaning but has added the

“theo” and “có” to provide clarity for readers

2.1.2.16 Omission

Omission involves deliberately leaving out certain words or elements from the source text in the translation when they are redundant or do not add value in the target language This is used to ensure fluency and clarity in the translation, particularly when the omitted elements are not essential to the message For example, the term “bag of words” is translated with the omission of the word “of” and turned into “túi từ”

2.1.2.17 Couplet

Newmark (1988) refers to couplet as "the combination of two translation procedures for one unit" (p 83) Moreover, triplets and quadruplets combine three or four of the

procedures for dealing with a single problem For example, to transfer the meaning

of the term “undirected probabilistic graphical model," the translator has to utilise a mixture of shift and paraphrase translation procedures to “mô hình đồ thị xác suất

vô hướng.”

2.1.3 Technical translation

2.1.3.1 Definition of technical translation

Machine learning in general, and deep learning in particular, are specialised technical fields Therefore, technical translation theory should be mentioned and applied in this analysis

Technical translation is a noteworthy concept mentioned in Peter Newmark’s research According to Peter Newmark (1995: 151), “Technical translation is one part of specialised translation, institutional translation, the areas of politics, commerce, finance, government, etc., is the other I think technical translation is potentially (but far from actually) non-cultural, therefore ‘universal’ Newmark also commented that “Technical translation is primarily distinguished from other forms

of translation by terminology, although terminology usually makes up about 5–10%

of a text."

2.1.3.2 Varieties of technical styles

According to Newmark (1995: 152), an in-depth overview of technical language can be found in his work on a variety of technical styles He proposes four distinct categories: scientific, workshop level, everyday usage level, and publicity/sales Newmark's classification can be succinctly summarised as including three discrete levels

- Academic: At this level, technical language is characterised by the inclusion

of Latin and Greek terminology, which is frequently encountered in scholarly articles, particularly in specialised fields These phrases frequently establish the basis for exact scientific discourse For example, the term “eigenvector”

is a term proposed by a German mathematician David Hilbert and others in the early 20th century, and "Vector" comes from the Latin word vector,

meaning "carrier" or "one who carries." In mathematics, a vector is a quantity with both magnitude and direction

- Professional: The professional level encompasses specialised vocabulary that

is employed solely by specialists and practitioners within a specific discipline The utilisation of such technical terminology has been tailored to meet the specific requirements of experts while upholding a notable level of specialisation and precision For instance, the term “supervised learning” refers to a branch of machine learning that trains algorithms to identify patterns and predict outcomes using labelled datasets, rather than some kind

of learning that is being monitored by any person in order to be executed correctly Therefore, the term is transferred into “học có giám sát” in Vietnamese

- Popular: In contrast to the more esoteric academic and professional levels, the popular level encompasses technical terminology that is easily comprehensible to the majority of people Frequently, it employs terminology that is readily understood and accessible to individuals without specialised knowledge, rendering technical material more comprehensible to the general public For example, the term “Artificial intelligence” or “AI” is quite common all over the world as computer systems capable of performing complex tasks that historically only a human could do The word with an equivalence in Vietnamese is “Trí tuệ nhân tạo”, which has been popularised

by news and media recently

During the process of translation, the translator encounters a significant decision-making process about the choice of technical terminology The presence of direct technical equivalents in the target language and the translator's dual role as a conduit for communication and a bridge between cultures both have an impact on the choice Newmark's methodology gives rise to two basic translation strategies:

- Technical Terms: In cases where a descriptive phrase in the source language has a corresponding technical term in the target language, the translator may

choose to employ the technical term This approach demonstrates the translator's mastery of the specialised lexicon in the target language, therefore improving communication with a specialised audience

- Descriptive Terms: On the other hand, in cases where a technical term in the source language does not have a direct equivalent in the target language, it is recommended for the translator to opt for a descriptive term instead The implementation of this particular approach guarantees a high level of clarity and comprehension for the intended audience while simultaneously minimizing any potential confusion that may arise due to the utilisation of unfamiliar terminology

The taxonomy of technical language that Peter Newmark proposed, which includes academic, professional, and popular levels, offers significant insights into the difficulties of technical translation The decision to use technical terms or descriptive variants depends on the availability of direct equivalents and the translator's goal of enhancing efficient communication Translators are able to navigate the intricacies of technical translation by understanding these nuances with finesse, preserving accuracy and accessibility in the target language

2.1.4 Deep learning terminology

2.1.4.1 Introduction of deep learning

Deep learning is a subfield of machine learning that focuses on the development of artificial neural networks capable of learning and making predictions from data The term "deep" in deep learning refers to the use of deep neural networks, which have multiple layers of interconnected artificial neurons This subfield has gained significant attention and popularity due to its remarkable success in a wide range of applications, including image and speech recognition, natural language processing, and autonomous systems The book "Deep Learning" by Ian Goodfellow, Yoshua Bengio, and Aaron Courville (2016) contains an introduction and explanation of deep learning The concept of “Deep learning” is introduced by the authors of this

book as a subset of machine learning, a methodology that enables computer systems

to enhance their performance through experience and data It is also mentioned that deep learning is a specific form of machine learning that attains significant capability and adaptability by acquiring the ability to describe the world as a hierarchical structure of interconnected ideas Each idea is defined in reference to simpler ideas, and more abstract representations are calculated based on less abstract ones

The following diagram, which is provided in the Deep Learning book, provides an overall view of deep learning as a subset of representation learning, machine learning, and article intelligence (AI)

Figure 1: A Venn diagram showing how deep learning is a kind of representation learning,which is in turn a kind of machine learning (Source: Deep Learning)

2.1.4.2 Areas of deep learning

Deep learning, a subset of machine learning, is an extensive field with numerous categorizations based on its diverse applications, methodologies, and theoretical foundations The book Deep learning (2016), which is used in this research, provided a classification of deep learning subsets and relating area which will be included in the book in the form of table below:

Figure 2: The high-level organization Deep learning An arrow from one chapter to anotherindicates that the former chapter is prerequisite material for understanding

the latter (Source: Deep Learning)

In summary, the "Deep Learning" book by Goodfellow, Bengio, and Courville serves as a foundational text that provides the necessary theoretical background and context to understand and appreciate the significance of these landmark research papers It provides vital knowledge about the fundamental principles and methodologies that support the progress showcased in these studies, making it an invaluable resource for individuals studying or working in the field of deep learning Hence, in my subjective viewpoint, the book is the optimal resource for this linguistic research

2.1.5 Previous studies

As stated, deep learning is an emerging field of study, offering plenty of study possibilities At the time this research was being conducted, investigations into deep learning and linguistic translation were limited to employing deep learning as a model for enhancing translation tools or for examining language usage Consequently, no studies have been conducted on the translation of deep learning terms To perform adequate research on this novel approach, the researcher required references to existing linguistic studies on terminology translation completed in several disciplines

2.2 Summary of the chapter

In this chapter, the theory on translation procedures by Newmark (1995) as well as technical translation with varieties of technical styles is stated The field of Deep learning is introduced and presented here with three subsets parts The

Later, the deep learning terms are divided in terms of varieties of technical styles, which includes: Academic (phrases frequently establish the basis for exact scientific discourse), Professional (employed solely by specialists and practitioners within a specific discipline), and Popular (technical terminology that is easily comprehensible to the majority of people)

Last but not least, deep learning terms are divided into technical terms and descriptive terms based on the theory of Newmark (1995)

CHAPTER 3: RESEARCH METHODOLOGY

This chapter will focus on introducing the methodology applied in the research based on Newmark’s (1995) framework of translation procedures Later on, this methodology will be used to apply to the translation process of deep learning terms from English to Vietnamese This approach will provide a thorough understanding

of primary deep learning terminologies currently being established and developed in the field

The research would be conducted in the form of a descriptive analysis, with a combination of quantitative and qualitative data collection and analysis The research plan, including the methodology, study participants, procedures, and analysis method, are also the primary components of this chapter

3.1 Methodology

Descriptive analysis in the translation of linguistic terminology involves systematic investigation and interpretation of language to document, compare, and comprehend the formation, usage, and translation of terms across different languages This method emphasises the analysis of linguistic aspects to deliver a comprehensive examination of language usage, particularly in specialised fields Descriptive analysis, however, does not seek to demonstrate the causation of test hypotheses Quantitative analysis is crucial in descriptive analysis, particularly in linguistic research that requires statistical evaluation of language data It emphasises the systematic quantification and numerical representation of linguistic features, allowing scholars to objectively discern patterns and trends This method is suitable for analysing large datasets, particularly when the research objective is to determine the predominant procedure employed in the translation process While it does not address causation, it offers precise insights that can guide subsequent qualitative research in linguistics

In addition to quantitative analysis, qualitative analysis is employed to understand non-numerical data, revealing meaning, context, and trends in language usage This

approach is especially adept at investigating the "how" and "why" of linguistic occurrences, offering depth and detail that enhances quantitative findings

The combined use of both qualitative and quantitative analysis in descriptive analysis enhances language research, particularly in terminology translation Quantitative statistics offer an accurate empirical basis, revealing patterns and trends that may not be obvious from qualitative observation alone Qualitative analysis enhances these discoveries by clarifying the reasons behind statistical trends and situating them within practical contexts By integrating these analyses, scholars gain a more comprehensive picture of the translation process, connecting quantifiable data with interpretive insights

In this research, after the data are collected and classified based on the translation procedures theory of Newmark (1995), then generalized into the most commonly used translation procedures As appeared in chapter two of the thesis, translation procedures are examined to find out the useful procedures to translate deep learning terms from English to Vietnamese In short, with the objective of examining the terminologies used in the source material and the translation utilized by dlbooksvn,

it was decided that the combination of quantitative and qualitative analysis would

be the best option for this research

3.2 Study participants

The participants in the thesis are three IT practitioners who work primarily with and utilise deep learning knowledge in their daily work The selected participants have different levels in the field

The first participant (pseudonym: Andy), is the most experienced, with over five years of experience working with deep learning He studied IT in a public university, and gained his knowledge in the field during the process of working Son also able to learn Deep Learning through articles and videos in English He did not have any knowledge about the book Hence, Andy would be familiar with the original version “Deep learning” the most

The second participant (pseudonym: Brian), has over three years of experience

working in the AI field and has the ability to work with multiple programming languages, including deep learning Brian studied at a private IT education institute, combining with self-learning online It is expected that this participant would have both experience and comprehensive academic background in this field

The third participant (pseudonym: Charlie) has the least amount of working experience but has the most solid academic background He studied in FPT University, and then later worked for FPT Software Both of which are a branch of FPT corporation The translated version “Học sâu” is carried out by the leading group of professionals from FPT Software Therefore, Charlie should be the person who would be familiar and accept the translation the most

All participants are Vietnamese who have studied deep learning in different proportions of academic and practical knowledge acquisition The selection of participants aims to have an overall view of terminology not only in theory but also

in real practice Participants are chosen to have a diverse range of working experience and academic path and accomplishments The verification of terminology that the participants offer is expected to provide a precise and detailed range of terminology as possible

3.3 Quantitative analysis

3.3.1 Quantitative data

There are two main sources of data in the thesis, both are verified and published The main verified source is from the English version “Deep Learning" (2016) and the translation delivered by the team DLBOOKVN “Học sâu" (2022)

Moreover, proceeding from the reality of being a translator of videos, documents, and learning courses and using literal translation as the only translation procedure, I decided to carry out this thesis to study the overall translation procedures with the references of the Vietnamese version “Học sâu" (2022)

3.3.1.1 The book “Deep Learning”

"Deep Learning" by Ian Goodfellow, Yoshua Bengio, and Aaron Courville is regarded as a foundational text in artificial intelligence This comprehensive text,

often dubbed the "Deep Learning Bible," serves as both an introduction and an in-depth exploration of the fundamental concepts and techniques of deep learning The authors' biggest strength is their capacity to harmonize the theoretical foundations of neural networks and deep learning architectures with practical insights and applications, rendering it an essential resource for both novices and seasoned practitioners in the domain

The importance of "Deep Learning" to the field cannot be overstated It serves as a cornerstone reference, encapsulating the collective understanding of deep learning

at a pivotal time in its development The book has been cited in countless research papers and has played a key role in educating a new generation of AI researchers and practitioners Its comprehensive coverage of the field’s foundational concepts and advanced topics ensures its relevance and longevity as a key resource in the AI community

Yann LeCun, a prominent figure in AI and Director of AI Research at Facebook, commends the book for its comprehensive and authoritative content He highlights the authors' ability to encapsulate the depth and breadth of the field, describing the book as "the definitive textbook on deep learning." This sentiment is echoed by Elon Musk, co-chair of OpenAI and CEO of Tesla and SpaceX, who emphasises the book's status as the "only comprehensive book on the subject." Musk's comment signifies the book's unparalleled scope in covering the myriad facets of deep learning, from basic principles to advanced techniques The book's educational value is particularly lauded by Geoffrey Hinton, a pioneering figure in neural networks and deep learning Hinton asserts that “'Deep Learning' will teach generations of students how to think about deep learning," emphasising the book's long-term educational impact This statement is not only a testament to the book's depth and clarity but also to its potential to influence future AI researchers and practitioners

3.3.1.2 The translation “Học sâu” by dlbooksvn

The translated version, “Học sâu”, was performed by a group of Vietnamese engineers who specialised in and are currently working on projects related to deep learning This book was a non-profit project to establish standardized Vietnamese deep learning material and terminologies The project was started with the aim of establishing a comprehensive Vietnamese resource in the domain of Deep Learning, specifically, and Artificial Intelligence, more broadly Its mission is to serve as a valuable reference and to provide solid groundwork Orientation regarding Vietnamese documents pertaining to the future of Artificial Intelligence One additional objective of the project is to aid in the standardisation of Vietnamese terminology in the realm of Artificial Intelligence This will serve as a foundation for official technical publications within the country

3.3.1.3 Data collection process

From the above-mentioned data source, I have collected 588 words altogether from the book The authors and experts who participated in the research have designated these terms as terminologies All data collected were from the book Deep Learning and its Vietnamese version

3.3.2 Qualitative data

While quantitative data was the primary source of analysis, the research also utilized qualitative data sources from questionnaires used to validate the quality of the input data and propose explanations on terminology translations procedures applied Since the Vietnamese version was performed by a fifty-people team from a single software technology company, the qualitative data was intended as a source

of verification that the data source is commonly used by the professionals in the field in general The questionnaire output was shown in the form of questions:

- Was the term commonly used in Vietnamese translation?

- Is the Vietnamese translation of the term provided in the terminology table

an appropriate translation?

- If the term provided is inappropriate, what is the correct translation used in the field?

3.4 Quantitative and qualitative data analysis

To form a quantitative data analysis, data preparation is the first task that must be performed The data collected from the source text will be converted into interpretable and helpful information In this thesis, the data sources are Deep learning terms extracted from the book Deep learning, with various intricate and uncommon words and phrases The choice of terminologies for this research data is verified in the appendix of the translation “Học sâu” and the confirmations from the experts took part in this research The terms are then classified into two criteria: Term classification and Translation procedure utilised In Term classification criterion, data are divided into two subtypes “Technical term’ and “Descriptive term” In terms of translation procedure, all terms are categorised into different translation procedures based on the framework of Newmark (1995)

3.5 Summary of the chapter

The goal of this chapter is to outline the research methods used to answer the research questions A quantitative method study is relevant in this study, in which a discussion of the procedure, study participants, data collection, conceptual relations, and translation procedures of analysis outline the specifics of how the study has been conducted The theory translation procedures by Peter Newmark (1995) have been used to help analyse quantitative data Qualitative data was also performed as

a validation tool for the terms in this research All of those procedures have contributed to this study to produce useful translation procedures of Deep learning terminologies from English to Vietnamese, which would be useful for professionals

in the field The results of the analysis following the methodology mentioned in chapter 3

CHAPTER 4: FINDINGS AND DISCUSSION

The findings of the descriptive analysis are presented in two separate types: technical terms and descriptive terms The researcher employs Newmark's (1988) theoretical framework on translation techniques to identify the frequently used technical and descriptive terms in the classification of deep learning terminology The findings of the thesis focused on showing only the translation procedures are used when translating terms Then, the results highlighted that the procedures Shift and Couplet are more used for technical terms, whereas the procedures of literal translation or shift are preferred for descriptive terms

The research findings sufficiently address both research questions Following this investigation, some implications have been identified, although certain limitations, along with additional studies are proposed to advance the research to a more sophisticated level

This chapter outlines the investigation into two principal sections: the translation procedure for deep learning technical terminology and translation procedures for deep learning descriptive terminology, together with the application of Newmark's (1988) procedures in their translation

4.1 Translation procedure for technical terms

The figure below shows the proportions of technical terms and descriptive terms in the total of 588 words under research

Figure 3: The number of deep learning terms compared to descriptive terms

Figure 4: Proportions of translation procedures applied

4.1.1 Literal translation

Newmark (1995) presented that “literal translation is a basic translation procedure, both in communicative and semantic translation, in which translation starts from there”

Although literal translation was utilised in the translated version both in technical terms and descriptive terms, only a small portion (43 out of 588 words) of technical terms are translated using Literal translation

Short common terms with simple meanings would be the primary subject for this translation procedure, for example:

English term Vietnamese

translation

Vietnamese Expert Comment

plateaus cao nguyên

Plateaus là tình trạng hàm mất mát dường như đang ở ở một điểm cố định trong một khoảng thời gian dài, khiến cho quá trình huấn luyện mô hình chậm lại và không tiến triển

product of experts tích của các chuyên

gia

Product of Experts (Product of Experts

là một mô hình xác suất đa biến) Skip Connections

through Time

kết nối nhảy cóc xuyên thời gian Bỏ kết nối theo thời gian

Newmark (1988) proposed this translation procedure as “the process of transferring

a SL word to a TL text as a translation procedure” In this research, there are a total

of 54 technical terms that are translated using the transference procedure I divide

terms employing this translation procedure into three main categories:

4.1.2.1 Terms that derive from persons

In the field of deep learning, there are terms denoting objects, methods, codes, that derive from their inventors, discoveries; in translation proper names are transferred while the common nouns are translated to produce a more understandable version

English term Vietnamese translation

binary Boltzmann machine máy Boltzmann nhị phân

block Gibbs sampling lấy mẫu Gibbs theo khối

Gaussian mixture model mô hình Gauss hỗn hợp

Hamming distance khoảng cách Hamming

Levenshtein distance Khoảng cách Levenshtein

4.1.2.2 Terms referring to modern concepts that are not technically lexicalized in Vietnamese

According to term creation theory, whenever a new concept appears or is transferred

to a new culture, a target culture term must be created to name it; however, developing a new term is difficult and time-consuming At times, it is exceedingly challenging to find a Vietnamese term that can accurately correspond to the English term Thus, it is customary to employ loan words in translation to denote novel concepts It is important to acknowledge that the status of being a "new concept" is temporary Over time, the concept ceases to be new, while the use of loan words continues until a suitable Vietnamese term is developed and widely accepted Below are a few examples:

English term Vietnamese translation

conjugate gradient gradient liên hợp

generative moment matching network mạng sinh mẫu khớp moment

log-likelihood độ hợp lý thang log

multilayer perceptron perceptron đa tầng

negative gradient gradient đối nghịch

softmax regression classifier bộ phân loại hồi quy softmax

word embedding vector vector nhúng từ

4.1.2.3 Translating letter-contained terms

In the process of categorization, there are several terms which have one or several letters in formation These words do not have a Vietnamese specific equivalent term, or an acronym for a term mentioned and explained prior in the source text Therefore, transferring these terms are reasonable

For example, the word “BFGS” is short for “Broyden–Fletcher–Goldfarb–Shanno”, which is an algorithm The term “V-structure” is a kind of structure of nodes in a Bayesian network, with the word V coming from the basis of the structure shaped like the letter “V”

discriminative RBM RBM phân biệt

4.1.2.4 Translating abbreviations and acronyms

There are terms that are mentioned, explained, and then replaced in the following parts of the text These terms are not considered to be widely recognized and still need footnotes to provide explanation and clarification: