GOOGLE TRANSLATE PERFORMANCE IN TRANSLATING ENGLISH PASSIVE VOICE INTO INDONESIAN

: A scant number of Google Translate users and researchers continue to be skeptical of the current Google Translate 's performance as a machine translation tool. As English passive voice translation often brings problems, especially when translated into Indonesian which rich of affixes, this study works to analyze the way Google Translate (MT) translates English passive voice into Indonesian and to investigate whether Google Translate (MT) can do modulation. The data in this research were in the form of clauses and sentences with passive voice taken from corpus data. It included 497 news articles from the online news platform ‘GlobalVoices,' which were processed with AntConc 3.5.8 software. The data in this research were analyzed quantitatively and qualitatively to achieve broad objectives, depth of understanding, and the corroboration. Meanwhile, the comparative methods were used to analyze both source and target texts. Through the cautious process of collecting and analyzing the data, the results showed that (1) GT (via NMT) was able to translate the English passive voice by distinguishing morphological changes in Indonesian passive voice (2) GT was able to modulate English passive voice into Indonesian base verbs and Indonesian active voice.


INTRODUCTION
As translation, both commercial and literary, is one of several activities that have been expanding in today's globalized world (Hatim & Munday, 2004), recently, humans are not the only ones who can be trusted to translate texts. Just as technology is constantly being developed, altered, and improved; machine translation (MT) arose and became one of the options for translating text. One of the most popular machine translations is Google Translate which was developed by Google Inc. By using Google Statistical Machine Translation (GSMT) in 2006, it is possible for everyone to translate a huge amount of data by just a single click away (Garcia, 2009). Unfortunately, SMT raised various problems in translation. Specific errors on translating Source Text (ST) to Target Text (TT) are hard to predict and fix by users. Consequently, machine translation was judged to be less acceptable and inaccurate in its early days (Komeili et al., 2011). Nevertheless, In late 2016, Google Translate then adopted Neural Machine Translation (NMT) which is called as Google Neural Machine Translation (GNMT). Compared with SMT, GNMT is capable of fixing translation difficulties and threats by providing a more fluent and legible translation by handling morphology and syntax five times better than SMT systems (Ramesh et al., 2021). Thus, GNMT translations were claimed to be more precise and fluent compared to translations of SMT systems. In addition, Bahdanau et al., (2014) stated that: "Unlike the traditional phrase-based translation system which consists of many small sub-components that are tuned separately, neural machine translation attempts to build and train a single, large neural network that reads a sentence and outputs a correct translation." Google Translate, then, has a number of flaws by supporting approximately 109 languages at various levels as of April 2021. Because of its development, Google Translate has been used by over 500 million people around the world with 100 billion words translated every day in 103 languages, when human translators were judged to be more expensive and took a lot of time (Aiken, 2019). Since most people begin to use Google Translate frequently, a scant number of scholars then became skeptical and conducted additional research focusing on Google Translate (Sun, 2014) to test its performance and accuracy. Amar (2017), for example, investigated the accuracy level of Google Translate especially in translating English text into Indonesian based on language error analysis and the use of equivalence strategy. He concluded that Google Translate can only translate English source text into Indonesian correctly if the appropriate equivalence translation strategy is just literal or transposition. In the same manner, Sutrisno (2020) examines the accuracy as well as the shortcomings of Google Translate in the context of English to Indonesian translations in order to critically engage the complaints made by Google users. Both the original sentences and their translated versions were analyzed using a sentence pair matrix to determine the machine's failings and areas for improvement. Through his research, he found that Google Translate has the capability to translate English to Indonesian sentences with an accuracy level reaching 60.37. Whereas Sianipar & Sajarwa (2021), by comparing the translation of passive voice in Indonesian research abstracts into English conducted by human translation vs. machine translation (Google Translate), they concluded that human translation is better than machine translation in translating English passive voice into Indonesian. All three studies show that Google Translate still has some drawbacks when it comes to translating certain texts.
However, as Google Translate continues to develop, we continue to verify its performance by analyzing the way Google Translate translates English passive voice into Indonesian which were tested by using a news corpus data set. Meanwhile, since humans are very intensive to do modulation, this research also tends to investigate whether Google Translate can do the same thing to create natural translation. Thus, passive voice was used as the variable of this research since it was positioned as the most common structure used in the written discourse, especially in the news and scientific writings construction (Keenan, 1985). Moreover, every language has a unique and different characteristic. In contrast to English active voice which is easy to translate, English passive voice is often difficult to translate into Indonesian due to Indonesian having some different affixes to use in passive construction. Besides, the roles between actor and agent which are called subject positions in the generative grammar, also need to be considered in the sentence construction. By using this approach, this research will be beneficial in looking at technological developments from a translation point of view.

Google Translate: How Does It Work?
Google Translate is a well-known free online translation engine that can translate not only numerous words, but also phrases, text fragments, and entire web pages (Karami, 2014). Along with Google Translate prominent heights, it is expanding to over 100 languages today and is used by most internet users around the world for translating texts (Koehn, 2020). In 2016, Google Inc. expanded their quality and released a Neural Machine Translation (NMT) system, which has the potential to address many of the shortcomings of traditional SMT. End-to-end Neural Machine Translation (NMT) has become the new standard method in actual machine translation systems in recent years (Tan et al., 2020). Google NMT can also solve the notoriously difficult language pair translation problem by taking the context of a word into account rather than simply translating each individual word. Its system can reduce translation errors compared to Google SMT's phrase-based translation. However, Google NMT can still make an amount of errors in languages with productive word creation, such as compounding and agglutination (Sennrich et al., 2015). This problem was used as the basis of our research and investigation on the translation of English into Indonesian using present Google NMT.

English Passive Voice
The use of passive voice is very common in English sentences and texts as one of the most fundamental elements of the English language. When the doer of an action is unknown or insignificant, or when the focus is "on the experiment or process being described", the passive voice is utilized (Hacker, 2003). In line with the definition, according to Apandi & Islami (2018), passive voice is used when the focus of the sentence is the outcome or the person affected by the action and it is not important or known who or what is performing the action. Furthermore, the passive voice is a grammatical form in which a head noun serving as the subject of a phrase, clause, or verb is impacted or acted upon by the verb's action (Scholastica, 2018). In passive voice, there are three markers: be, -ed, and by, each with its own meaning and significance. Passive with agent and passive without agent, or agentive passive and non-agentive passive, are the two most common types of passive. The agent will not appear in the agentive passive, but will be implied in the context. The rules and usage of the passive voice differ between languages.
In English, the passive voice can be constructed in many different forms. The short dynamic be-passive pattern with '[be-verb+Past Participles (Verb 3)]' construction is the most fundamental passive pattern in English grammatical structure (Biber et al., 1999), e.g. "is stolen, was caught, were written", etc.). Nevertheless, sometimes English only used past participles to mark passive voices, e.g."The book written by the lecturer is now in the well-known publisher". In this case, the passive voice used in the sentence does not use the "be-verb" formula, but simply by using past participle verbs only used the .
Another feature of the passive in English is the use of "by phrases" at the end of the clause (for example: "The book is written by my father").

Indonesian Passive Voice
In addition, Indonesians regularly utilize passive voice as well. According to Alwi et al. (2003), there are several ways to construct passive voice in Indonesia, those are: 1) by adding prefix di-into the base verb; 2) by adding prefix ter-into the base verb; and 3) by using the verb base itself. The first and most common method of forming passive constructions in Indonesian is to use a base verb combined with the prefix di-with . This construction is commonly used if the subject/agent is a noun or noun phrase. Furthermore, if the action is unintended, the prefix ter-is used instead of di-. The construction is [prefix ter-+verb base], e.g termakan, tertabrak, etc, and [prefix ter-+ base verb + suffix -i, nya, kan], e.g tertuliskan (is written), terbawanya (is bought), terwakili (is represented), such as in the sentence "The girl was hit by a car"; the translation became " Gadis itu tertabrak mobil".
Based on the example, it means that the car accidentally hits someone.

Modulation: An Overview
Translating a text is not only a matter of finding the relevant words in the target language and applying the correct target language grammar when translating a text (Putranti, 2018), it is also a matter of generating the most natural translation of the source language message into the target language (Pinchuck in Machali, 2009;Nida & Taber, 1982). Nevertheless, creating the closest natural equivalent was not easy to handle. One of translation techniques which can be applied by translators is known as modulation (Catford, 1965;Newmark, 1988;Vinay & Darbelnet, 1955). In our research, we focused our investigation on passive voice which caused modulation.

METHOD
In this research, mixed methods were used to obtain breadth and depth of understanding, as well as corroboration. According to Nassaji (2015), qualitative data can also be analyzed quantitatively. This occurs when the researcher examines qualitative data to identify relevant themes and ideas before converting them to numerical data for further comparison and evaluation. The quantitative method by using descriptive statistics in this research was used to reveal (1) the frequency as well as the number of passive voice in the news corpus; (2) the number of affixations in the target texts (Indonesian); and (3) the number of modulations. Meanwhile, descriptive deals with qualitative method was used to describe the patterns of the translation of English passive voice into Indonesian, as well as modulation techniques conducted by Google Translate.
The data in this research were in the form of clauses and sentences containing passive voice structure taken from news corpus data which were downloaded from Parallel Global Voices (http://nlp.ilsp.gr/pgv/). The corpus data for this research was accessed in April 12 th , 2021 containing 497 news articles with a total 17,069 sentences, 665,664-word tokens; and 36,763-word types. All the data, then, were manually entered into Google Translate to serve output data in Indonesian language.

Figure 1. The translation of the sentence "It was just ten days ago that part of the country was submerged by waters" using Google Translate
A content analysis method was used to select the data from the corpus since the data were in the form of texts. Meanwhile, purposive sampling was applied focusing on the emergence of passive voice in the SL's data sets. Then, the data were input into Google Translate from April to May 2021 and were classified based on their morphological changes and the probability of modulation. To collect the data, AntConc 3.5.8 was used as the data instrument which was downloaded from https://www.laurenceanthony.net/software/antconc/. It was frequently utilized by researchers all around the world for corpus-based research tools since it was freely accessible to scholars. The overall distribution of the research items was displayed under "Concordance Plot" during the inputting process, and the particular contexts of each retrieved word were shown via "File Views". However, we discovered that AntConc 3.5.8 processed and counted some words or phrases that have the same formula as the Indonesian passive structure (such as possessive, i.g, dirinya (himself). To facilitate the analysis, the corpus data were reduced and eliminated by selecting and focusing only on sentences containing English passive voices, as well as data that were indicated to contain modulation. Using AntConc 3.5.8, the data were reduced into 1,098 data of passive sentences with 1,550 passive verbs.
As the basis in conducting this research, we stand on the theory that Indonesian is similar to English in terms of structure (S-V-O). Nevertheless, as stated by Sutrisno (2020), there are certain rules in both languages that may cause interference, such as: (1) Indonesian does not have tenses; (2)  During the analyzing process, the entire data set was analyzed and evaluated using Sudaryanto's (1993) comparative methods. Any differences and similarity in the source texts and the target texts were fully observed. As the results, two strands of research (both quantitative and qualitative) were served at the interpretation stage or discussion. Both quantitative and qualitative tend to complement each other and receive equal emphasis in the findings. Investigator triangulation by repeatedly checking the data, theoretical triangulation by linking back to some relevant theories, and methodological triangulation by using appropriate methods were employed in order to achieve credibility, dependability, transferability and conformability (Moleong, 2001). Triangulation in this research was used since subjectivity becomes one of problems during collecting and analyzing the data in the form of language, social and humanities approaches.

Findings
Through the cautious process of collecting and analyzing the data, we focused our results and discussions on the morphological changes as the impact of translation

Discussion
We segregated the discussion into four sections which were dealing with (1)

The Translation of English Passive Voice into Indonesian Passive Voice
Through the investigation, there are roughly two basic forms of passive voice used in Indonesia presented by Google Translate. The first is distinguished by the prefix di-, while the second is distinguished by the prefix ter-. The results of inputting and translating English passive voice into Indonesian passive voice are described below.

The Translation of English Passive voice into Indonesian Passive Voice, marked by Prefix di-.
From the analyses of 1,297 data of passive verbs, we found that overall data were effectively translated from the English passive voice into the Indonesian passive voice marked by prefix di-. As our consideration, we found that each verb construction is most fully transferred and there is no specific change in terms of meanings. Nevertheless, we neglected the terms of accuracy in this research along with our research limitations.
Furthermore, we also identified other Indonesian passive voices formed with the prefix [di-+root verb+suffix (-i,-kan)]. This following table showed the tendency of the translation of English passive voice into Indonesian passive voice, basically marked by prefix di-. the prefix di-is more appropriate. Therefore, it would be better to translate it as "dituduh"

Table 1. English Passive Voice (SL) and Their Translation into Indonesia Passive Voice (TL) Using [prefix di-+root verb], and [prefix di-+root verb+suffix -i, -kan] Affix(es) Source Language(s) Target Language(s) (Google
instead of "tertuduh [ter-+ root verb]". Relating to aspect, the modal word "have" in the source language is directly translated into the adverb "telah" in the target language. In line with what Alwi et al. (2003) said, adverbs in Indonesian can be used as markers of aspect, modality, quantity, and quality of the categories of verbs, adjectives, numerals, and other adverbs. Meanwhile, the adverb "telah" in excerpt (1) is a perfective aspect marker which indicates that the event has already started in the past and continues in the present. This translation process also proves that Google Translate has been using literal translation while translating passive voice with aspects.
Then, from the excerpt (2), the phrase "was not asked [to be (past)+not+V3]", (negative passive) is also found to be translated into Indonesian negative passive construction " tidak ditanyai [negation+di-+root verb+-i]". We detected that Google Translate has a tendency to translate "was not asked" into ditanyai instead of ditanya because Indonesian suffix -i can be used to change the form of a verb from intransitive to its transitive meaning. Then, in excerpt (3), the English passive verb "are reported [to be+V3]" is translated as "dilaporkan [di-+root verb+-kan]" using confix-affixes that function to form passive verbs. It has functioned to state the causative meaning of causing something to happen, and stating the meaning of an act done for someone else. So, in this case, "dilaporkan" means that the event is reported by other people to someone else.
At a glance, our findings show that Google Translate's translation of passive voice has improved significantly since its inception. Because Google Translate is educated on hundreds of millions of pre-translated words, phrases, and even material from the internet, it will operate and give the more generic translation if one version of passive voice exists several times. As a result, Google Translate's meaning is solely determined by the program's internal logic.

The Translation of English Passive Voice into Indonesian Passive Voice, Marked by Prefix ter-
When comparing Indonesian passive voice with prefix di-and prefix ter-, it is clear that the use of prefix ter-implies such unintended factors. The use of prefix terimplies that the action is done unintentionally. The passive voice that is translated into Indonesian with the prefix ter-are shown in Table 2.

Table 2. English passive voice (SL) and their translation into Indonesia passive voice (TL) using [prefix ter-+root verb], and [prefix ter-+root verb+suffix -i, -kan, -nya] Affix(es) Source Language(s) Target Language(s) (Google Translate)
ter-4) That way, object both persons will be spared from having to go through renewing or not renewing the expirable marriage license. Passive Construction: Modal Auxilary (will) + be + past participle (spared) The overall data showed that when the meaning of a passive voice verb in Indonesian includes an unintentional action, the prefix ter-is used instead of di-. As seen in the excerpt (4), GT has translated the phrase "will be spared [will+be+V3]" into "akan terhindar [modal (akan)+ter-+root verb]". It showed that since the action represented in the verb "spared" is done unintentionally, the use of the prefix ter-is more appropriate.
In this case, the word "will" in the source language is translated into "akan" in the target language. The word "akan" in Indonesian is an adverb that can function as both an aspect and a modality marker. In Indonesian, the adverb "akan" was used as an aspect marker to indicate that the event would take place in the future.
Meanwhile, in excerpts (5) and (6) (6) is also one of the various confix-affixes functioning to form passive verbs. This Indonesian passive translation "terwakili and terpinggirkan" possessed as passive stative verb which expresses the stative condition that something or someone is involved in a certain situation. Relying on the findings, the suffixes -i and -kan serve distinct functions based on the context of the sentences. Suffix -kan serves as a causal function, whilst suffix -i serves as a repetitious function.

The Translation of English Passive voice into Indonesian Base Verbs
Some English passive voice also were translated into Indonesian root verbs (without prefix di-, and ter-). Based on the analysis, we found that there is a tendency of omitting affixes in both Indonesian passive constructions. These phenomena are in line with what Alwi et al. (2003) has said that the sentences which use the root based in Indonesian are essentially as one way in expressing the passive voice in Indonesian. The examples of the data were presented in Table 3. Passive construction using root verb is commonly used in Indonesia when the verb (Prefix me-/meN-+root verb+suffix) in active voice is changed to passive by omitting its prefix and suffix, e.g Active sentence such as Saya sudah mencuci mobil [I have washed the car] change in to Passive sentence such as Mobil sudah saya cuci [The car has been washed by me]. There is a shift in perspective in this example because it is unusual to say "mobil sudah dicuci oleh saya" in Indonesia. As a result, to make the sentence construction sound natural, Indonesian use the passive construction without the prefix diand ter-.
We discovered that Google Translate was able to recognize passive constructions as well by shifting point of view from ST to TT (Vinay & Darbelnet, 1955). Took a look back into the examples, excerpts (8) and (9) show that the phrase "is published [to be (present) +V3]" is translated into "terbit [root verb]" and "was leaked [to be (present) +V3]" is translated into "bocor [root verb]". This kind of shifting was called a modulation technique.

The Translation of English Passive Voice into Indonesian Active Voice
According to the data, Google Translate intends to translate the English passive voice into Indonesian active voice. Newmark (1988) and Vinay & Darbelnet (1955) called the changes from passive form into active form as another kind of modulation. Here, Vinay & Darbelnet (1955) added modulation in order to produce the natural translation.
Hence, the modulation technique is considered to be the best option to hold the original meaning in the source language. Because Indonesian has a specific word order, this issue frequently occurs in English to Indonesian translations which were shown in Table 4. Based on our analysis, we found that the active voice is characterized with the verb preceded with the prefix ber-, me-and meN-. Here, the passive phrase "has been motivated [has+been+V3]" is translated into "bermotif [ber-+root verb]" which belongs to an active verb in Indonesian. The verb "bermotif" sounds more natural rather than "dimotifkan". The other example, the phrase "is eaten [to be (present)+V3]" also translated by Google Translate into "memakannya" [me-+root verb (active)+-nya]. This case showed that the role of context also influenced the choice of translation. It is more acceptable or natural to say "bagaimana memakannya'' instead of "bagaimana itu dimakan". Furthermore, the phrase "have been displaced [have+been+V3]" is translated into "telah mengungsi [meN-/meng-+ ungsi]" which belongs to the Indonesian conditional allomorph of MeN-. Overall, Google Translate tends to convert active translation into passive formulations because the action is carried out by the sentence's agent. The current study showed that Google Translate intends to use modulation to resolve some translation issues, particularly when producing natural translation.

Conclusios
From the depth analysis, we conclude that Google Translate using Neural Machine Translation (NMT) was able to translate English passive voice into Indonesian by distinguishing morphological changes in Indonesian passive voice through the use of affixes (such as the use di-, di-kan, di-i, ter-, ter-i, ter-kan, ter-nya). Furthermore, we also evaluate that today, Google Translate by the performance of NMT was able to modulate or change English passive voice into Indonesian active voices appropriately given their context by using some kinds of affixes, such as ber-, and me-, meN-or by the change the point of view using root (base) verbs in Indonesia passive constructions. Thus, our findings then were used as a follow up from the parliamentary findings which concluded that a statistical machine translation (SMT) does not yet have the capability of modulation.

Suggestions
Along with the rapid development and improvement of Google Translate, this means that the scope of this field of study is overly vague. Hence, this study provides an opportunity for future researchers to expand further research on the performance of Google Translate over time, for instance; seeing the accuracy level of passive voice translation conducted by Google Translate or by testing using other types of sentences. It is deemed essential to test Google Translate's accuracy in translating English passive voice into Indonesian using accuracy evaluation methods such as manual or automatic evaluation (e.g., BLEU (Bilingual Evaluation Understudy) scores (Aiken, 2019;Ramesh et al., 2021), CompareMT (Neubig et al., 2019), MTComparEval, Memsource criteria (see www.memsource.com), translation closeness metric, and etc.).