The Creativity and Limitations of AI Neural Translation

The creativity and limitations of AI neural
machine translation
A corpus-based study of DeepL’s English-to-
Shanghai Jiaotong University (jbid141041) IP: 180.166.197.110 On: Sun, 20 Aug 2023 03:27:24
Chinese translation of Shakespeare’s plays
Hu Kaibao and Li Xiaoqian

Shanghai International Studies University
This study examines the performance of the neural machine translation

system DeepL in translating Shakespeare’s plays Coriolanus and The
Merchant of Venice. The aim here is to explore the strengths and limitations
of an AI-based English-Chinese translation of literary texts. Adopting a
corpus-based approach, the study investigates the accuracy and fluency
rates, the linguistic features, and the use of various methods of translation in
the Chinese translations of Shakespeare’s plays conducted via DeepL. It
compares these to the translations by Liang Shiqiu, a well-known Chinese
translator. The study finds that DeepL performs well in translating these
works, with an accuracy and fluency rate of above 80% in sampled texts,
showing the potential of the use of neural machine translation in translating
literary texts across distant languages. Our research further reveals that the
DeepL translations exhibit a certain degree of creativity in their use of trans-
lation methods such as addition, explicitation, conversion and shift of
perspective, and in the use of Chinese sentence-final modal particles, as well
as Chinese modal verbs. On the other hand, the system appears to be
limited in that a certain amount of translation errors are present, including
literal translations.
Keywords: AI, translation, corpus, limitations, Chinese, Shakespeare,

neural machine translation, DeepL
1. Introduction
Machine translation (hereafter MT) refers to using computer software to auto-

matically translate a text or a speech from one language to another language. It
has gone through three stages of development: rule-based MT, statistical MT, and
https://doi.org/10.1075/babel.00331.hu | Published online: 24 July 2023

Babel ISSN 0521-9744 | E‑ISSN 1569-9668
© 2023 Fédération Internationale des Traducteurs (FIT) Revue Babel
[2] Hu Kaibao and Li Xiaoqian
neural MT. Rule-based MT is the oldest approach among these and relies on a
set of linguistic rules (morphological, syntactic, and semantic rules), as well as
contextual knowledge of both the source and target language (Al-Batineh and
Rabadi 2019). Rule-based MT operates on the premise that the source text is run
through a process of analyzing and parsing, to then render the source text into
the target text. Statistical MT is considered as a second-generation type of trans-
lation in that it relies on statistically analyzing previously translated texts (so-
called parallel texts). It then generates target texts “based on a probabilistic model
of the translation process, the parameters of which are estimated from parallel
texts” (Yang and Min 2015, 201). Finally, neural MT, also known as AI translation,
uses neural network models to learn statistical models for machine translation. It
attempts to build and train a single, large neural network that reads a sentence and
outputs a correct translation. As opposed to rule-based MT and statistical MT,
neural MT does not run a set of predefined rules but rather takes into account
the whole input sentence, and incorporates context from the training data, source
text, and unfolding target text to deliver a fast and accurate translation.
With the advances made in the past few years, neural MT has been widely
used in translating non-literary texts, including general texts, news texts and
legal texts. However, the question of whether a neural MT system can be used
in translating literary texts has been controversial. Besacier (2015) argued that
MT, followed by post-editing, can be a useful low-cost alternative for translating
literary works, albeit at the expense of translation quality. On the contrary, Voigt
and Jurafsky (2012, 18) claimed that the use of MT in literary translation might
seem a “near-contradiction in terms,” because MT has traditionally been used for
informative texts only.
In view of these differing accounts, our article assesses the performance of
DeepL, a neural MT system, and does this by drawing upon corpus methodology.
We aim to check what strengths and weaknesses neural MT displays in translating
Shakespeare’s plays from English to Chinese in order to assess whether it is suit-
able to the translation of literary texts. It also aims to explore whether, and if so
to what extent, the translations of literal texts by an AI translator are limited or,
in contrast, may indicate a measure of creativity compared to human translation.
In this article, ‘limitations’ in translation refer to the fact that a translator tends to
follow the structure of the source text closely and translates it literally or word-
for-word as a result of being unable to take account of its linguistic context or
cultural context, which in general leads to ungrammatical or unidiomatic target
texts. Creativity in translation here denotes the fact that a translator renders or
makes explicit the meaning, style, and feelings of the source text by using a variety
of translation methods other than literal translation.
The creativity and limitations of AI neural machine translation [3]
In this article, we will compare the Chinese translations of Shakespeare’s plays

Coriolanus and The Merchant of Venice by DeepL on the one hand, and by Liang
Shiqiu, a renowned Chinese translator, on the other hand. To discuss the perfor-
mance of DeepL, we have considered the translation errors and identified the
linguistic features and the use of translation methods. These are then compared
to the same features in the translations by Liang Shiqiu. Therefore, this article
addresses the following research questions: (1) Can MT be successfully applied
to the translation of literary texts? (2) Do the translations by an AI literary trans-

lator or a neural MT system exhibit signs of creativity or are they limited in some
manner? (3) If the analysis indicates the presence of creativity, what is its nature?
2. MT in the translation of literary texts
In the past decade, several studies have been conducted to explore the applica-
bility of MT to literary translation. One of the earlier studies of this kind includes
Jones and Irvine (2013), which used existing MT systems to translate samples of
French literature (prose and poetry) into English. This study is of relevance to the
current study because the authors used qualitative analysis grounded in transla-
tion theory on the MT output in order to determine the potential of MT in literary
translation and to address what makes the translation of literary works particu-
larly difficult. They argued that the main errors involved using English syntactic
structures and expressions, instead of their French equivalents, and not consid-
ering specific cultural references.
Toral and Way (2015) examined whether MT can be used to translate literary
texts by analyzing parallel corpora and measuring the degree of freedom of the
translations and the narrowness of the domain. The study revealed that MT could
be helpful in this regard.
In Matusov’s research (2019), an analysis was conducted on the translations
of fiction stories by neural MT systems. These literary texts were translated from
English to Russian and German to English. The analysis showed that these trans-
lations have a richer vocabulary, up to 30% of machine-translated sentences have
acceptable quality, and there are very few severe syntactic errors, even in complex
sentences in the translations.
More recently, Webster et al. (2020) analyzed the output of the neural
machine system of Google Translate and DeepL when applied to four classic
novels translated from English into Dutch. The analysis revealed that a large
proportion of the translated sentences contained errors. Moreover, lexical rich-
ness and local cohesion were lower in the neural MTs than in human translations.
It is important to note that these studies have shown that MT can indeed be
applied to literary texts. However, the data analyzed in these studies have been
limited to translations between closely related languages, in particular between
European languages. Few studies have explored the possibility of using MT in
translating literature across languages that are more distant. Thus, it remains
unclear whether MT can be used in translating literature across languages that
are more distant and linguistically remote from each other, such as Chinese and
English, and whether an AI literary translation or literary translation by neural

MT can exhibit a degree of creativity in the translation it produces or may be
limited in other ways.
3. Research design
3.1 The corpus for the research
As mentioned, this article aims to investigate the performance of DeepL in the

translations of literary texts from English into Chinese by using corpus method-
ology. The data is drawn from the English-Chinese Parallel Corpus of The
Merchant of Venice and Coriolanus. The corpus thus comprises these two plays by
Shakespeare, their corresponding Chinese translations by Liang Shiqiu, and the
translations by DeepL. The Merchant of Venice and Coriolanus were selected as
the research data because these specific plays can be assumed to be representative
of Shakespeare’s work to a certain extent. In addition, the former is a comedy and
the latter a tragedy, which also adds to their overall representativeness of Shake-
speare’s body of work. Further, the translations by Liang Shiqiu were selected
because they are similar to the translations by DeepL in adopting literal transla-
tion, as revealed by Hu (2015). Table 1 provides the details of our corpus.
Table 1. The English-Chinese Parallel Corpus of The Merchant of Venice and Coriolanus
Files Tokens Types TTR STTR STTR Std. dev. STTR basis
Source text 52475 5625 10.72 40.75 56.32 1000.00
DeepL 58722 5870 9.49 39.64 58.01 1000.00
Liang 62266 7424 11.92 43.61 54.89 1000.00
Table 1 clearly shows that the number of tokens in the translations by Liang
Shiqiu exceeds the number of tokens in the translations by DeepL, with 62,266
compared to 58,722. The STTR of the latter is 39.64 per 1,000 words, which is
lower than that of the former (43.61 per 1,000 words). Comparatively, the transla-
tions by Liang Shiqiu have a lower lexical variety than the translations by DeepL.
This is confirmed by Vanmassenhove et al. (2019), who argues that MT causes a
loss in terms of lexical richness and diversity when compared to human-generated
or human-translated texts.
3.2 Research procedures

To explore the possibility of using MT in literary translation and examine possible

levels of creativity as well as limitations in the translations by DeepL, we
conducted three types of analyses. First, an error analysis was conducted on
the Chinese translations by DeepL. This type of analysis, which is frequently
used for assessing translation quality, aims to answer the first research question
we formulated. What is the potential of MT for literary translation? Second, in
order to examine whether creativity is exhibited in the Chinese translations by
DeepL, or whether and to what extent these are limited somehow, a corpus-
based analysis was conducted. Its aim was to identify the linguistic features of the
translated texts by DeepL, as compared to the translated texts by Liang Shiqiu.
The use of sentence-final modal particles and modal verbs was also investigated
here. Sentence-final modal particles and modal verbs are used to convey mood,
attitude, or feelings. Analyzing their use in translated texts may therefore reveal
whether the target text has faithfully reproduced the mood, attitude or feelings
of the source text and thus provide clues as to whether a translator is subject to
the constraints of the source text or exercises his or her creativity in translation.
And finally, the translation methods adopted by DeepL and Liang Shiqiu, respec-
tively, were examined based on the analysis of some translation examples. This
type of analysis may provide direct evidence as to whether the translations by
DeepL exhibit a sense of creativity, and/or in what way they may be limited.
3.3 Theoretical framework
In this study, we have analyzed the use of Chinese sentence-final modal particles,
modal verbs, and translation methods. We address these below.
3.3.1 Sentence-final modal particles

A modal particle, which is a type of grammatical particle, is used to reveal how the
speaker thinks that the content of the sentence relates to the participants’ common
knowledge, or how it adds mood to the meaning of a sentence. In the Chinese
language, sentence-final modal particles include 啊, 吧, 啦, 吗, 呢, and 呀. Among
these modal particles, 吗 and 呢 are used to indicate doubt or uncertainty. 吧 and
啦 are not only used to express doubt or uncertainty but also to convey a request
or an order, whereas 啊 and 呀 express exclamation.
3.3.2 Modal verbs

Modal verbs are part of the modal system which includes modal adjectives,
modal adverbs, and personal pronouns. According to Palmer (1990), modal verbs
express three kinds of modality: epistemic, deontic, and dynamic modality. Epis-
temic modality is concerned with making judgments about the truth of a propo-
sition. Deontic modality refers to the subject’s ability to influence actions, states,
events, permissions, or directives. And finally, dynamic modality is “concerned
with the ability or volition of the subject of the sentence, rather than the opinions
(epistemic) or attitudes (deontic) of the speaker” (Palmer 1990, 36). Moreover,
modal verbs can be used to express a judgment on different degrees of truth-value
or obligation. Their use can commit the addressee in varied degrees to a course of
action. Halliday (2000: 76) proposes a system of three degrees of value for modal
verbs: high-value, median-value, and low-value modal verbs. English language
high-value modal verbs include must, ought to, need and have to. Median-value
modal verbs involve will, shall, and should, whilst can, could, may, and might are
low-value modal verbs.
Chinese language modal verbs include 要 (will), 可以 (can), 能 (can), 能
够 (can), 会 (can), 必须 (must), 应 (should), 应该 (should), and 应当 (should).
Among them, 要 (will), 可以 (can), 必须 (must), 应 (should), 应该 (should),
and 应当 (should) express deontic modality. 会 conveys epistemic modality,
while 能 (can) and 能够 (can) indicate dynamic modality. 必须 is a high-value
modal verb, 要, 应, 应该, and 应当 are median-value modal verbs, and 可以,
能, 能够, and 会 are low-value modal verbs. Generally, Chinese modal verbs are
rendered from equivalent English modal verbs, but they can also be translated
from non-equivalent English modal verbs, or from other structures. For example,
the Chinese modal verb 应该 in the Chinese translation of Shakespeare’s play The
Merchant of Venice by DeepL is not only rendered from its English equivalent –
the modal verb “should” – but is also translated from the English modal verb
“shall,” the verb “deserve,” and other structures. The non-equivalent rendition of
modal verbs involves the use of free translation and thus reveals a certain degree
of creativity in translation, which will be discussed in more detail in Section 4
below.
3.3.3 Translation methods

Translation methods primarily include word-for-word translation, literal trans-
lation, and free translation. Word-for-word translation is a method whereby the
source text is rendered into the target text mechanically, which means it is trans-
lated without changing the word order of the source text. Literal translation
refers to a method whereby a text is rendered from one language into another
by following the source language closely. This means that, unlike word-for-word
translation, literal translation provides a fluent and accurate translation that the
readers of the target language can more easily comprehend. Free translation, then,
is a method used to deliver the intended meaning of the source text by over-
coming the constraints imposed by the structure of the source text and going
beyond its literal meaning. It involves such translation methods as addition,

explicitation, conversion, shift of perspective, and other methods. Generally, the
use of free translation signifies more creativity in how something is translated
than the use of word-for-word translation and literal translation. In contrast, the
use of word-for-word translation reveals much more constraints in translation
than the use of other translation methods.
4. Results and discussion
4.1 Errors in the Chinese translations of Shakespeare’s Plays by DeepL
Errors in translation may be categorized as being either accuracy errors or fluency

errors. Accuracy errors can be identified by comparing the source text and the
target text, whereas fluency errors can be detected by looking at the target text
alone. An accuracy error refers to the translator’s failure to reproduce the
meaning, style, or attitude of the source text in a faithful manner. An accuracy
error is classified into lexical error and sentential error. The former refers to
the translator’s failure to provide a correct translation for the source-language
word, whereas the latter denotes the fact that the meaning of the source-language
sentence is not faithfully reproduced. A fluency error refers to the translator’s
failure to produce a grammatical, idiomatic, and fluent target text.
To explore and assess the performance of DeepL in translating Coriolanus
and The Merchant of Venice, we conducted a manual analysis of the translation
errors occurring in the Chinese translations of Act I and Act V of Coriolanus
and Act I, Act IV and Act V of The Merchant of Venice. Around 900 sentences
and their Chinese translations were selected for the analysis from the beginning
and the end of the two plays respectively. For Coriolanus, the number of selected
sentences was 811, accounting for 31% of the total number of the sentences in
the play, with 346 sentences from Act I, and 465 sentences from Act V. For The
Merchant of Venice, the number of the selected sentences was 993, amounting to
38% of the total number of the sentences, with 349 sentences from Act I, and 644
sentences from Act IV and Act V.
The analysis reveals that both accuracy and fluency errors are present in these
Chinese translations.
Table 2. Errors in the Chinese translations of Shakespeare’s Plays by DeepL

Coriolanus The Merchant of Venice
Item Act I Act V Act I Act IV-V
Number of sentences 346 465 349 644

Accuracy Lexical error 41 67 30 46
Syntactic error 6 20 25 29
Error rate 13.58% 18.7% 15.76% 11.65%
Accuracy rate 86.42% 81.3% 84.24% 88.35%
Fluency Fluency error 32 75 33 65
Error rate 9.25% 16.13% 9.46% 10.09%
Fluency rate 90.75% 83.87% 90.54% 89.91%
As indicated in Table 2, 47 out of 346 sentences and 87 out of 465 sentences of

the Chinese translations by DeepL of Act I and Act V of Coriolanus contain accu-
racy errors. This accounts for 13.58% and 18.7%, respectively. The accuracy rate
here is 86.42% and 81.3%. In addition, 32 out of 346 sentences and 75 out of 465
sentences contain fluency errors, which makes for 9.25% and 16.13% respectively.
The fluency rate is 90.75% and 83.87%.
In the Chinese translations by DeepL of Act I and Act IV-V of The Merchant
of Venice, 55 out of 349 sentences and 75 out of 644 sentences contain accuracy
errors, accounting for 15.76% and 11.65%. The accuracy rate is 84.24% and 88.35%.
Moreover, 33 out of 349 sentences and 65 out of 644 sentences have fluency errors,
taking up 9.46% and 10.09%. The fluency rate here is 90.54% and 89.91%. These
percentages are both surprisingly high when compared with van Brussel et al.’s
(2018) work on general domain neural machine translation from English into
Dutch, which indicates that 33% of their neural MT sentences were error-free.
Given the high accuracy and fluency rates, DeepL seems to perform well in trans-
lating Shakespeare’s plays. Thus, neural MT can be claimed to be useful, at the
very least, or even successful in the translation of literary texts.
Table 2 further shows that DeepL tends to commit lexical and syntactic
errors. In looking deeper into this, our analysis detected more lexical errors than
syntactic ones. Upon probing the lexical errors in the translations of Act I and
Act V of Coriolanus and The Merchant of Venice, we found that most of these
were caused by DeepL’s failure to take account of, and correctly understand, the
linguistic and cultural contexts pertaining to the use of a word. For instance,
“call” in the English sentence “Who bids thee call?” was rendered into its Chinese
equivalent 电话 [a telephone call], but of course this was wholly erroneous as,
for starters, the telephone was not invented when Shakespeare wrote his plays.
Instead, “call” in the English sentence here means “ask somebody to come by
shouting or speaking.” Thus, “call” should be translated into 喊 or 叫.
4.2 Linguistic features of the Chinese translations by DeepL

4.2.1 The use of Chinese sentence-final modal particles

We compared the use of Chinese sentence-final particles in the translations by
DeepL and by Liang Shiqiu. Six Chinese sentence-final modal particles were iden-
tified in the corpus.
Table 3. Chinese sentence-final modal particles identified in the corpus

Chinese sentence-final modal particles DeepL Liang Shiqiu
啊 72 122
呀 0 65
吗 188 72
呢 58 144
吧 376 407
啦 1 141
Total Frequency 695 951
Frequency per 10000 words 11.84 15.27
Chi-square 26.6147
p 0.000
As shown in Table 4, Chinese sentence-final modal particles occur with a

frequency of 11.84 times per 10000 words in the translations by DeepL, compared
to a frequency of 15.27 times per 10000 words in the translations by Liang. The
frequency of these particles in Liang’s work is significantly higher than in DeepL’s
translation, but the frequency of ma is higher in the DeepL output, as compared
to the human translator’s work. 呀 and 啦 occur 122 times and 141 times in the
latter, but 0 and 1 time in the former. It should be noted that Chinese sentence-
final particles were often added in the translations by DeepL, since there is no
English counterpart for them. These particles are used to reproduce doubt, uncer-
tainty, request, or exclamation conveyed by the syntactic structure of the source
text, thereby making explicit the mood, attitude or feeling that is implicit in the
source text.
We analyzed the correspondence between the sentences containing the above

particles in the translations by DeepL and their English source text, in order
to examine whether the translations by DeepL had reproduced the feelings and
attitudes of the source text. We found that 啊 is generally used in the Chinese
sentences rendered from English exclamatory sentences, whereas 吗 and 呢 are
often used in the Chinese sentences translated from English interrogative
sentences. However, 呢 (ne) is sometimes used in the Chinese sentences rendered
from English affirmative sentences in order to emphasize a fact or a point of view.

This is exemplified in the following excerpt:
(1) ST: Aufidius, I am attended at the cypress grove: I pray you —
DeepL: 奥菲迪乌斯我在树林里等着你呢。我请你……
Back translation: Aufidius, I am in the forest waiting for you. I invite you…
Liang Shiqiu: 奥有朋友在柏树林里等候着我。
Back translation: Oh, there are friends in the cypress waiting for me.
In Example (1) above, 呢 is featured in a Chinese affirmative sentence in the

Chinese translations by DeepL to emphasize the fact that “I’m waiting for you at
the cypress grove.” In contrast, this article is not found in the Chinese translations
of the above example by Liang Shiqiu. Overall, DeepL has managed to faithfully
reproduce or make explicit the feelings or attitudes implicit in the source text by
adding the particle. However, before selecting appropriate sentence-final modal
particles to reproduce these feelings or attitudes, a translator has to determine
what kind of feelings, attitude, mood, or interpersonal relationship the source text
conveys, since there is no English counterpart for these modal particles. In partic-
ular, a translator has to analyze what feeling or attitude is implicit in the source
text first, whether this be doubt, surprise, or request, and he or she has to assess
this based on the analysis of the context of the source text. Then, he or she has
to decide which sentence-final particle is to be added to the target text. In doing
so, the translator tends to go beyond the restrictions imposed by the source text
and adopts addition and explicitation. Arguably, the use of Chinese sentence-final
modal particles here suggests a degree of creativity in the Chinese translations by
DeepL.
4.2.2 The use of Chinese modal verbs

We examined the use of modal verbs in Chinese translation by DeepL and Liang
Shiqiu. Eight Chinese modal verbs were identified in the corpus, and the corre-
spondence between these words and their English source text was analyzed, as
shown in Table 4.
Table 4. Chinese modal verbs identified in the corpus

Chinese Frequency Frequency
modal The (frequency per of Percentage of Chinese modal
verbs translator 10000 words) addition verbs rendered by using addition
应该 DeepL 97(16.52) 10 10.30%
(should) Liang 23(3.69) 10 43.48%
Shiqiu
应当 DeepL 4 (0.68) 3 75%
(should) Liang 4 (0.64) 2 5%
Shiqiu
应 DeepL 7 (1.19) 0 0
(should) Liang 9 (1.45) 7 78%
Shiqiu
能够 DeepL 8 (1.36) 1 12.5%
(can) Liang 12 (1.93) 2 16.67%
Shiqiu
能 (can) DeepL 177 (30.14) 40 22.60%
Liang 172 (27.62) 85 49.42%
Shiqiu
可能 DeepL 47 (8.00) 12 25.53%
(may/ Liang 21 (3.30) 12 57.14·%
might) Shiqiu
必须 DeepL 76 (12.94) 5 65.79%
(must) Liang 44 (7.07) 5 11.36%
Shiqiu
会 (will) DeepL 410 (69.82) 103 25.12%
Liang 163 (26.18) 54 33.13%
Shiqiu
Total DeepL 826 (140.66) 174 21.07%
Liang 448 (71.95) 177 39.50%
Shiqiu
As shown in Table 4, Chinese modal verbs occur 826 times with a frequency
of 140.66 times per 10,000 words in the translated texts produced by DeepL,
compared to 448 times with a frequency of 71.95 times per 10,000 words in the
translations by Liang Shiqiu. The frequency of use of all modal verbs is higher in
the DeepL translation as compared to the human translation, except in the cases
of the modal verbs 应 (should) and 能够 (can). The frequencies of 应 (should)

and 能够 (can) are 1.19 per 10,000 words and 1.36 per 10,000 words in DeepL’s
version, but 1.45 and 1.93 per 10,000 words in Liang’s translation. This suggests
that, comparatively, DeepL tends to use Chinese modal verbs more frequently
than Liang to convey the interpersonal meaning of the source text. However, the
percentage of Chinese modal verbs rendered by using addition is higher in the
translations by Liang Shiqiu than in the translations by DeepL. In the former,
177 Chinese modal verbs without corresponding forms were added, accounting
for 39.50% of the total number of Chinese modal verbs. In the latter, 174 out of
448 modal verbs were rendered by using addition, accounting for 21.07%. To be
specific, 应该 (should), 应当 (should), 能够 (can), 能 (can), 可能 (may/might),
必须 (must) and 会 (will) occurred with a higher frequency in the translations
by DeepL than in the translations by Liang Shiqiu, but the percentages of 应
(should) and 能够 (can) were higher in the latter than in the former. Overall,
both DeepL and Liang Shiqiu tended to add Chinese modal verbs in the Chinese
translations from English. To choose appropriate Chinese modal verbs for the
translation of English modal verbs, a translator has to determine what kind of
modality is expressed by the English modal verbs, whether it is epistemic, deontic,
or dynamic modality, and what degree of values is conveyed by these modal verbs,
whether it is low value, median value or high value. Moreover, before s/he adds
appropriate Chinese modal verbs to the target text, a translator has to decide what
kind of feelings or attitudes are conveyed by the source text. In doing so, the trans-
lator is not bound or constrained by the construction or literal meaning of the
source text but bases the translation on his or her holistic understanding of the
contexts and intrinsic meaning of the source text. As a matter of fact, the use of
Chinese modal verbs in the translations by DeepL, be it rendered from English
modal verbs or added to the target text, suggests a clear sense of creativity in
its translation, although it is not as noticeable as in the translated texts by Liang
Shiqiu.
4.3 The use of translation methods in the translations by DeepL
In the above sections, the use of literal translation and addition in the Chinese
translations by DeepL has been examined by analyzing the use of sentence-final
modal particles and modal verbs. The analysis shows that the overwhelming
majority of modal verbs have been rendered literally from the English source text,
whilst all the sentence-final modal particles have been added to the target text.
Clearly, limitations and creativity co-exist in the translations by DeepL, with the
former manifested by the use of literal translation, and the latter evidenced by the
use of addition.
In this section, we investigate the use of translation methods by DeepL as

compared to Liang Shiqiu, based on the analysis of some translation examples
retrieved from the English-Chinese Parallel Corpus of The Merchant of Venice
and Coriolanus. We found that explicitation, conversion and shift of perspective
have been used in the translations by DeepL.
4.3.1 Explicitation
Explicitation refers to making explicit the meaning implied in the source text
through lexical means in the target text. This method can be used to make an
expression in translated texts clear and easy to comprehend, but it requires more
processing effort on part of the translator. In particular when using this method,
a translator is supposed to determine what the implied meaning of the source text
is, including the pragmatic and metaphorical meanings, as well as the referent
of pronouns. Following this, the translator needs to decide how to make this
meaning explicit by lexical means in the target text.
(2) ST: If I can catch him once upon the hip, I will feed fat the ancient grudge I
bear him.
DeepL: 如果我能够抓到他的把柄，我就会把我对他的古老怨恨喂得肥肥
的。
Back translation: If I could catch his fault, I would feed my ancient grudge
toward him fat.
Liang: 如期我能一旦抓到他的后腰，我要痛痛快快的报这一段旧仇。
Back translation: As expected, once I can catch his hip, I will in a straightfor-
ward manner repay this old vendetta.
In Example (2), Liang Shiqiu rendered the English expression “catch him once
upon the hip” literally into the Chinese expression 抓到他的后腰 (grab the back
of his waist). The Chinese expression is clear and definite in meaning but it
does not constitute a logical relationship with the Chinese sentence that follows
我就会把我对他的古老怨恨喂得肥肥的. Unlike Liang Shiqiu, DeepL rendered
“catch him once upon the hip” into 抓到他的把柄, which makes explicit the prag-
matic meaning of the English expression, which is “have got something on him”
or “find his weakness.” This example supports the argument that the use of explic-
itation in the translations by DeepL has produced a logical and coherent target
text, while, in contrast, the use of literal translation in the translations by Liang
Shiqiu has led to an incoherent target text.
4.3.2 Conversion
Conversion is the method by which the source-language word belonging to one
part of speech is rendered as a target-language word that belongs to another
part of speech. Alternatively, a source-language clause is translated into a target-
language word or a phrase. In most cases, this method is used to produce a more
idiomatical, fluent, or vivid target text, but it is also used in cases when rendering
a source-language word into a target-language word without changing parts of
speech may produce an ungrammatical or unidiomatic target text due to the

differences between the source language and the target language. Generally, an
English noun comprising of a verb and the suffix “-er” not only refers to a person
but also to an action or activity, whereas a Chinese noun composed of a verb and
suffix 者 or 家 merely refers to a person. Given this difference, an English noun
consisting of a verb and the suffix “-er” can be rendered by using conversion as
a Chinese verb describing an action or activity. For instance, in the translations
of The Merchant of Venice by DeepL and Liang Shiqiu the sentence “In truth, I
know it’s a sin to be a mocker.” is rendered into 说实话，我知道嘲弄人是一种
罪过。 by DeepL versus 凭良心讲，我知道取笑人家是一椿罪过。 by Liang,
where mocker was translated into the Chinese verb 嘲笑 and 取笑, which means
“to mock.”
(3) ST: they have had inkling this fortnight what we intend to do, which now we’ll
show’em in deeds.
DeepL: 这两星期来，他们已经隐约知道我们的打算，现在我们要用行动
告诉他们。
Back translation: For two weeks, they have vaguely known our plan, now we
should use actions to tell them.
Liang：我们要做的事他们在两星期前就已得到了消息，现在我们要以
行动向他们表示。
Back translation: What we will do they two weeks ago have got a message.
Now we will show it to them by action.
In Example (3), the English verb + noun structure, that is “has inkling,” was
rendered into the Chinese adverb + verb structure 隐约知道 (vaguely know) by
using conversion in the translations by DeepL. Here, 知道 (know) is a Chinese
verb, and 隐约 (vaguely) functions as an adverb modifying the verb. Moreover,
the nominal clause “what we intend to do” was rendered as a noun phrase, that is
我们的打算 (our plan). In the translations by Liang Shiqiu, however, the English
verb + noun structure “has inkling” was translated literally into the Chinese verb
+ noun structure 得到消息 (get the message), and the clause “what intend to do”
was literally rendered into the Chinese nominal clause 我们要做的事 (the thing
we will do).
4.3.3 Shift of perspective

Perspective refers to the angle from which a person, a thing or a topic is described
or discussed or the focus of a discussion. By shift of perspective, we mean the
translation method by which the perspective or the focus of a statement or a
discussion in the source text is changed into another perspective or focus in the
target text. This method is often used when the translator wants to emphasize a
topic of discussion that is different from that of the source text. This is shown in
the following example:
(4) ST: Our business is not unknown to the senate.
DeepL: 元老院对我们的事并不陌生。
Back translation: The senate to our business is not strange.
Liang：我们的事元老院不是不知道。
Back translation: Our business the senate is not unknown.
In Example (4), “our business,” which functions as the subject of the source text,
is the perspective for the source text. However, in the translation by DeepL, the
expression 元老院 functions as the subject and constitutes its angle or perspec-
tive, thus enabling a shift of perspective from “our business” to “the senate.” Unlike
DeepL, Liang Shiqiu translated the source text literally, whereby the perspective
remained the same as in the target text.
Such differences might be explained by Liang Shiqiu’s concern for faithfulness
to the source text and the substantial advances in neural MT technology. In
the preface to the Chinese version of the Complete Works of Shakespeare, Liang
asserted that his main purpose in translating Shakespeare’s plays is to arouse the
readers’ interest in the source text (Liang 1968, 18). Thus, Liang gives priority to
being faithful to the source text and follows the structure of the source text closely.
Unlike Liang, DeepL does not have an explicit translation purpose. After incor-
porating context into its model for translation, it has learned to rearrange the
perspective of the source text from huge amounts of language data.
5. Conclusion
This study has investigated the limitations and possible creativity displayed by
AI-based machine translation in translating literary works by Shakespeare. It
has done this by considering different types of translation errors, the error rate,
linguistic features, and the use of translation methods in the Chinese translations
by DeepL in comparison to the translations by Liang Shiqiu. Our analysis of the
DeepL translations has revealed that the accuracy and fluency rates are above
80% in the translations of Act I and Act V of Coriolanus, and those of Act I and
Act IV-V of The Merchant of Venice. Modal verbs are more frequently used in the
translations by DeepL than in the translations by Liang Shiqiu, while sentence-
final modal particles are more frequently used in the DeepL translation versus
the human translation expert. DeepL tends to employ literal translation, but uses
addition, explicitation, conversion, and shift of perspective now and then. This
means that in contrast to human translation, neural MT is subject, at least to a
greater extent, to the restrictions of the syntactic structure of the source-language
sentence. This is so since the contextual factor that may impact its use of trans-
lation methods is merely linguistic context at the sentence level. In most cases, it
cannot make adjustments or alterations to the source text based on the analysis
of the linguistic context beyond the sentence level or the cultural contexts of the
source text. In the English translations by DeepL, this has led to the frequent use
of literal translation and the resulting high occurrence of modal verbs rendered
from their English equivalents. We argue that neural MT can thus be used in
literary translation. The translation errors and the frequent use of literal transla-
tion in the translations by DeepL indicate that it is to some extent limited when
it comes to translating literary works. However, the use of addition, explicitation,
conversion, and shift of perspective demonstrates a certain degree of creativity in
its translations, although these translation methods are used less frequently than
literal translation. It is hoped that these findings may contribute to a better under-
standing of the limitations and the degree of creativity of an AI literary trans-
lator. Arguably, future studies in related areas can be conducted on the linguistic
features of the translations by an AI literary translator, based on the analysis of the
use of syntactic structures and discourse markers.
References
Al-Batineh, Mohammed, and Reem Ibrahim Rabadi. 2019. “Will the Machine Understand
Literary Translation? A Glimpse of the Future of Literary Machine Translation through
the Lenses of Artificial Intelligence.” Studies in Translation 5 (1): 151–169.
Besacier, Laurent, and Lane Schwartz. 2015. “Automated Translation of A Literary Work: A
Pilot Study.” In Proceedings of NAACL-HLT Fourth Workshop on Computational
Linguistics for Literature, 114–122. Denver, CO: Association for Computational
Linguistics.
Halliday, Michael, and Alexander Kirkwood. 2000. An Introduction to Functional Grammar.
Beijing: Foreign Language Teaching and Research Press.
Hu, Kaibao 胡开宝. 2015. Jiyu yuliaoku de Shashibiya xiju hanyi1yanjiu 基于语料库的莎士比
亚戏剧汉译研究 [A corpus-based study of the Chinese translations of Shakespeare’s
plays]. Shanghai: Shanghai jiaotong daxue chubanshe.
Jones, Ruth, and Ann Irvine. 2013. “The (Un)Faithful Machine Translator.” In Proceedings of
the 7th Workshop on Language Technology for Cultural Heritage, Social Sciences, and
Humanities, edited by Piroska Lendvai and Kalliopi Zervanou, 96–101. Sofia: Association
for Computational Linguistics.
Liang, Shiqiu 梁实秋. 1968. “Guanyu Shashibiya de fanyi” 关于莎士比亚的翻译 [On the
translation of Shakespeare’s plays]. In Shashibiya danchen sibai zhounian jinian ji 莎士比

亚诞辰四百周年纪念集 [An anthology in honor of the 400th anniversary of
Shakespere’s birth], edited by Shiqiu Liang, 562. Taipei: Guoli bianyi guan.
Matusov, Evgeny. 2019. “The Challenges of Using Neural Machine Translation for Literature.”
In Proceedings of the Qualities of Literary Machine Translation, edited by James Hadley,
Maja Popović, Haithem Afli, and Andy Way, 10–19. Dublin: European Association for
Machine Translation.
Palmer, Frank Robert. 1990. Modality and the English Modals. London: Longman Group
Limited.
Toral, Antonio, and Andy Way. 2015. “Machine-assisted Translation of Literary Text: A Case
Study .” Translation Spaces 4 (2): 241–268.
Vanmassenhove, Eva, Dimitar Shterionov, and Andy Way. 2019. “Lost in Translation: Loss and
Decay of Linguistic Richness in Machine Translation.” In Proceedings of Machine
Translation Summit XVII Volume 1: Research Track, edited by Mikel Forcada, Andy Way,
Barry Haddow, and Rico Sennrich, 222–232. Dublin: European Association for Machine
Translation.
Van Brussel, Laura, Arda Tezcan, and Lieve Macken. 2018. “A Fine-grained Error Analysis of
NMT, PBMT and RBMT Output for English-to Dutch.” In Proceedings of the Eleventh
International Conference on Language Resources and Evaluation, edited by
Nicoletta Calzolari et al., 3799–3844. Miyazaki: European Language Resources
Association.
Voigt, Rob, and Dan Jurafsky. 2012. “Towards a Literary Machine Translation: The Role of
Referential Cohesion.” In The 2012 Conference of the North American Chapter of the
Association for Computational Linguistics: Human Language Technologies (NAACL-HLT
2012), edited by Eric Fosler-Lussier, Ellen Riloff, and Srinivas Bangalore, 18–25. Montréal:
Association for Computational Linguistics.
Webster, Rebecca, Margot Fonteyne, Arda Tezcan, Lieve Macken, and Joke Daems. 2020.
“Gutenberg Goes Neural: Comparing Features of Dutch Human Translations with Raw
Neural Machine Translation Outputs in a Corpus of English Literary Classics.”
Informatics 7 (3): 32.
Yang, Liu, and Min Zhang. 2015. “Statistical Machine Translation.” In The Routledge
Encyclopedia of Translation Technology, edited by Chan Sin-wai, 201–213. New York:
Routledge
Résumé
Afin d’explorer les capacités et les limites d’une traduction anglais-chinois basée sur l’IA, cette
étude examine les performances du système de traduction automatique neuronal DeepL dans
la traduction des pièces de Shakespeare : Coriolanus et Le Marchand de Venise. Adoptant une
approche basée sur un corpus, cet article étudie le taux de précision, le taux de fluidité, les
caractéristiques linguistiques et l’utilisation de diverses méthodes de traduction dans les traduc-
tions chinoises des pièces de Shakespeare effectuées via DeepL, et les compare aux traductions
de Liang Shiqiu, un traducteur chinois bien connu. L’étude montre que DeepL donne de bons
résultats dans la traduction de ces œuvres, avec un taux de précision et de fluidité supérieur à
80% dans les textes échantillonnés, ce qui suggère le potentiel de l’utilisation de la traduction
automatique neuronale dans la traduction de textes littéraires dans des langues éloignées. Nos
recherches révèlent en outre que les traductions de DeepL présentent un certain degré de créa-
tivité dans l’utilisation de méthodes de traduction telles que l’ajout, l’explicitation, la conversion
et le changement de perspective, ainsi que dans l’utilisation des particules modales finales des
phrases chinoises, de même que des verbes modaux chinois. D’autre part, le système semble
être limité dans la mesure où il présente des erreurs de traduction et des traductions littérales.
Mots clés : traduction automatique neuronale, textes littéraires, créativité, limites
Address for correspondence
Hu Kaibao
Institute of Corpus Studies and Applications
1550 Wenxiang Road
Shanghai
P. R. China
[email protected]
Co-author information
Li Xiaoqian
[email protected]
Publication history
Date received: 30 June 2022

Date accepted: 12 February 2023
Published online: 24 July 2023

The Creativity and Limitations of AI Neural Translation

Uploaded by

Copyright:

Available Formats

The Creativity and Limitations of AI Neural Translation

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Creativity and Limitations of AI Neural Translation

Uploaded by

Copyright:

Available Formats

The creativity and limitations of AI neural

Chinese translation of Shakespeare’s plays

Hu Kaibao and Li Xiaoqian

This study examines the performance of the neural machine translation

Keywords: AI, translation, corpus, limitations, Chinese, Shakespeare,

Machine translation (hereafter MT) refers to using computer software to auto-

https://doi.org/10.1075/babel.00331.hu | Published online: 24 July 2023

In this article, we will compare the Chinese translations of Shakespeare’s plays

to the translation of literary texts? (2) Do the translations by an AI literary trans-

2. MT in the translation of literary texts

English, and whether an AI literary translation or literary translation by neural

3.1 The corpus for the research

As mentioned, this article aims to investigate the performance of DeepL in the

3.2 Research procedures

To explore the possibility of using MT in literary translation and examine possible

3.3 Theoretical framework

3.3.1 Sentence-final modal particles

3.3.2 Modal verbs

3.3.3 Translation methods

beyond its literal meaning. It involves such translation methods as addition,

4. Results and discussion

4.1 Errors in the Chinese translations of Shakespeare’s Plays by DeepL

Errors in translation may be categorized as being either accuracy errors or fluency

Table 2. Errors in the Chinese translations of Shakespeare’s Plays by DeepL

Number of sentences 346 465 349 644

As indicated in Table 2, 47 out of 346 sentences and 87 out of 465 sentences of

4.2 Linguistic features of the Chinese translations by DeepL

4.2.1 The use of Chinese sentence-final modal particles

Table 3. Chinese sentence-final modal particles identified in the corpus

As shown in Table 4, Chinese sentence-final modal particles occur with a

We analyzed the correspondence between the sentences containing the above

from English affirmative sentences in order to emphasize a fact or a point of view.

In Example (1) above, 呢 is featured in a Chinese affirmative sentence in the

4.2.2 The use of Chinese modal verbs

Table 4. Chinese modal verbs identified in the corpus

of the modal verbs 应 (should) and 能够 (can). The frequencies of 应 (should)

4.3 The use of translation methods in the translations by DeepL

In this section, we investigate the use of translation methods by DeepL as

speech may produce an ungrammatical or unidiomatic target text due to the

4.3.3 Shift of perspective

translation of Shakespeare’s plays]. In Shashibiya danchen sibai zhounian jinian ji 莎士比

Mots clés : traduction automatique neuronale, textes littéraires, créativité, limites

Address for correspondence

Date received: 30 June 2022

You might also like