Integration of machine translation and human translation

Machine translation quality has reached a higher plateau in recent years due to the emergence of neural machine translation (NMT) since it entered the public domain in 2016. DePalma and Lommel (2018, p. 7) summarize the development of NMT as follows. (For details regarding traditional MT models and MT in general, see Arnold, Balkan, Lee Humphreys, Meijer, & Sadler, 1994; Hutchins & Somers, 1992; Koehn, 2010; Nirenburg, Somers, & Wilks, 2003; Sin-wai, 2015.)

Machine translation based on neural networks is the newest entrant to the field.

These systems use machine-learning techniques that emulate the behavior and function of biological neurons. The theory is that this approach more closely reflects how human translators perform their tasks. These techniques advanced

11

rapidly in the last two years, and now dominate MT development and commercial deployments. (DePalma & Lommel, 2018, p. 7)

From a linguistic point of view, what makes NMT most promising is that it can produce more natural-sounding translations. Compared to Statistical MT (SMT), NMT rates much higher in terms of fluency, which increases user satisfaction with its output (DePalma & Lommel, 2018, p. 7). The shortcoming of NMT, on the other hand, is that debugging and troubleshooting is almost impossible, and NMT technology remains unproven, thus still unreliable.

Nevertheless, the integration of NMT and HT is a growing trend in the translation industry. There are two main scenarios: (1) an MT system translates the source content and humans post-edit (PE) the MT output (MT works as translators, humans as the revisers, a combination known as MTPE); and (2) humans translate source content referring to MT suggestions (humans work both as translators and revisers). But whichever the scenario, what matters the most to human translators is the unit price per source word or character.

In the first scenario, there are mainly two ways for LSPs to decide the translation rate: (1) some LSPs might request translators to perform a sample check or test on MT output and establish if translators are willing to accept the discount rate proposed by the LSP. If not, either a revised discount rate can be agreed through discussion or the translators will not work for a particular project/account; and (2) some LSPs simply present the discount rate to the translators, and no negotiation is possible. Translators will have to decide whether to sell their services at a lower rate irrespective of whether MT can improve their productivity to a reasonable level. As for how a discount rate is decided, LSPs often need to evaluate the content to be translated and the language pairs first to decide if MT is suitable for a particular job. If the evaluation result turns out

12

positive, LSPs will further perform productivity tests with their language teams in order to set a discount rate by language pair.

It is widely believed in the translation industry that very few NMT systems support JA to ZH-CN translation to a PE threshold level at the moment. Based on the present author’s experience, when JA to ZH-CN is the language pair and the quality expectation from the buyer is high, the MTPE process will not be recommended by LSPs as it is unlikely to be able to improve productivity, nor are LSPs willing to promise

“perfect” quality at a discounted price. (See Koponen (2016) for the general trend of MTPE research, and Yamada (2019) for productivity gains and cognitive effort associated with PE of NMT.)

As for the second scenario where humans work both as translators and revisers with MT output as a reference, when JA is the source and ZH-CN the target and quality translation is expected for content such as CSR reports, the MT result is, again, often of no use, especially when the source JA sentences are long or ill-structured. Table 1.1 shows the EN and ZH-CN Google Translate results of a sample sentence from a JA CSR report.

Table 1.1 Example of Google Translate from JA into EN and ZH-CN a. JA sentence from a CSR report

XXX は、写真分野で培ったコラーゲン技術を基盤に開発した、細胞培養に必要な細 胞外マトリックス（リコンビナントペプチド：RCP）「cellnest」を研究用試薬として 発売しているほか、2014年当時、国内で唯一、自家培養表皮や培養軟骨等の再生医療 製品を上市していたYYYを連結子会社化。

b. Google translation of a to EN (as of March 21, 2020)

XXX has released cellnest, an extracellular matrix (recombinant peptide: RCP) required for cell culture, developed as a research reagent, based on collagen technology

13

cultivated in the field of photography. Only YYY, which marketed regenerative medicine products such as self-cultured epidermis and cultured cartilage, became a consolidated subsidiary.

c. Google translation of a to ZH-CN (as of March 21, 2020)

XXX 发布了 cellnest，这是细胞培养所需的细胞外基质（重组肽：RCP），它是根据摄 影领域培养的胶原技术开发的研究试剂。只有销售自养表皮和软骨的再生医学产品的 YYY 成为合并子公司。

d. Google translation of c to EN for comparison with b (as of March 21, 2020) XXX released cellnest, an extracellular matrix (recombinant peptide: RCP) required for cell culture, which is a research reagent developed based on collagen technology cultivated in the field of photography. Only YYY, which sells regenerative medicine products of autotrophic epidermis and cartilage, becomes a consolidated subsidiary.

The Google NMT output of a JA sentence from a CSR report shown in Table 1.1 demonstrates that: (1) it would be less confusing and straightforward for translators to translate from scratch without struggling to work upon a bad ZH-CN NMT result; and (2) the ill-structured JA has a bad impact on EN translation as well. In any case, it will take translators considerable time to read such a long source sentence and make sense of the complicated structure. This shows the utility of the suggestions (already summarized in Section 1.1.2) that Lommel (2016) makes to buyers on source content optimization to achieve better outcome, even if they are general suggestions that are not specific to either HT or MT. By definition, post-hoc solutions are contingency measures which cannot improve the quality of translation from JA systematically. JA source optimization through a joint effort of buyers and LSPs (possibly, academics as well) is the solution calculated to achieve long-term success. This is echoed by Kelly (2010, p. 1) who observes that “translation and localization professionals often point out that quality starts at the source.”

14

Back to the integration of MT and HT which is predominating the workflow in the translation industry day by day, Yamada (2018, para. 1) has pointed out that the assumption behind the MT+PE model is that the competences of the machine and those of the human are complementary, but this assumption is no longer valid. The reason is that technological advances have created NMT systems that more closely mimic HT, including human-like errors, “thus increasing the effort required for post-editors to detect and correct errors” (para. 1). Yamada (2019, p. 87) also concludes that NMT+PE does not contribute to reducing the post-editors’ cognitive effort or error correction rate and that NMT+PE requires almost the same competence as ordinary HT. Although Yamada’s (2019) experiment involved college students who might arguably be less familiar with the task, the present author observes that the NMT+PE model poses the same challenges in a professional setting.