Where Have All the Paragraphs Gone?
Detecting and Exposing Censorship in Chinese Translation

Abstract

Translated literature often gets cut or altered before being published in China. Worse yet, both readers and the original literature authors are not informed of such removal or changes most of the time. Such erasures can change the fundamental meaning of texts, and readers are left unaware of vital context.

In this work, we propose an NLP-based approach to detecting translation censorship automatically and effectively. We further conduct a case study to analyze censorship in translated literature. We build a demonstration website and argue that an effective way to combat this form of censorship is to actively trigger the Streisand effect: by highlighting what text has been censored and making it easy to access, we hope the censored text reaches a broader audience, including those who would not have read the censored version in the first place.

1. Introduction

“The past was erased, the erasure was forgotten, the lie became the truth.”

Like other media, China censors translated books and literature before they are published in China. This censorship is often invisible to readers as a deliberate choice by the censor to avoid triggering Streisand effect [18, 4]. That is, telling readers that a book is (partially) censored provokes their curiosity and interest, and the censored content is thus more likely to be sought out by more people [7, 8, 9]. While this type of quiet censorship could be caught by painstakingly translating the original text for comparisons, it is often too expensive to do so in practice [12, § “No One Is Checking”].

Surprisingly, it is often the publisher (and not the government) who self-censors translated books, as failure to do so could result in various forms of punishment, including harsh scrutiny of future publications [12, § “Censorship on the Mainland”], confiscating the published books [20, § “Books published and then banned”], and immediate suspension of the publishing rights [21, 5]. This type of self-censorship based on fear is effective, because without clear and exact rules, a publisher errs on the side of conservatively censoring anything that might be interpreted as sensitive, resulting in more restrictive censorship than if it were done by the censoring government themselves [4, 18]. In many cases, authors were not aware that the translation of their work had been cut until well after its publication [12, § “In the Dark”]. In other cases, authors were aware of the censorship, but had to make a choice between publishing a censored version [2, 11], or not publishing at all [13].

In this work, we propose a low-cost and effective way to automatically detect when a translated book or text has omitted or substantially changed from the original work. We then present the original text and its translation side-by-side on a dedicated website, with censored contents highlighted. By highlighting these erasures and making them easy to access for the general public, we frustrate censors' attempts to suppress information and make its censorship backfire.

Our tool operates by performing an automatic translation of the translated text (e.g. in Chinese) back to the original language (e.g. English), and comparing the paragraphs using state-of-the-art natural language processing (NLP) tools. If a paragraph is either missing or is substantially different in meaning, our tool can alert the reader to the deletion or change. By using NLP, our tool can gracefully handle the subtle differences between automatic and manual translation that do not change the semantic meaning of paragraphs. Our hope is to ultimately apply this tool to many translated works; we report on our prototype tool and preliminary results here.

2. Methodology

Program to detect censorship in translation

The way we detect censorship in translation can be framed as a typical bitext word alignment task in the NLP domain, where we align two texts and look for discrepancies. In particular, our program first uses the Google translation API [6] to machine-translate the Chinese book back into English. We then break the bitext (the machine translation and the English version) into paragraphs to form a bipartite graph, where each paragraph is a vertex, and the weight of each edge is the similarity between the two paragraphs. To calculate the weight of each edge, we first embed each paragraph using the NLP model distiluse-base-multilingual-cased-v1 [15] and assign the negation of the cosine similarity of two paragraphs as the weight [16]. We set a similarity threshold, and when two paragraphs' similarity is below the threshold, we do not consider them as similar at all by assigning a positive number to the weight. Finally, we calculate the linear sum assignment of the bipartite [19]. The result tells which machine-translated paragraph matches which original paragraph.

Once the program computes which paragraph in the machine translation matches which paragraph in the original version, it applies the same approach recursively to determine, for each pair of matched paragraphs, which translated sentence matches which original sentence. Finally, the program marks unmatched sentences as possibly being cut, and sentences with a low similarity as possibly being altered.

Human analysis

As the first step, the program analysis provides a similarity score for each sentence. It allows us to quickly and effectively detect and identify potential censorship in translation. When the similarity is lower than a certain threshold, we manually confirm the missing text and identify potential meaning changes in context.

Limitations

The current approach does not consider the cases where two machine-translated sentences correspond to one original English sentence, or vice versa. In Appendix B, we introduce this issue and a solution to it in detail. To reduce the potential bias introduced in human analysis, we will minimize the involvement of humans and measure the inter-rater reliability when human analysis is necessary.

3. Case Study

In this section, we conduct a case study to help readers better understand this form of censorship. In particular, we use Chapter 5 of Chinese Literature: A Very Short Introduction by Sabina Knight as a demonstration [10], comparing its original English version with its Chinese version translated by Li Yongyi side-by-side. We strongly encourage readers to visit the demonstration website for an interactive experience.

In Table 1, we categorize the censored topics and count the censorship actions in the Chinese translation of Chapter 5 of Chinese Literature: A Very Short Introduction. In total, we identified seven censored topics, including 2000 Nobel Prize in Literature laureate Gao Xingjian and his work, the plasma economy in Henan province, China, the Tiananmen Square Massacre in 1989, and Mao Zedong and the political movements he led: the Great Leap Forward and the Cultural Revolution. Interestingly, as discussed previously, a discussion of the book censorship itself is cut, showing another effort by the censor to avoid triggering the Streisand effect.

4. Discussion

Alternative approaches

While our approach to detecting censorship in translation is based on state-of-art NLP models and is proved to be efficient and effective, it is by no mean the only approach. For example, an alternative approach we tried is bitext mining. However, possibly due to our limited understanding and skills, we have not been able to use bitext mining to yield a detection result that is as good as the current approach [17, 3]. Appendix C logs that ChatGPT identified a missing sentence in translation correctly and reasoned on it [1]. We note that having multiple independent approaches to detecting censorship in translation will help to improve the overall robustness of detection, as one can compare and evaluate each method's output against the others'.

5. Conclusion

In this work, we explore censorship in translation. In particular, we propose an NLP-based approach to automatically and effectively detecting translation censorship. We further conduct a case study to analyze censorship in translated literature. In addition, we build a demonstration website and argue that an effective way to combat this form of censorship is to actively trigger the Streisand effect. That is, by highlighting the censored text and making it easy to access, we hope the censored text reach a broader audience, even including the curious who would not have read the censored version of the books in the first place.

Acknowledgments

The authors would like to express their sincere gratitude to Jeffrey Knockel and the anonymous reviewers for their insightful feedback and constructive comments. We also appreciate Xiao Qiang and Michael Carl Tschantz for their initial discussions on this work. We thank David Fifield for his guidance, comments, feedback, and support throughout the work.

This study was funded by the NSF grant 1953786, and by the Young Faculty Award program of the Defense Advanced Research Projects Agency (DARPA) under the grant DARPA-RA-21-03-09-YFA9-FP-003. The views, opinions, and/or findings expressed are those of the authors and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government.

Availability

To maintain reproducibility and stimulate future work, we make our code and data public to the maximum extent that does not harm our anonymity. The code, data, and the project webpage are available at: https://gfw.report/projects/translation-censorship/.

References

1. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, 1877–1901.
2. Liz Bury. 2013. Author bows to Chinese censorship of his Deng Xiaoping biography. The Guardian, October 2013.
3. Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, and Wei Wang. 2020. Language-agnostic BERT sentence embedding.
4. David Fifield. 2022. Reading group: Censored: Distraction and Diversion Inside China's Great Firewall. August 2022.
5. Nam fung Chang. 2008. Censorship in translation and translation studies in present-day China. In Translation and censorship in different times and landscapes, 229–240. Cambridge Scholars.
6. Suhun Han. Googletrans.
7. William R. Hobbs and Margaret E. Roberts. 2018. How sudden censorship can increase access to information. American Political Science Review 112(3):621–636.
8. Sue Curry Jansen and Brian Martin. 2003. Making censorship backfire.
9. Sue Curry Jansen and Brian Martin. 2015. The Streisand effect and censorship backfire.
10. Sabina Knight. 2012. Chinese Literature: A Very Short Introduction. Oxford University Press.
11. Lily Kuo. 2019. Edward Snowden says autobiography has been censored in China. The Guardian, November 2019.
12. Alexa Olesen, Katy Glenn Bass, Suzanne Pettypiece, et al. 2015. Censorship and conscience: Foreign authors and the challenge of Chinese censorship. May 2015.
13. Natalie Ornell. 2014. Evan Osnos on not publishing his book in China. China Digital Times, May 2014.
14. George Orwell. Nineteen Eighty-Four. Project Gutenberg Australia.
15. Nils Reimers. Pretrained models.
16. Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics.
17. Nils Reimers and Iryna Gurevych. 2020. Making monolingual sentence embeddings multilingual using knowledge distillation. arXiv preprint arXiv:2004.09813.
18. Margaret E. Roberts. 2018. Censored. Princeton University Press.
19. SciPy. SciPy: linear sum assignment function.
20. Yuwu Song. 2022. Book censorship in Post-Tiananmen China (1989–2019). Journal of East Asian Libraries 2022(175):4.
21. Zaixi Tan. 2015. Censorship in translation: The case of the People's Republic of China. Neohelicon 42(1):313–339.

A. The demonstration website

Figure 1 is a screenshot of the demonstration website. We encourage readers to visit the website for an interactive experience: https://gfw.report/projects/translation-censorship.

B. An example of the one-on-one mapping issue

The current detection program does not consider cases where two machine-translated sentences correspond to one original English sentence, or vice versa. For example, when translating the sentence “[w]ear whatever you want, don't wait until you are my age”, the translator may break it into two sentences “[w]ear whatever you want” and “[d]on't wait until you are my age”. While no sentence is cut or altered, the current program will only match “[w]ear whatever you want” with the original text, and incorrectly mark “[d]on't wait until you are my age” as missing.

To address this problem, one can improve the program by heuristically merging a missing sentence with its context and check if the similarity score goes up. If so, it may consider the sentence and its context together correspond to one sentence. In this example, the program would have first merged “[d]on't wait until you are my age” with “[w]ear whatever you want”, and then check if the similarity score goes up.

C. A conversation with ChatGPT

In this section, we log a conversation with ChatGPT [1], which suggests a promising alternative approach to detecting censorship in translation in future work.

As shown in Table 2, in the first session, we provided the chatbot with an original text and its partially censored translation. The chatbot identified the missing sentence correctly and reasoned on it.

Table 3 shows the second conversation session with ChatGPT. As a control group, we provided the chatbot with the same original text and completed the translation ourselves. This time, the chatbot correctly stated that the translation was complete.

D. Ethics

This work does not involve human subjects.