When reading English papers, the most common approach is to copy a passage, paste it into a chat box, and get some Chinese back. That works for abstracts, but with long texts it starts to lose things: paragraphs go missing, terminology is translated inconsistently, and formulas and symbols get misaligned. MuchSci's paper translation feature breaks this into two steps — "preserve paragraph structure" and "side-by-side comparison." The goal is not prettier translation, but letting you check it.
This article unpacks the paper translation module in the app collection: after uploading a PDF, why doesn't it just output one large block of translated text, but instead places the original and translation side by side?
01 Start with the interface: what side-by-side comparison solves
▲ Figure 1 | The comparison interface presents three layers of information at once: the original text, the translation, and terminology constraints
When you open the translation interface, the original is on the left and the translation is on the right, aligned paragraph by paragraph. MuchSci calls this design "preserve the original paragraph structure." This does not simply mean adding line breaks; it means the Nth paragraph of the translation always corresponds to the Nth paragraph of the original. The value of this alignment is locating: when a translated sentence seems off, you don't need to move your eyes away from the current line to scan the corresponding original.
There are two other easily overlooked points in the interface. One is the status notice under the "Translation Preview" tab; the other is the "Stop Translation / Delete" button. The former indicates that the current translation is still an editable intermediate product; the latter indicates that users can actively interrupt a translation run they are not satisfied with. Translation here is treated as a progressive task, not a one-off black-box output.
▲ Figure 2 | The real interface: source language English, target language Chinese, a PDF being translated, with the original and translation displayed in sync side by side
02 The difference between feeding in the whole document and comparing paragraph by paragraph
▲ Figure 3 | The essential difference between the two translation approaches is not fluency, but traceability
If you copy all the text from a PDF to a model at once, it can usually produce a passage of Chinese that reads smoothly. The problem is that you get the result but lose the frame of reference. Paper translation differs from everyday translation in that readers often need to go back to the original to confirm: does this "it" refer to the experimental subject or the model? Have I misread the unit for this value? Is this term translated differently elsewhere?
Paragraph-by-paragraph side-by-side comparison preserves this ability to trace back. Next to each translated paragraph is its original, paragraph numbering stays the same, and the structure remains intact. Terminology consistency also becomes achievable under this structure: if "embedding" is translated as both “embedding” and “vectorization” in the same passage, the side-by-side view immediately exposes it; with whole-document output, this problem gets buried in two passages far apart from each other.
03 Translation workflow: from PDF to comparison file
▲ Figure 4 | After a PDF is uploaded, the system executes in the order of parsing, identifying terminology, locking translations, translating paragraph by paragraph, side-by-side comparison, and export
The first step in the workflow is not translation, but parsing the PDF. The system needs to separate titles, body paragraphs, formulas, and lists in academic papers to ensure there is no misalignment in the subsequent comparison. The second and third steps are terminology work: first scan the full text to list high-frequency technical term candidates, then let the user or rules determine a unified translation. Only after these two steps are done does paragraph-by-paragraph translation begin.
Why lock terminology before translating? Because large language models are not stable in how they translate the same technical term, especially in a long document. Providing a glossary in advance is equivalent to imposing constraints on every subsequent paragraph translation. At export, the translation that preserves paragraph structure can be used directly for close reading, annotation, or further rewriting.
04 How terminology consistency is ensured
▲ Figure 5 | The glossary is not decoration; it is a constraint on consistency across the full text
Terminology consistency is the most hidden problem in paper translation. When the same word is translated differently in different places, readers will mistake it for two concepts. MuchSci's approach is to use the glossary as a preceding constraint: determine the translations of high-frequency terms before translation begins, and during translation the model is constrained to use this fixed set of terms.
This mechanism is equally effective for references and symbols. English papers contain many references such as it, the above, and this approach. If you translate paragraph by paragraph out of context, it is easy to get the reference wrong. Side-by-side comparison lets you immediately return to the original paragraph to check when the translation seems off. Formulas, superscripts and subscripts, Greek letters, and other symbols also depend on paragraph structure to keep their positions; otherwise, symbols may be dropped or misaligned.
05 Boundaries
① This article discusses the feature's placement and mechanism, and does not constitute a promise of results. Translation quality depends on the complexity of the original text, the completeness of the glossary, and the academic field. Side-by-side comparison only reduces the cost of checking; it does not guarantee zero errors in the translation.
② Output results require human review; professional judgment cannot be outsourced. The accuracy of key terms, data, formulas, and conclusions still needs to be confirmed by readers against the original text. Machine translation cannot be treated directly as the final version.
③ Translation accuracy and academic integrity are subject to the current regulations of the relevant institution. When it comes to paper writing, submission, and citation, please comply with the specific requirements of your school or journal regarding AI-assisted translation. For professional texts involving medicine, law, and other fields, review by qualified personnel is recommended.