Inside NoveLM: 16 genre styles, 7-category NER with gender inference, 5-dimension quality scoring. How TeaNovel's AI translation engine works.
Every AI translation tool runs on large language models. ChatGPT, DeepL, Google Translate — they all use neural networks trained on massive datasets to convert text from one language to another. The models are impressive. The results, for general text, are often excellent.
But novel translation is not general text. A 500-chapter xianxia novel contains thousands of recurring proper nouns that must be translated identically every time they appear. It uses genre-specific registers that shift between formal cultivation prose, casual modern dialogue, and classical wuxia cadence depending on the scene. And it demands quality transparency — a reader who does not speak Chinese needs to know whether a translation is trustworthy before investing hours reading it.
General-purpose translation handles none of this. NoveLM was built to handle all of it.
This article explains how NoveLM works — not at the code level, but at the level that matters to readers: what it does differently, why those differences exist, and how they affect the English text you read.
Take a passage from a typical xianxia novel:
萧炎的斗气如同烈焰般在经脉中奔涌，九转玄功的第三层终于突破。他猛然睁开双眼，周身的气势骤然攀升，周围的空气都仿佛凝固了一般。
A general-purpose translator processes this as a standalone text block. It has no information about what genre this passage belongs to, no memory of how "斗气" was translated in the previous 200 chapters, no knowledge of whether "萧炎" is male or female, and no way to evaluate whether its own output is any good.
NoveLM processes the same passage through four layers:
Each layer solves a problem that general translation ignores. Together, they produce output that reads like genre fiction rather than machine translation.
When a novel enters NoveLM's pipeline, the first decision is genre classification. NoveLM supports 16 genre-specific translation styles:
| Genre | Register | Example Adaptation |
|---|---|---|
| Xianxia | Formal, classical-inflected | "Dou Qi," "tribulation," "meridians" |
| Xuanhuan | Epic fantasy, dramatic | "Essence power," "divine realm" |
| Wuxia | Classical martial, elegant | "Sword qi," "internal force," "jianghu" |
| Romance | Conversational, emotionally textured | "Heart raced," natural dialogue rhythm |
| Historical | Formal period prose | Court titles, period-appropriate phrasing |
| Ancient | Archaic, ceremonial | Classical sentence structures |
| Urban | Modern casual, grounded | Contemporary references, slang-adjacent |
| Modern | Clean contemporary | Neutral modern English |
| Fantasy | Western fantasy-inflected | "Mana," "enchantment," familiar fantasy terms |
| Sci-fi | Technical, precise | Scientific terminology, measured cadence |
| Horror | Tense, atmospheric | Short sentences, sensory language |
| Comedy | Light, rhythmic | Timing-aware phrasing, punchline preservation |
| Mystery | Controlled, deliberate | Information pacing, clue-conscious word choice |
| Slice of Life | Warm, observational | Everyday language, gentle pacing |
| Thriller | Clipped, urgent | Short paragraphs, tension-building |
| Military | Functional, authoritative | Rank-specific language, tactical precision |
The system also recognizes 10 genre aliases to handle common variations in how novels are tagged:
Genre routing is not cosmetic. It changes the fundamental character of the translation output. The same Chinese sentence translated in xianxia style and modern style will produce noticeably different English — different vocabulary, different sentence structure, different rhythm. Here is a concrete example:
Original: 他凝聚全身真气，一掌击出。
Xianxia style: He gathered his True Qi throughout his body and struck with a single palm.
Modern style: He concentrated all his energy and threw a palm strike.
Wuxia style: He condensed his inner force to its peak and unleashed a palm strike.
Same meaning. Different register, different terminology, different reading experience. The xianxia reader expects "True Qi." The modern fiction reader expects "energy." The wuxia reader expects "inner force." NoveLM delivers the right term for the right genre.
The second layer runs before translation begins. NoveLM's NER system scans the source text and identifies entities across 7 categories:
| Category | Examples | Why It Matters |
|---|---|---|
| Character | 萧炎 → Xiao Yan | Names must be consistent across 500+ chapters |
| Location | 天云山 → Heavenly Cloud Mountain | Readers need to recognize recurring locations |
| Organization | 天云宗 → Heavenly Cloud Sect | Sects, guilds, and factions appear in every chapter |
| Skill | 九转玄功 → Nine Revolutions Mysterious Art | Technique names are genre-critical terminology |
| Item | 火莲瓶 → Fire Lotus Bottle | Artifacts and items drive plot points |
| Title | 大长老 → Grand Elder | Rank and title consistency affects power hierarchy clarity |
| Race | 妖族 → Demon Race | Species and race names must be stable |
Chinese third-person pronouns present a unique challenge for translation. In spoken Chinese, 他 (he), 她 (she), and 它 (it) are pronounced identically (tā). In older or informal writing, 他 is often used generically regardless of gender. This means a translator — human or AI — must infer gender from context.
NoveLM performs gender inference using a weighted voting system across 6 basis types:
Each basis type casts a weighted vote. When multiple signals agree, the system assigns gender with high confidence. When signals conflict, the system marks the character as ambiguous and uses neutral constructions until more evidence accumulates.
The result: pronouns are correct in the vast majority of cases, including for characters whose gender is not immediately obvious. This is a problem that general translation tools consistently fail at because they lack any mechanism for cross-chapter entity tracking.
Once an entity mapping is established — "萧炎 = Xiao Yan (male, character)" — it persists through every subsequent chapter. Chapter 1 and chapter 500 use the same translation for the same entity. There is no drift, no variation, and no need for the reader to figure out whether "Xiao Yan" and "Shaw Yan" are the same person.
This is the single most impactful difference between NoveLM and general translation tools. A long web novel may introduce 200+ named entities. Without persistent tracking, every entity is an opportunity for inconsistency. With tracking, the reader experiences one coherent story instead of a patchwork of disconnected chapters.
With genre routing and NER data in place, the actual translation step begins. NoveLM generates English text using genre-specific prompting that incorporates:
The cultural reference decision is worth examining. In xianxia fiction, "斗气" is better kept as "Dou Qi" because the xianxia reading community uses that term universally. Translating it to "fighting spirit" is technically accurate but strips away genre identity. In a modern urban novel, the same concept might be better rendered as "fighting energy" because the modern fiction reader does not have the same genre-specific vocabulary expectations.
NoveLM makes these decisions systematically based on genre. A xianxia translation preserves genre-standard transliterations. A modern translation favors natural English equivalents. A wuxia translation uses classical martial arts terminology. These are not random choices — they are the accumulated conventions of decades of genre translation, encoded into the translation pipeline.
Here is the same passage rendered in three different genre styles:
Original: 她轻轻握住他的手，心中涌起一股暖意。这个世界再冷，有他在身边就足够了。
Romance style:
She gently took his hand, warmth flooding through her chest. The world could be as cold as it wanted — having him beside her was enough.
Xianxia style:
She lightly clasped his hand, a surge of warmth rising in her heart. No matter how cold this world grew, his presence at her side was sufficient.
Modern style:
She held his hand softly, a warm feeling welling up inside her. Even if the whole world turned cold, having him here was all she needed.
The romance style prioritizes emotional immediacy — "warmth flooding through her chest" is visceral and intimate. The xianxia style uses slightly formal phrasing — "his presence at her side was sufficient" carries the dignity of cultivation fiction. The modern style lands in between — natural, warm, but not heightened. All three are accurate. Each reads like it belongs in its genre.
After translation, NoveLM evaluates its own output across five dimensions with weighted scoring:
| Dimension | Weight | What It Measures |
|---|---|---|
| Accuracy | 30% | Does the translation faithfully convey the original meaning? |
| Fluency | 25% | Does the English read naturally and grammatically? |
| Style | 20% | Does the translation match the genre's register and tone? |
| Terminology | 15% | Are proper nouns and genre terms translated consistently? |
| Format | 10% | Are paragraphs, dialogue, and structural elements preserved? |
Scores follow a 100-point rubric:
| Score Range | Rating | Interpretation |
|---|---|---|
| 90-100 | Exceptional | Publication-quality translation |
| 75-89 | Good | Minor issues, fully readable |
| 60-74 | Acceptable | Noticeable but non-critical issues |
| 40-59 | Poor | Significant problems, may require review |
| 20-39 | Very Poor | Major inaccuracies or readability issues |
| 0-19 | Untranslated | Translation failed or was not attempted |
Quality scoring serves three purposes:
Reader confidence. A chapter that scores 94 can be read without concern. A chapter that scores 68 on Terminology signals that some proper nouns may have been handled incorrectly — and the reader can review those terms alongside the chapter in the wider TeaNovel workflow.
Translation transparency. No AI translation is perfect. The question is whether the tool acknowledges its limitations or hides them. NoveLM scores every chapter visibly. If a chapter struggled — perhaps because the source text contained unusual dialect, code-switching between literary and vernacular Chinese, or ambiguous pronouns — the score reflects that. The reader knows before investing time.
Continuous improvement signal. Scoring patterns reveal which types of content translate well and which do not. Chapters heavy in classical Chinese poetry consistently score lower on Fluency because poetic forms resist natural English prose. Chapters with extensive dialogue consistently score higher because conversational Chinese maps well to conversational English. These patterns inform pipeline improvements over time.
To make the four layers concrete, here is what happens when the same passage goes through NoveLM versus a general translation pipeline:
Original (xianxia scene):
萧炎的斗气如同烈焰般在经脉中奔涌，九转玄功的第三层终于突破。他猛然睁开双眼，周身的气势骤然攀升，周围的空气都仿佛凝固了一般。
General translation (no genre routing, no NER, no scoring):
Xiao Yan's fighting spirit surged through his meridians like a blazing fire. The third level of the Nine Turns Mysterious Technique finally achieved a breakthrough. He suddenly opened his eyes wide, and the momentum around his body sharply increased. The air around him seemed to solidify.
NoveLM (xianxia style, NER active, scored 93):
Xiao Yan's Dou Qi surged through his meridians like raging flames as the third layer of the Nine Revolutions Mysterious Art finally broke through. His eyes snapped open, his aura climbing sharply, and the very air around him seemed to freeze solid.
What changed:
| Dimension | General Translation | NoveLM |
|---|---|---|
| Genre terminology | "fighting spirit" | "Dou Qi" (genre standard) |
| Skill name | "Nine Turns Mysterious Technique" (varies) | "Nine Revolutions Mysterious Art" (locked by NER) |
| Prose structure | 4 short sentences | 2 flowing compound sentences |
| Register | Neutral informational | Dramatic xianxia cadence |
| Quality score | None (unknown confidence) | 93 (Exceptional) |
The general translation is not wrong. Every word is defensible. But it reads like a translation. The NoveLM output reads like xianxia fiction — because the genre routing selected xianxia register, the NER locked skill names, and the style profile shaped sentence rhythm to match genre expectations.
Every claim in this article is derived from NoveLM's actual codebase:
These are not marketing numbers. They are the actual parameters of the system that translates your chapters. Every genre style exists as a distinct configuration. Every NER category has dedicated extraction logic. Every quality dimension has a defined weight and rubric.
NoveLM's four layers work together to solve a problem that general translation tools do not even attempt: making a 500-chapter Chinese web novel read like a coherent, genre-appropriate English novel from start to finish.
Genre routing means your xianxia novel reads like xianxia — not like a business email or a Wikipedia article. NER means "Xiao Yan" is always "Xiao Yan" and never "Shaw Yan" or "Little Flame." Quality scoring means you know which chapters to trust and which to double-check. And the immersive reader means you consume all of this in an interface designed for fiction, not for reviewing translation output.
No translation tool is perfect. NoveLM will not produce output identical to what a professional human translator would write. But it produces output that respects the genre, maintains consistency, scores its own confidence, and improves chapter by chapter as the NER database grows.
Try it at read.teanovel.com — 1,000 free credits per month, enough to see the difference between general translation and genre-aware translation for yourself.
No. NoveLM is TeaNovel's translation pipeline, not a single language model. It includes genre routing, named entity recognition, genre-specific prompting, and quality scoring layers built on top of AI models accessed through the Vercel AI Gateway. The layers that make NoveLM different — 16 genre styles, 7-category NER, 5-dimension scoring — do not exist in any raw LLM.
It depends on the source material. Genre routing works best when the style matches the content. The most dramatic improvements over general translation appear in xianxia, wuxia, and romance — genres with strong community-standard terminology and distinct registers. Modern and urban novels show smaller but still meaningful improvements in natural phrasing.
Genre routing is applied at the novel level based on the primary genre tag. For novels that blend genres — a xianxia novel with romance subplots, or an urban novel with cultivation elements — the primary genre style handles the majority of content well. NoveLM does not currently switch styles mid-chapter, though individual passages naturally adapt because the underlying model adjusts to conversational vs. narrative context.
Current public workflow uses source-platform metadata and automatic genre routing. There is no public manual genre-override UI in the extension today.
Quality scoring adds minimal processing time to each chapter. It runs as a post-translation evaluation step and does not affect the translation itself. The score is generated and displayed alongside the translated chapter in the reader — there is no additional wait for the reader.