The Voice of Style in Rewriting: Why Rules Don't Capture Tone
Automated news rewriting systems break style down into aspects: structure, tone, vocabulary, headlines, and other components. Each aspect analyzes sample texts from the target media outlet and generates rule descriptions. These rules are compiled into a prompt generation profile. This approach accurately reproduces formal characteristics but falls short in conveying the text's 'voice'—the implicit lexical choices that create emotional impact.
The problem emerged during testing on a publication with a distinct identity. The system generates correct news stories but lacks the characteristic tone.
Testing on a Real Case
We trained the model on the style of 'Fontanka'. Input text: the detention of a 16-year-old teenager in Ufa for planning a terrorist attack in a church and recruiting high school students. Original from the Investigative Committee:
In Ufa, a teenager was detained who recruited high school students and planned a terrorist attack in an Orthodox church. The FSB detained a 16-year-old boy from Bashkortostan...
System generation:
In Ufa, a 16-year-old teenager was detained for planning a terrorist attack in an Orthodox church. The FSB detained a 16-year-old boy from Bashkortostan. According to the Investigative Committee, the teenager recruited high school students...
Google AdInline article slot
Original 'Fontanka':
In Ufa, a schoolboy was detained; he is suspected of preparing a terrorist attack in a church on the orders of handlers. The boy, according to law enforcement, had been recruited.
The rewrite is structurally correct and factual but loses the key effect: 'boy' instead of 'teenager' creates a contrast between a childlike image and a terrorist threat. This isn't a stylistic rule but a contextual word choice.
What Aspects Capture and What They Miss
Aspects describe:
- Sentence length and lead structure.
- Frequency of passive voice.
- Rules for attributing sources.
- Lexical preferences by frequency.
They don't capture:
- Synonym choice for emotional contrast ('boy' vs 'teenager').
- Embedded distance ('according to the investigation' as a signal of unobtrusiveness).
- Clashing registers to enhance tone.
The tone aspect outputs 'neutral-informational with conversational elements'—formally correct but lacking nuance.
Rule Limitations in Conveying Voice
Rules work for constraints: 'Petersburg instead of Saint Petersburg', attribution once per text. They set structure: lead with a fact, headline with a result. But voice is about local decisions: a word in the context of a sentence.
Attempts to formalize ('less formality next to high formality') fail: it depends on topic and context. Universal instructions don't exist—voice can't be packaged into rules.
Solutions and Pipeline Improvements
Complete voice transmission is unattainable. Current measures:
- Increasing examples to 10–15 per topic (crime to crime).
- Checking style prohibitions (previously ignored).
- Controlling fact-extraction completeness (verifying all facts).
- Selecting examples by tags, not randomly.
These fixes measurably improve structure and accuracy. Voice is imitated through examples but not guaranteed.
For MVP: accurate structure + approximated voice is better than a full illusion of identity.
Key Takeaways
- The aspect approach is ideal for mid-tier publications without a strong 'voice'—their style reduces to rules.
- Top publications with identity ('Fontanka') reveal the ceiling: imitation, not replication.
- Examples are more effective than abstract descriptions for lexical nuances.
- Voice is editorial intuition, developed over years; models copy but don't understand.
- The pipeline is evolving: from rules to demonstration on thematic samples.
Final Conclusions
Rewriting systems have achieved accuracy in structure and facts. Transitioning to 'just like the original' requires a different approach: focus on examples over rules. Testing on complex styles immediately showed limits, allowing product expectations to be adjusted.
— Editorial Team
No comments yet.