关于Stanford NLP英文统计共指消解系统的训练语料及小说适配性问询

Great question! Let's break this down clearly for you:

1. Training Corpus #

Stanford NLP's statistical coreference resolution system is trained primarily on OntoNotes 5.0. This is a widely used multi-domain corpus that includes text from:

• Newswire articles (like CNN, AP)
• Broadcast news transcripts
• Conversational telephone speech
• Web text
• Encyclopedia entries (from Wikipedia)

It’s worth noting that OntoNotes contains very little literary or novel-style text, which directly impacts how well the system performs on that type of content.

2. Effectiveness on Novel Text #

The short answer is: it works as a basic baseline, but has notable limitations—here’s a deeper dive:

• Straightforward cases hold up: For simple, linear narrative segments (e.g., pronouns like "she" referring to a recently named character, or repeated mentions of a main character’s name), the system will often get things right. It’s a solid starting point for basic coreference tasks in novels.
• Struggles with literary complexity: Novels are full of elements the system wasn’t trained to handle, leading to errors:
  • Shifts in narrative perspective (e.g., switching between third-person limited viewpoints of different characters)
  • Metaphorical or symbolic references (e.g., a character being called "the raven-haired stranger" instead of their name, when multiple unknown characters are present)
  • Frequent introduction of minor characters with minimal context
  • Non-standard, stylized sentence structures common in literary prose
• Can be optimized for novels: If you need better performance, you could fine-tune the system on a novel-specific corpus (like public domain texts from Project Gutenberg) or add rule-based post-processing to handle common literary patterns the statistical model misses.

内容的提问来源于stack exchange，提问作者Axel Clerici