Introduction
Writing a persuasive argumentative essay is a complex task that requires more than just grammatical proficiency; it demands the strategic use of evidence. Anecdotes—short, illustrative stories about specific individuals—are among the most effective forms of evidence. However, writers often struggle to recall or find stories that perfectly align with their specific argumentative goals.
This paper addresses the "semantic gap": the distance between the objective facts contained within a story (e.g., the events of Steve Jobs' life) and the subjective intent of a writer (e.g., arguing for "The Value of Life"). To solve this, the authors introduce a system that automatically extracts anecdotes from a large corpus and recommends them based on a user's chosen topic.
System Architecture
The authors propose a two-stage framework consisting of Anecdote Recognition and Anecdote Recommendation. This pipeline transforms raw essay text into a searchable database of structured evidence.
Anecdote Recognition
Instead of treating anecdotes as unstructured text, the system represents them as a structured tuple: <person, story, implication>. This structure is key to bridging the semantic gap by explicitly linking a narrative to its argumentative purpose.
The Extraction Process
- Discourse Role Identification: Using a Linear-chain Conditional Random Field (CRF), the system labels sentences with roles such as Thesis, Main Idea, Support, or Conclusion. This allows the system to distinguish between a writer stating an opinion and a writer providing narrative evidence.
- Story Extraction: The system identifies "human mentions" (names, titles, pronouns) and extracts event chains (agent-predicate-recipient) specifically from sections labeled as "Support."
- Implication Linking: This is the system's "secret sauce." The authors assume the implication of a story is found in the Thesis or Main Idea sentences it supports. A logistic regression classifier links stories to these nearby blocks based on paragraph distance and shared keywords.
Anecdote Recommendation
Once a database is populated (the authors extracted 26,060 anecdotes from 16,618 Chinese essays), the system uses Learning to Rank (LTR) to suggest them to users. They compared two primary strategies:
- Query-Story (QS): Matching the user's topic to the narrative text of the story.
- Query-Implication (QI): Matching the user's topic to the extracted "point" or implication of the story.
Key Results
The evaluation revealed that the Query-Implication (QI) approach vastly outperformed the Query-Story (QS) approach. Using BM25, the QI method achieved a Mean Average Precision (MAP) of 0.738, compared to just 0.357 for the QS method. This confirms that factual descriptions of events rarely contain the abstract keywords (like "persistence" or "sacrifice") that writers use when searching for topics; the extracted implication serves as the necessary bridge.
Insights and Limitations
Human evaluations judged over 60% of the automatically extracted anecdotes as "Good" (understandable and correctly paired with an implication). However, the system faces several challenges:
- Stance Detection: The system currently identifies relevance but cannot distinguish if an anecdote supports or attacks a specific topic.
- Extraction Errors: Failures in Named Entity Recognition (NER) or pronoun resolution accounted for most "Poor Story" results.
- Rhetorical Queries: The system struggles with highly abstract or metaphorical topics that lack direct keyword or embedding matches.
By turning unstructured text into interpretable <person, story, implication> tuples, this research provides a significant step forward in supporting the planning and composition phases of writing.
Summary updated