Abstract: The complex sentences of news wire reports contain floating content units that appear to be opportunistically placed where the form of the surrounding text allows. We present a corpus analysis that identified precise semantic and syntactic constraints on where and how such information is realized. The result is a set of revision tools that form the rule base for a report generation system, allowing incremental generation of complex sentences.
1. Introduction
Generating the "lead" sentence of a news story is a significant challenge for Natural Language Generation (NLG). In sports journalism, these sentences are often incredibly dense, averaging 21 to 46 words, and packed with both fixed facts (e.g., final score, date) and "floating" facts (e.g., historical context, season highs, or streaks).
Traditional NLG systems struggle because floating content lacks a fixed syntactic home; it appears wherever the surrounding structure has "room." Researchers at Columbia University propose a "draft-and-revise" strategy to address this. Instead of attempting to generate complex structures in a single pass, the system generates a basic draft and then iteratively revises it to include optional, interesting details.
2. Methodology: The Revision Toolbox
The authors performed a detailed corpus analysis of 800 basketball summaries from the UPI newswire to understand how humans perform these revisions. The core of their method involves Semantic Decrement Pairing, where they identified pairs of sentences that were semantically identical except for a single piece of information.
By comparing these pairs, they identified the minimal syntactic transformations required to turn a simple sentence into a complex one. These transformations were categorized into a "Revision Toolbox":
- Simple Revisions: Additions that preserve the original structure, such as Adjoin (adding an adjective) or Conjoin (adding a clause).
- Complex Revisions: Non-monotonic changes that rebuild the sentence structure, such as Nominalization (e.g., turning the verb "defeated" into the noun "defeat" to accommodate modifiers).
- Side Transformations: "Clean-up" tasks like Reference Adjustment, which modifies phrasing to avoid repetition after new information is added.
3. Structural Insights: The Hub Model
A key discovery of the analysis is that both simple and complex sentences share a common top-level "skeleton." Sentences are typically built around two main clusters: the notable statistic cluster (detailing player performance) and the game result cluster (detailing the outcome).
Floating information is not placed randomly; it gravitates toward one of these two hubs based on the "local constraints" of the draft. For instance, if the draft uses a verb, a revision might use an adverbial phrase; if it uses a noun, an appositive might be employed. This opportunistic placement suggests that human writers may use a similar revision-based process to manage the cognitive load of composing dense reports.
4. Dataset and Results
The primary dataset consisted of 800 basketball game summaries, with a focused sub-corpus of 300 lead sentences containing four fixed content units (winner, loser, score, date) plus various floating units.
The analysis yielded 270 surface decrement pairs, forming the rule base for the STREAK generation system. Key findings include:
- Tool Frequency: Conjoin was the most frequent tool (127 instances), followed by Adjoin (88).
- Information Density: Floating content accounts for over 40% of the information in lead sentences, demonstrating why its inclusion is vital for high-quality summaries.
- Cross-Domain Potential: Preliminary analysis suggests these revision tools are domain-independent, as similar linguistic constructions appear in stock market reporting.
5. Conclusion and Future Work
While the current analysis was largely manual, future work aims to automate the process using regular expression-matching tools like CREP. Additionally, the authors plan to test the STREAK system's revision tools in other domains, such as financial news, to further validate the abstract nature of these linguistic resources. Although the paper establishes a robust rule base, quantitative evaluation of the system's output against human writing remains a goal for future research.
Summary updated