Patterns of Wisdom: Discourse-Level Style in Multi-Sentence Quotations

Kyle Booten, Marti A. HearstNAACL 2016

TL;DR

This paper investigates the stylistic patterns within multi-sentence quotations, moving beyond individual sentence features to analyze how elements are ordered across sentences. By identifying specific linguistic and syntactic patterns that are more likely to appear in the first or second sentence of a quotation, the authors demonstrate that quotations exhibit a more predictable structure than general text, suggesting that these patterns contribute to their memorability and "quotability."

What Problem Does This Paper Address?

Quotations are often memorable and impactful, but the linguistic features that contribute to this memorability are not fully understood. While previous work has characterized the stylistic properties of individual memorable phrases, this paper tackles a more nuanced question: how is the discourse-level style of multi-sentence quotations structured? Specifically, it asks whether there are predictable patterns in the ordering of linguistic elements across the sentences within a quotation, and if these patterns are unique to quotations compared to other forms of text. This research aims to empirically investigate the "rhetorical arts" of ancient scholars by analyzing contemporary digital text.

Methods

The authors employ a feature-based approach to analyze quotations. They define and extract a set of features that capture both low-level (unigrams) and high-level (syntactic parse tree structures, abstract/general word usage) stylistic properties. The core of their methodology involves analyzing the positional tendencies of these features within two-sentence quotations. They use statistical tests (specifically, chi-squared tests) to determine if a feature is significantly more likely to appear in the first or second sentence. To evaluate the distinctiveness of these patterns, they design a "Quote Ordering Task" where a classifier attempts to distinguish between original and reversed two-sentence quotations. This task is framed as a measure of how "formulaic" or predictable the sentence order is for quotations compared to non-quotations.

Datasets

The paper utilizes two primary types of datasets:

  • Quotations:
    • Quotations 1 (Tumblr): A dataset of two-sentence quotations collected from Tumblr, filtered using the "#quote" hashtag. (n=4237)
    • Quotations 2 (Tumblr): A separate test set of two-sentence quotations from Tumblr, also filtered with "#quotation." (n=1846)
  • Non-Quotes (Comparison Corpus):
    • Non-Quotes 1 (Brown Corpus): Two-sentence paragraphs from the Brown Corpus, used for cross-validation. (n=1846)
    • Non-Quotes 2 (Brown Corpus): Sequential pairs of sentences randomly chosen from paragraphs longer than two sentences in the Brown Corpus. (n=1846)

Key Insights and Ideas

  • Quotations have a predictable sentence order: The core insight is that the sequence of sentences in a quotation is not random. Certain stylistic patterns are consistently found in either the first or second sentence, making the order more predictable than in general text.
  • Negative-to-Positive constructions are common: Quotations frequently start with a negative statement or commandment and follow with a positive one, often employing parallel structures like antimetabole. This suggests a rhetorical strategy of setting up a contrast or limitation before offering a resolution or elaboration.
  • Conjunctions signal cohesion, but with a quotable twist: While conjunctions like "but" and "and" are generally used to link sentences, they appear significantly more often in the second sentence of quotations than in non-quotations. This might create a "dramatic pause" or emphasize the connection between the two parts of the quote.
  • Sweeping declarations often start quotations: Quotations tend to begin with broad, abstract concepts or generalizations (e.g., "people," "life," "truth"), with the second sentence elaborating on or illustrating these initial declarations.
  • The "formulaic" nature of quotations: The success of the Quote Ordering Task suggests that quotations, as a genre, adhere to underlying stylistic schemas that make their structure more discernible and predictable than that of general text. This predictability is posited as a contributor to their memorability.

Key Results

  • Quote Ordering Task Performance:
    • In a 5-fold cross-validation on balanced subsets, classifiers achieved:
  • 62.6% accuracy on Quotations 1.
  • 56.0% on Non-Quotes 1.
  • 52.9% on Non-Quotes 2.
  • On a separate test set (Quotations 2), performance was 63.7% for unigrams and 63.9% for all features.
  • The performance on quotations was consistently higher than on non-quote corpora (p<.01), indicating a statistically significant difference in predictability.
  • Feature Tendencies:
    • Unigrams like "it," "but," and "they" were significantly more likely to appear in the second sentence.
  • High-level syntax patterns, particularly those involving conjunctions (e.g., CC + NP + VP + .), were significantly more likely to occur in the second sentence of quotations.
  • Patterns indicating questions (e.g., WHNP + SQ + .) were more frequent in the first sentence.
  • Abstract and General Nouns as subjects were more prevalent in the first sentence of quotations.

Limitations and Future Work

  • Limitations:
    • The study primarily focuses on two-sentence quotations, which may not represent the full diversity of multi-sentence quotes.
  • The "Quote Ordering Task" relies on automated classification, and it's unclear how consciously aware users are of these patterns.
  • Integrating the analyzed features into classification tasks proved challenging, with only slight benefits observed for non-quotes.
  • Future Work:
    • Compare stylistic patterns across a wider variety of text genres.
  • Investigate the relationship between these stylistic patterns and specific rhetorical goals.
  • Conduct user studies to determine if people are consciously aware of these stylistic elements when posting or consuming quotations.
  • Explore why people choose to share specific quotations.
  • Apply discourse-level stylistic analysis to applications like automatically culling multi-sentence quotations and improving "creative text retrieval."
    • Use these patterns to recommend stylistically similar quotations to users on social media.

Summary updated

BibTeX
@inproceedings{booten2016patterns,
  author = {Kyle Booten and Marti A. Hearst},
  title = {{Patterns of Wisdom: Discourse-Level Style in Multi-Sentence Quotations}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2016},
}

Loading related content...

Figure