From Shakespeare to Twitter: What are Language Styles all about?

Wei Xu2017 · DOI

Abstract: As natural language processing research is growing and largely driven by the availability of data, we expanded research from news and small-scale dialog corpora to web and social media. User-generated data and crowdsourcing opened the door for investigating human language of various styles with more statistical power and real-world applications. In this position/survey paper, I will review and discuss seven language styles that I believe to be important and interesting to study: influential work in the past, challenges at the present, and potential impact for the future.

Introduction

For decades, Natural Language Processing (NLP) research was largely confined to "clean" data environments, such as professionally edited news articles and small-scale, controlled dialogue corpora. However, the explosion of user-generated content on the web has forced a paradigm shift. Researchers must now confront the reality that human language is not monolithic; it varies significantly based on the author's identity, the intended audience, and the medium of communication. This paper addresses the ambiguity of "style"—an umbrella concept covering genre, domain, register, and personal attributes—and explores how we can computationally identify, analyze, and generate these varied linguistic forms.

The Seven Dimensions of Style

To provide structure to the broad concept of style, this survey categorizes research into seven distinct dimensions, each presenting unique computational challenges:

  1. Text Simplification: Moving beyond simple sentence compression, simplification requires a complex balance of paraphrasing and re-wording to reduce cognitive load while maintaining core meaning. Key benchmarks include Simple English Wikipedia and the professionally edited Newsela corpus.
  2. Historical Language: Research into historical archives, such as the Primeros Libros corpus or Shakespearean plays, faces challenges ranging from Optical Character Recognition (OCR) errors to evolving grammar and vocabulary.
  3. Internet and Colloquial Speech: The use of the Twitter streaming API and the PIT-2015 corpus has allowed researchers to study the informal, rapidly evolving language of social media.
  4. Personalized Attributes: This involves profiling authors based on age, gender, and personality. A major insight here is the difficulty of isolating "pure" style from topical preference (e.g., do men and women write differently because of style, or because they discuss different subjects?).
  5. Politeness: Politeness is not merely the absence of rudeness; it is often encoded in specific "function words" and markers (e.g., "I appreciate that"). Current classifiers can predict politeness with approximately 80% accuracy.
  6. Framing and Persuasion: On platforms like Reddit’s ChangeMyView, stylistic choices often have more predictive power regarding a user's persuasiveness than the actual logical arguments, though this remains difficult to model (approx. 60-65% accuracy).
  7. Abusive Language: Detecting bullying is a "needle in a haystack" problem. Research shows that "bullying traces" appear in only 0.02% to 0.73% of random social media samples, making data collection and detection exceptionally difficult.

Methodological Evolution

The field has transitioned from early rule-based systems to statistical machine learning, and most recently, to neural generation models. A primary focus of modern methodology is Style Transfer—the task of altering the style of a sentence (e.g., modern to Shakespearean) while preserving its underlying semantics. This often involves multi-task learning to isolate stylistic features from topical ones and the use of semantic parsing for instructional language.

The "Three Datas" Challenge

Despite recent progress, three primary obstacles—the "Three Datas"—hinder the field:

  • Data Shortage: There is a chronic lack of parallel stylistic data (the same content written in two different styles).
  • Data Fusion: Style and topic are often inextricably linked, making it difficult for models to learn stylistic markers without being biased by the subject matter.
  • Data Annotation: Stylistic labels are inherently subjective, leading to inconsistencies in human-labeled datasets.

Limitations and Future Directions

A significant bottleneck in style research is Evaluation. Standard automatic metrics like BLEU are often insufficient for capturing the nuance of stylistic change. While human evaluation using Likert scales remains the gold standard, it is prone to bias.

Looking forward, the development of neural generation models that require less parallel data is a priority. Furthermore, there is significant potential for "social justice" applications, such as using stylistic obfuscation to preserve user anonymity or simplifying complex legal and medical documents to make essential information more accessible to the general public.

Summary updated

Abstract

As natural language processing research is growing and largely driven by the availability of data, we expanded research from news and small-scale dialog corpora to web and social media. User-generated data and crowdsourcing opened the door for investigating human language of various styles with more statistical power and real-world applications. In this position/survey paper, I will review and discuss seven language styles that I believe to be important and interesting to study: influential work in the past, challenges at the present, and potential impact for the future.
BibTeX
@misc{xu2017from,
  author = {Wei Xu},
  title = {{From Shakespeare to Twitter: What are Language Styles all about?}},
  year = {2017},
  doi = {10.18653/v1/w17-4901},
}
Figure