A Case Study of Analysis of Construals in Language on Social Media Surrounding a Crisis Event

Lolo Aboufoul, Khyati Mahajan, Tiffany Gallicano, Sara Levens, Samira ShaikhACL 2021

TL;DR

This paper explores how people's language on social media shifts between abstract and concrete terms when discussing a major event, using Construal Level Theory. By applying topic modeling with concreteness as a covariate, they found that abstract language dominated discussions before and after the Charlottesville protests, while concrete language became more prevalent during and immediately after the event. This offers a novel way to computationally analyze psychological distance in real-world social media discourse.

What Problem Does This Paper Address?

Understanding how people perceive and communicate about events, especially crisis events, is crucial. Construal Level Theory (CLT) suggests that psychological distance influences how we mentally represent things, leading to more abstract thinking for distant events and more concrete thinking for proximal ones. While CLT has been studied in lab settings, its application to spontaneous, real-world social media discourse, particularly around significant events, remains less explored. The authors aim to bridge this gap by investigating whether language patterns on social media reflect these theoretical predictions about psychological distance and construal levels, and if these patterns can be computationally identified.

Methods

The researchers employed a novel application of topic modeling combined with psycholinguistic measures. Specifically, they used the Structural Topic Model (STM), which allows for the inclusion of covariates. The key innovation here is incorporating concreteness values of words as a covariate in the topic models.

Here's a breakdown:

  • Topic Modeling: This technique identifies latent themes (topics) within a collection of documents (tweets). The STM model was chosen for its ability to incorporate external information (covariates) to influence topic distributions.
  • Construal Level Theory (CLT): This theoretical framework guides the interpretation of language. CLT posits that as events become psychologically closer (e.g., in time or space), people tend to think about them more concretely, and as they become more distant, thinking becomes more abstract.
  • Concreteness Lexicon: The authors utilized an existing lexicon (Brysbaert et al., 2014a) to assign concreteness scores to words. A crucial methodological choice was to focus on words with extreme concreteness (high or low) to align with prior research suggesting extreme valence is highly correlated with emotion and memory.
  • Covariate Integration: By treating "concreteness" as a covariate in the STM, the model can analyze how the prevalence of different topics is influenced by the average concreteness of the language used within those topics.

Datasets

The study utilized a dataset of 526,102 tweets collected via the GNIP Historical Powertrack Twitter API. The data was gathered around the Charlottesville protests of August 2017, with a longitudinal collection period from February 7, 2017, through October 11, 2017. The data was collected using a carefully curated set of keywords and hashtags designed to capture discussions related to the event and its aftermath, split into two conditions to ensure comprehensive coverage.

Key Insights and Ideas

  • Social Media as an Ecological Valid Lab: The study highlights that social media provides a more ecologically valid environment for studying CLT compared to traditional lab settings, as communication is spontaneous and interactive.
  • Concreteness as a Distinguishing Factor: Incorporating concreteness values as covariates in topic models significantly enhances the analysis, allowing for a more nuanced understanding of language surrounding events.
  • Temporal Shifts in Construal: The most striking finding is the temporal shift in language construal. Abstract language was more prevalent before and after the event, while concrete language saw a surge during and immediately after the Charlottesville protests.
  • Interpreting Abstract vs. Concrete Language: Concrete terms (e.g., "statues," "trophies") allow for easier mental imagery and engagement, while abstract terms (e.g., "secession," "confederate") represent broader ideas or emotions.

Key Results

The study identified distinct topics within the tweet corpus, some characterized by high concreteness and others by low concreteness.

  • Topic 3 was identified as the least concrete, featuring terms like "outrage," "lead," and "nationalist."
  • Topic 9 was identified as the most concrete, with terms related to specific people or objects.
  • Temporal Analysis: Figure 3 visually demonstrates that the proportion of Topic 9 (concrete) increased significantly after the protests, peaking in the period immediately following the event. Conversely, Topic 3 (abstract) declined during the protests and remained lower afterward. This suggests a shift towards more concrete language when discussing the event in its immediate aftermath.

Limitations and Future Work

The authors acknowledge several limitations:

  • Single Event Focus: The analysis is limited to one event (Charlottesville protests). Generalizability to other crisis events needs further investigation.
  • Deeper Concrete Term Analysis: While concreteness was used as a covariate, a more in-depth study of the specific concrete vs. abstract terms within each topic could yield further insights.
  • Language Limitations: The study is limited to English. Replicating findings in other languages is necessary for broader generalizability, especially given the availability of multilingual concreteness lexicons.

Future work could involve applying this methodology to other crisis events, such as the COVID-19 pandemic, and conducting more granular analyses of the specific words contributing to concreteness scores within topics.

Summary updated

BibTeX
@inproceedings{aboufoul2021case,
  author = {Lolo Aboufoul and Khyati Mahajan and Tiffany Gallicano and Sara Levens and Samira Shaikh},
  title = {{A Case Study of Analysis of Construals in Language on Social Media Surrounding a Crisis Event}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure