Sudden Semantic Shifts in Swedish NATO discourse

Brian Bonafilia, Bastiaan Bruinsma, Denitsa Saynova, Moa JohanssonACL 2023

TL;DR

This paper uses word embeddings to track how the meaning of words related to NATO shifted in Swedish Twitter discourse following Russia's invasion of Ukraine. They found that while word embeddings can capture these sudden semantic shifts, distinguishing genuine signals from noise is challenging due to data sparsity and the need for domain expertise.

What Problem Does This Paper Address?

Understanding how public opinion changes, especially during significant geopolitical events, is crucial. This paper investigates "sudden semantic shifts" – rapid changes in word meaning that reflect evolving public sentiment. Specifically, they focus on Sweden's decision to apply for NATO membership in 2022, a pivotal moment influenced by the Russian invasion of Ukraine. The challenge lies in accurately detecting and interpreting these shifts using computational methods, particularly when public engagement with the topic might be low, making it hard to separate meaningful signals from random fluctuations in language.

Methods

The authors employ word embeddings, specifically Word2Vec (both Continuous Bag-of-Words and Skip-gram architectures), to analyze semantic shifts. They train custom models on Swedish Twitter data, fine-tuning them from a general Swedish pre-trained model. To track shifts, they compare the similarity rankings of words across different time periods (pre-invasion, post-invasion, pre-application, post-application). They use Rank-Biased Overlap (RBO) to quantify the similarity between word rankings from different periods and the pre-trained model, indicating the degree of semantic shift. Their analysis involves a multi-stage word selection process, combining domain knowledge, identification of polarizing terms, and statistical measures like RBO to pinpoint words exhibiting significant changes.

Datasets

The primary dataset consists of Swedish-language Twitter posts ("tweets") collected over a year (September 11, 2021, to September 11, 2022) that contained keywords related to NATO. This resulted in 1,188,556 tweets from 64,315 users. For pre-training their Word2Vec models, they utilized a large corpus of Swedish media text (Göteborgs-Posten, SVT, and Wikipedia) from 2003 to 2014, totaling 759 million tokens.

Key Insights and Ideas

  • Sudden events drive semantic shifts: The paper demonstrates that major geopolitical events, like the invasion of Ukraine, can indeed trigger rapid changes in how words are understood and associated.
  • Word embeddings as a lens for public opinion: Word embeddings, when applied to social media data, can serve as a valuable tool to track evolving public sentiment and understanding of complex topics.
  • The challenge of signal vs. noise: Distinguishing genuine semantic shifts from random linguistic variations is a significant hurdle, especially with limited public engagement on a topic. Domain knowledge is crucial for interpretation.
  • Different models capture different nuances: The CBOW and Skip-gram architectures, while both word embedding methods, capture different aspects of semantic relationships, leading to distinct insights.
  • Context matters for interpretation: Even seemingly neutral words can acquire new meanings or associations during specific events, as seen with "förskolor" (preschools) and its connection to nuclear weapons discussions.

Key Results

The study identified several words that showed significant semantic shifts in Swedish Twitter discourse related to NATO. For instance, the word "natoansökan" (NATO application) became strongly associated with "destabiliserande" (destabilizing) in the post-invasion period, reflecting concerns about the implications of Sweden's potential membership. The word "säkerhet" (security) became linked to "suveränitet" (sovereignty) as the debate shifted towards Sweden's neutrality. The paper also noted that the word "azov" shifted from referring to geographical locations to being associated with the Azov Battalion and the Siege of Mariupol.

Limitations and Future Work

The authors acknowledge that the sparsity of their dataset, particularly the non-representative distribution of NATO-related words, made it challenging to study search words effectively. They suggest extending the search criteria to capture a broader representation of language. They also propose exploring alternative embedding models like FastText and GloVe, which might offer different perspectives. Practically, they note that their assumption of NATO membership being a major electoral issue was not fully realized, and Sweden's aspirations remain unfulfilled. Therefore, they recommend extending the data collection period to gain a more comprehensive view of any shifts.

Summary updated

BibTeX
@inproceedings{bonafilia2023sudden,
  author = {Brian Bonafilia and Bastiaan Bruinsma and Denitsa Saynova and Moa Johansson},
  title = {{Sudden Semantic Shifts in Swedish NATO discourse}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure