Uncovering Semantic Shifts in Climate Discourse among Dutch Parliamentarians: A Dynamic Bernoulli Embeddings Approach

Traditionally, the detection of semantic change has been studied primarly by historians and linguists. However, the advent of new computational methods has enabled more efficient analysis of word usage in large corpora. Previous studies in this domain have relied mainly on Static Word Embeddings Models, which are faced with the challenge of solving the alignment problem when comparing independently trained embeddings. This paper contributes to the field of Natural Language Processing by employing Dynamic Bernoulli Embeddings (D-EMBs) on a Dutch corpus, effectively avoiding the alignment problem.

The objective of this work is to investigate semantic shifts in the climate discourse within the Dutch Parliament from 1995 to 2019. Given the influential role of parliamentarians in formulating national policies and shaping public perceptions, it is crucial to analyze how their language surrounding the climate debate changes or evolves over time. For this work, we consider three target words: (1) `klimaat' (climate), (2) `klimaatverandering' (climate change), and (3) `klimaatbeleid' (climate policy).

The findings obtained using D-EMBs indicate that a significant semantic shift has occurred in the climate debate. Notably, the analysis reveals that `klimaat' (climate) underwent the most substantial changes among the three words. In 1995, its nearest neighbors encompassed diverse contexts such as `school climate` and `fickleness`, while in 2019, the nearest neighbors predominantly consisted of terms related to climate (change) policies, such as `sustainability', `energy policy', and `environmental policy'. We ought to note, however, that while various evaluation methods were employed alongside a more in-depth analysis (through the nearest neighbours), the absence of a definitive gold standard test for detecting semantic shifts, particularly in the presence of polysemy and homonymy, remains a challenge. Further research is recommended to enhance these evaluation methods and develop more robust techniques for semantic change detection.