ReFRAME or Remain: Unsupervised Lexical Semantic Change Detection with Frame Semantics

TL;DR

The paper investigates unsupervised lexical semantic change detection using Frame Semantics, arguing that shifts in the frames a word participates in (as a trigger or frame element) reflect diachronic meaning changes. It employs FrameNet parsing via a Frame Semantic Transformer to extract frame distributions for English target lemmas across two SemEval-2020 time slices, comparing frame-element only (FE) and frame-trigger+frame-element (FTFE) distributions with Jensen–Shannon Divergence. The approach yields competitive results on Subtask 2 (ranking) and Subtask 1 (binary change), with FTFE achieving a Spearman correlation of $0.306$ and FE $0.249$, and binary accuracy of $0.622$, outperforming many baseline embeddings while offering rich, per-frame interpretability. Qualitative analyses demonstrate that frame shifts align with plausible semantic changes, while also highlighting limitations related to frame inventory coverage and parser reliability, suggesting frame-based signals as a valuable, complementary tool to neural methods for diachronic semantics.

Abstract

The majority of contemporary computational methods for lexical semantic change (LSC) detection are based on neural embedding distributional representations. Although these models perform well on LSC benchmarks, their results are often difficult to interpret. We explore an alternative approach that relies solely on frame semantics. We show that this method is effective for detecting semantic change and can even outperform many distributional semantic models. Finally, we present a detailed quantitative and qualitative analysis of its predictions, demonstrating that they are both plausible and highly interpretable

Figures (5)

• Figure 1: Changes in the frequencies of 6 frames that the English noun plane participated in between the 2 sub-corpora for English of the SemEval 2020 shared task 1 schlechtweg_semeval-2020_2020. Plane is annotated as semantically changed in the dataset
• Figure 2: The JSD contributions of different frames of the English noun prop. Green bars indicate an increase in relative frequency; red bars indicate a decrease.
• Figure 3: The JSD contributions of different frames of the English noun tree. Green bars indicate an increase in relative frequency; red bars indicate a decrease.
• Figure 4: The JSD contributions of different frames of the English noun quilt. Green bars indicate an increase in relative frequency; red bars indicate a decrease.
• Figure 5: The JSD contributions of different frames of the English noun ball. Green bars indicate an increase in relative frequency; red bars indicate a decrease.