Sparse AutoEncoders (SAEs) have recently been employed as a promising unsupervised approach for understanding the representations of layers of Large Language Models (LLMs). However, with the growth in model size and complexity, training SAEs is computationally intensive, as typically one SAE is trained for each model layer. To address such limitation, we propose Group-SAE, a novel strategy to train SAEs. Our method considers the similarity of the residual stream representations between contiguous layers to group similar layers and train a single SAE per group. To balance the trade-off between efficiency and performance, we further introduce AMAD (Average Maximum Angular Distance), an empirical metric that guides the selection of an optimal number of groups based on representational similarity across layers. Experiments on models from the Pythia family show that our approach significantly accelerates training with minimal impact on reconstruction quality and comparable downstream task performance and interpretability over baseline SAEs trained layer by layer. This method provides an efficient and scalable strategy for training SAEs in modern LLMs.
State-of-the-art approaches for metaphor detection compare their literal - or core - meaning and their contextual meaning using metaphor classifiers based on neural networks. However, metaphorical expressions evolve over time due to various reasons, such as cultural and societal impact. Metaphorical expressions are known to co-evolve with language and literal word meanings, and even drive, to some extent, this evolution. This poses the question of whether different, possibly time-specific, representations of literal meanings may impact the metaphor detection task. To the best of our knowledge, this is the first study that examines the metaphor detection task with a detailed exploratory analysis where different temporal and static word embeddings are used to account for different representations of literal meanings. Our experimental analysis is based on three popular benchmarks used for metaphor detection and word embeddings extracted from different corpora and temporally aligned using different state-of-the-art approaches. The results suggest that the usage of different static word embedding methods does impact the metaphor detection task and some temporal word embeddings slightly outperform static methods. However, the results also suggest that temporal word embeddings may provide representations of the core meaning of the metaphor even too close to their contextual meaning, thus confusing the classifier. Overall, the interaction between temporal language evolution and metaphor detection appears tiny in the benchmark datasets used in our experiments. This suggests that future work for the computational analysis of this important linguistic phenomenon should first start by creating a new dataset where this interaction is better represented.
Understanding differences of viewpoints across corpora is a fundamental task for computational social sciences. In this paper, we propose the Sliced Word Embedding Association Test (SWEAT), a novel statistical measure to compute the relative polarization of a topical wordset across two distributional representations. To this end, SWEAT uses two additional wordsets, deemed to have opposite valence, to represent two different poles. We validate our approach and illustrate a case study to show the usefulness of the introduced measure.