Shared Global and Local Geometry of Language Model Embeddings

• URL: http://arxiv.org/abs/2503.21073v2
• Date: Thu, 24 Apr 2025 02:54:29 GMT
• Title: Shared Global and Local Geometry of Language Model Embeddings
• Authors: Andrew Lee, Melanie Weber, Fernanda Viégas, Martin Wattenberg,
• Abstract summary: We find that token embeddings of language models exhibit common geometric structure.<n>We show that tokens with lower intrinsic dimensions often have semantically coherent clusters, while those with higher intrinsic dimensions do not.<n>Perhaps most surprisingly, we find that alignment in token embeddings persists through the hidden states of language models.
• Score: 46.33317507982751
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Researchers have recently suggested that models share common representations. In our work, we find that token embeddings of language models exhibit common geometric structure. First, we find ``global'' similarities: token embeddings often share similar relative orientations. Next, we characterize local geometry in two ways: (1) by using Locally Linear Embeddings, and (2) by defining a simple measure for the intrinsic dimension of each token embedding. Our intrinsic dimension demonstrates that token embeddings lie on a lower dimensional manifold. We qualitatively show that tokens with lower intrinsic dimensions often have semantically coherent clusters, while those with higher intrinsic dimensions do not. Both characterizations allow us to find similarities in the local geometry of token embeddings. Perhaps most surprisingly, we find that alignment in token embeddings persists through the hidden states of language models, allowing us to develop an application for interpretability. Namely, we introduce Emb2Emb, a simple method to transfer steering vectors from one language model to another, despite the two models having different dimensions.

Related papers

• The structure of the token space for large language models [1.5621144215664768]
  Large language models encode the correlational structure present in natural language by fitting segments of utterances (tokens) into a high dimensional ambient latent space upon which the models then operate. We present estimators for the dimension and Ricci scalar curvature of the token subspace, and apply it to three open source large language models of moderate size. We find that the dimension and curvature correlate with generative fluency of the models, which suggest that these findings have implications for model behavior.
  arXiv  Detail & Related papers  (2024-10-11T17:07:15Z)
• Concept Space Alignment in Multilingual LLMs [47.633314194898134]
  We show that multilingual models suffer from two familiar weaknesses: generalization works best for languages with similar typology, and for abstract concepts. For some models, prompt-based embeddings align better than word embeddings, but the projections are less linear.
  arXiv  Detail & Related papers  (2024-10-01T21:21:00Z)
• Lexinvariant Language Models [84.2829117441298]
  Token embeddings, a mapping from discrete lexical symbols to continuous vectors, are at the heart of any language model (LM) We study textitlexinvariantlanguage models that are invariant to lexical symbols and therefore do not need fixed token embeddings in practice. We show that a lexinvariant LM can attain perplexity comparable to that of a standard language model, given a sufficiently long context.
  arXiv  Detail & Related papers  (2023-05-24T19:10:46Z)
• Duality-Induced Regularizer for Semantic Matching Knowledge Graph Embeddings [70.390286614242]
  We propose a novel regularizer -- namely, DUality-induced RegulArizer (DURA) -- which effectively encourages the entities with similar semantics to have similar embeddings. Experiments demonstrate that DURA consistently and significantly improves the performance of state-of-the-art semantic matching models.
  arXiv  Detail & Related papers  (2022-03-24T09:24:39Z)
• All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality [5.203329540700176]
  We call into question the informativity of such measures for contextualized language models. We find that a small number of rogue dimensions, often just 1-3, dominate similarity measures.
  arXiv  Detail & Related papers  (2021-09-09T16:45:15Z)
• The Low-Dimensional Linear Geometry of Contextualized Word Representations [27.50785941238007]
  We study the linear geometry of contextualized word representations in ELMO and BERT. We show that a variety of linguistic features are encoded in low-dimensional subspaces.
  arXiv  Detail & Related papers  (2021-05-15T00:58:08Z)
• Quadric hypersurface intersection for manifold learning in feature space [52.83976795260532]
  manifold learning technique suitable for moderately high dimension and large datasets. The technique is learned from the training data in the form of an intersection of quadric hypersurfaces. At test time, this manifold can be used to introduce an outlier score for arbitrary new points.
  arXiv  Detail & Related papers  (2021-02-11T18:52:08Z)
• Learning Contextualised Cross-lingual Word Embeddings and Alignments for Extremely Low-Resource Languages Using Parallel Corpora [63.5286019659504]
  We propose a new approach for learning contextualised cross-lingual word embeddings based on a small parallel corpus. Our method obtains word embeddings via an LSTM encoder-decoder model that simultaneously translates and reconstructs an input sentence.
  arXiv  Detail & Related papers  (2020-10-27T22:24:01Z)
• Learning Universal Representations from Word to Sentence [89.82415322763475]
  This work introduces and explores the universal representation learning, i.e., embeddings of different levels of linguistic unit in a uniform vector space. We present our approach of constructing analogy datasets in terms of words, phrases and sentences. We empirically verify that well pre-trained Transformer models incorporated with appropriate training settings may effectively yield universal representation.
  arXiv  Detail & Related papers  (2020-09-10T03:53:18Z)
• LNMap: Departures from Isomorphic Assumption in Bilingual Lexicon Induction Through Non-Linear Mapping in Latent Space [17.49073364781107]
  We propose a novel semi-supervised method to learn cross-lingual word embeddings for bilingual lexicon induction. Our model is independent of the isomorphic assumption and uses nonlinear mapping in the latent space of two independently trained auto-encoders.
  arXiv  Detail & Related papers  (2020-04-28T23:28:26Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.