Linear Algebraic Structure of Word Senses, with Applications to Polysemy

2016/01/14 by Sanjeev Arora, Yuanzhi Li, Arora, Sanjeev +7 · 46 citations
Computer Science · Mathematics · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #cs.CL #cs.LG #stat.ML

paper · pdf · doi:10.48550/arxiv.1601.03764

Appear in the Transactions of the Association for Computational Linguistics 2018, link: https://transacl.org/ojs/index.php/tacl/article/view/1346

arxiv created 2018/12/07 · arxiv updated 2018/12/10

Abstract

Word embeddings are ubiquitous in NLP and information retrieval, but it is unclear what they represent when the word is polysemous. Here it is shown that multiple word senses reside in linear superposition within the word embedding and simple sparse coding can recover vectors that approximately capture the senses. The success of our approach, which applies to several embedding methods, is mathematically explained using a variant of the random walk on discourses model (Arora et al., 2016). A novel aspect of our technique is that each extracted word sense is accompanied by one of about 2000 "discourse atoms" that gives a succinct description of which other words co-occur with that word sense. Discourse atoms can be of independent interest, and make the method potentially more useful. Empirical tests are used to verify and support the theory.

Cited by

• A Geometric Unification of Concept Learning with Concept Cones
• Us-vs-Them bias in Large Language Models
• Unboxing the Black Box: Mechanistic Interpretability for Algorithmic Understanding of Neural Networks
• Efficient Toxicity Detection in Gaming Chats: A Comparative Study of Embeddings, Fine-Tuned Transformers and LLMs
• The Geometry of Reasoning: Flowing Logics in Representation Space
• From Defender to Devil? Unintended Risk Interactions Induced by LLM Defenses
• AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features
• Microsaccade-Inspired Probing: Positional Encoding Perturbations Reveal LLM Misbehaviours
• Binary Sparse Coding for Interpretability
• Negative Pre-activations Differentiate Syntax
• Analysis of Variational Sparse Autoencoders
• Un-Doubling Diffusion: LLM-guided Disambiguation of Homonym Duplication
• Neural Scaling Universality: If Exponents Are Fixed, Time to Understand Coefficients
• AdaptiveK Sparse Autoencoders: Dynamic Sparsity Allocation for Interpretable LLM Representations
• Attention Layers Add Into Low-Dimensional Residual Subspaces
• Measuring Media Criticism with ALC Word Embeddings
• Feature learning is decoupled from generalization in high capacity neural networks
• The Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability?
• Exploring Task Performance with Interpretable Models via Sparse Auto-Encoders
• Persona Features Control Emergent Misalignment
• Securing Language Models Against Vulnerability Encoding
• FaithfulSAE: Towards Capturing Faithful Features with Sparse Autoencoders without External Dataset Dependencies
• On the Limits of Sparse Autoencoders: A Theoretical Framework and Reweighted Remedy
• Dense SAE Latents Are Features, Not Bugs
• Transferring Linear Features Across Language Models With Model Stitching
• Interpretation Meets Safety: A Survey on Interpretation Methods and Tools for Improving LLM Safety
• Attention-Only Transformers via Unrolled Subspace Denoising
• From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit
• Localizing Persona Representations in LLMs
• Decom-Renorm-Merge: Model Merging on the Right Space Improves Multitasking
• Towards Interpretability Without Sacrifice: Faithful Dense Layer Decomposition with Mixture of Decoders
• The Origins of Representation Manifolds in Large Language Models
• Analyzing Fine-Grained Alignment and Enhancing Vision Understanding in Multimodal Language Models
• On the creation of narrow AI: hierarchy and nonlocality of neural network skills
• Understanding Task Representations in Neural Networks via Bayesian Ablation
• Where did the ambiguity go? Examining how multimodal models interpret polysemous words
• Superposition Yields Robust Neural Scaling
• Are Sparse Autoencoders Useful for Java Function Bug Detection?
• Geometry of Semantics in Next-Token Prediction: How Optimization Implicitly Organizes Linguistic Representations
• Beyond Input Activations: Identifying Influential Latents by Gradient Sparse Autoencoders
• Do Sparse Autoencoders Capture Concept Manifolds?
• When Does LeJEPA Learn a World Model?
• Towards Understanding the Nature of Attention with Low-Rank Sparse Decomposition
• Disentangling MLP Neuron Weights in Vocabulary Space
• Neuron Populations Exhibit Divergent Selectivity with Scale
• The Information Geometry of Softmax: Probing and Steering

Related

• A Simple and Efficient Method To Generate Word Sense Representations
• A Simple Approach to Learn Polysemous Word Embeddings
• Learning Word Sense Embeddings from Word Sense Definitions
• Geometry of Polysemy
• Understanding and Improving Multi-Sense Word Embeddings via Extended Robust Principal Component Analysis