TLDR: A new study on RoBERTa-base, a transformer language model, reveals that its static word embeddings encode a wide range of semantic information. By clustering these embeddings and testing them against psycholinguistic measures like valence and concreteness, researchers found strong evidence that LLMs maintain a “lexical store” of word meanings, challenging theories that suggest meaning is solely derived from context.
Large language models (LLMs) have transformed how we interact with artificial intelligence, but a fundamental question remains: do these models truly “understand” the words they use? A recent research paper delves into this intriguing question, exploring how word meanings are represented within the core architecture of transformer-based LLMs.
The study, titled “Word Meanings in Transformer Language Models,” by Jumbly Grindrod and Peter Grindrod, investigates whether these AI systems employ something akin to a “lexical store,” where each word has an entry containing rich semantic information. This research challenges the idea that LLMs might process language without inherently storing word meanings, a concept sometimes referred to as “meaning eliminativism.”
At the heart of transformer models are two types of word representations: static embeddings and contextualized embeddings. Static embeddings are the invariant representations assigned to each word in the model’s vocabulary, serving as initial input. Contextualized embeddings, on the other hand, are dynamic representations that capture a word’s meaning based on its specific use in a given text. This paper focuses specifically on the static embeddings to see if they hold semantic information before any contextual processing occurs.
Exploring the Static Embedding Space
To investigate this, the researchers extracted the static token embedding space from RoBERTa-base, a widely used open-source transformer model. They then used a technique called k-means clustering to group similar words into 200 distinct clusters. The subsequent analysis involved two main studies.
The first study involved a manual inspection of these 200 clusters. While some clusters contained less interesting elements like special symbols or word parts, a significant number revealed a clear sensitivity to the meanings of the terms. For instance, clusters were found to group words by categories such as first names, work sectors, negative and positive terms, colors, tools, musical terms, medical terms, sports, companies, religious terms, clothing, financial terms, stages of life, official roles, political names, and even abstract concepts like “combination” or “division.” This manual review provided strong initial evidence that semantic information is indeed encoded within these static embeddings.
Also Read:
- How Language Models Learn Across Languages: The Role of Unified Representations
- Unpacking LLM Strategic Reasoning: A New Benchmark Approach
Quantitative Analysis with Psycholinguistic Measures
For their second study, the researchers employed a more quantitative approach, testing the clusters’ sensitivity to five well-established psycholinguistic measures. These measures are often used in human language studies and relate to semantic features:
- Valence: This refers to the pleasantness or emotional tone of a word. The study found that 27 clusters were sensitive to this attribute, meaning words with similar emotional tones tended to group together.
- Concreteness: This measures how much a word refers to a perceptible entity (e.g., “bicycle” is concrete, “justice” is abstract). A substantial 60 clusters showed sensitivity to concreteness, indicating that the model differentiates between concrete and abstract terms at the static level.
- Iconicity: This explores the perceived resemblance between a word’s form (sound or appearance) and its meaning. While intriguing, only 9 clusters were sensitive to iconicity, and the researchers expressed some skepticism about these findings, suggesting potential correlations with surface features like syllable count.
- Taboo: This measures the extent to which a word is considered a “swear” or “curse” word. Despite a relatively small dataset for this measure, 6 clusters showed sensitivity, including groups related to medical terms (e.g., “cancer”) and negative events, which can carry high taboo values.
- Age of Acquisition (AoA): This is the age at which a word is typically learned. While not directly a semantic feature, AoA correlates with some semantic properties like imageability. 36 clusters were found to be sensitive to AoA.
Overall, the results from both studies strongly suggest that the static embeddings within transformer models are rich with a wide array of semantic information. This finding challenges the “meaning eliminativist” view, which posits that LLMs might not need to store invariant semantic information for individual words, instead generating meaning entirely from context. The research indicates that LLMs do, in fact, rely on a form of lexical store containing semantic data as part of their process for understanding text.
This research provides valuable insights into the internal workings of large language models, suggesting that their “understanding” of words goes beyond mere statistical patterns and includes a foundational layer of stored meaning. For more details, you can read the full research paper available here.
