The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities
Zhaofeng Wu, Xinyan Velocity Yu, Dani Yogatama, Jiasen Lu, Yoon Kim
2024-11-11
Summary
This paper discusses the Semantic Hub Hypothesis, which suggests that language models can understand and connect meanings across different languages and types of data, like text and images.
What's the problem?
Language models today can process various languages and data formats, but it's unclear how they manage to do this effectively. Often, these models might learn to associate meanings in a way that isn't straightforward, making it hard to understand their inner workings and how they relate different types of input.
What's the solution?
The authors propose the Semantic Hub Hypothesis, which states that language models create a shared space for understanding meanings across different languages and modalities (like text and images). They demonstrate that when the model processes similar concepts in different languages, it represents them in similar ways within its internal structure. This means that the model can leverage its understanding from one type of data to improve its performance on another type. The study includes experiments showing that interventions in one type of data can affect the model's output in another, confirming the existence of this shared representation space.
Why it matters?
This research is significant because it helps us understand how language models work more deeply. By identifying that these models share a central understanding of meaning, we can develop better AI systems that are more flexible and capable of handling multiple languages and data formats. This could lead to advancements in translation, image recognition, and other applications where understanding across different modalities is crucial.
Abstract
Modern language models can process inputs across diverse languages and modalities. We hypothesize that models acquire this capability through learning a shared representation space across heterogeneous data types (e.g., different languages and modalities), which places semantically similar inputs near one another, even if they are from different modalities/languages. We term this the semantic hub hypothesis, following the hub-and-spoke model from neuroscience (Patterson et al., 2007) which posits that semantic knowledge in the human brain is organized through a transmodal semantic "hub" which integrates information from various modality-specific "spokes" regions. We first show that model representations for semantically equivalent inputs in different languages are similar in the intermediate layers, and that this space can be interpreted using the model's dominant pretraining language via the logit lens. This tendency extends to other data types, including arithmetic expressions, code, and visual/audio inputs. Interventions in the shared representation space in one data type also predictably affect model outputs in other data types, suggesting that this shared representations space is not simply a vestigial byproduct of large-scale training on broad data, but something that is actively utilized by the model during input processing.