Hi Milan,
concerning the fist question, I'm using only three dimension to simplify the annotation process. This space could have more dimensions, offering a more rich description at emotional level.
Concerning the second question, in the examples the emotional values were shown at the token (word) level. However, this is a simplified representation of a more complex process. While individual tokens have their own emotional embeddings, these are not used in isolation. The model integrates these token-level embeddings with their context. This integration happens through the attention mechanism, which considers the relationships between all tokens in a sequence.
The overall emotional evaluation of a sentence arises from the interaction of its individual tokens through the attention mechanism. This enables the model to capture subtle emotional variations that result from the combining of words, which may deviate from a simple aggregation of individual word emotions. The λ parameter in our attention mechanism allows the model to adaptively weight the importance of emotional information relative to semantic content.
A key challenge in improving model alignment is enabling these models to understand and respond properly to the emotional and contextual aspects of human communication. This post presents an emotion-informed valuation method for LLMs that aims to improve alignment by increasing their ability to process and respond to emotional context.
Standard LLMs, while proficient in processing semantic content, often struggle with:
These limitations can lead to responses that, while semantically correct, may be emotionally tone-deaf or contextually inappropriate, potentially causing miscommunication or even harm in sensitive situations.
Drawing inspiration from V.J. Wukmir's psychological... (read 1662 more words →)
While semantic and emotional information flows start in parallel, they are not fully parallel throughout the entire process. They update each other iteratively, enabling it to capture intricate connections between semantic content and emotional tone. This has the potential to enhance the model's comprehension of the input text, resulting in a more refined understanding.