The Double Helix inside the NLP Transformer

TL;DR

This paper presents a framework for analyzing information types in NLP Transformers, revealing a helix-shaped positional information pattern and how different layers encode syntactic and semantic features.

Contribution

It introduces a novel analysis method distinguishing four information layers and proposes a Linear-and-Add approach for positional embedding, uncovering the helix pattern of positional data.

Findings

Positional information forms a helix in deep layers.

Encoder layers generate Part-of-Speech clusters.

Decoder layers reveal PoS clusters via bigram analysis.

Abstract

We introduce a framework for analyzing various types of information in an NLP Transformer. In this approach, we distinguish four layers of information: positional, syntactic, semantic, and contextual. We also argue that the common practice of adding positional information to semantic embedding is sub-optimal and propose instead a Linear-and-Add approach. Our analysis reveals an autogenetic separation of positional information through the deep layers. We show that the distilled positional components of the embedding vectors follow the path of a helix, both on the encoder side and on the decoder side. We additionally show that on the encoder side, the conceptual dimensions generate Part-of-Speech (PoS) clusters. On the decoder side, we show that a di-gram approach helps to reveal the PoS clusters of the next token. Our approach paves a way to elucidate the processing of information through the deep layers of an NLP Transformer.