# How Sequence-to-Sequence Models Perceive Language Styles?

## TL;DR

This paper introduces a novel perspective on how sequence-to-sequence models perceive language style by analyzing style matrices derived from semantic vectors, leading to a flexible, learning-free style transfer method that performs competitively.

## Contribution

It presents the first explanation of style encoding in seq2seq models using style matrices and proposes a new, learning-free style transfer algorithm based on these matrices.

## Key findings

- Style matrices effectively capture language style information.
- The proposed style transfer algorithm outperforms some existing methods.
- The method transfers style even for out-of-domain sentences.

## Abstract

Style is ubiquitous in our daily language uses, while what is language style to learning machines? In this paper, by exploiting the second-order statistics of semantic vectors of different corpora, we present a novel perspective on this question via style matrix, i.e. the covariance matrix of semantic vectors, and explain for the first time how Sequence-to-Sequence models encode style information innately in its semantic vectors. As an application, we devise a learning-free text style transfer algorithm, which explicitly constructs a pair of transfer operators from the style matrices for style transfer. Moreover, our algorithm is also observed to be flexible enough to transfer out-of-domain sentences. Extensive experimental evidence justifies the informativeness of style matrix and the competitive performance of our proposed style transfer algorithm with the state-of-the-art methods.