摘要：Composing the representation of a sentence from the tokens that it comprises is difficult, because such a representation needs to account for how the words present relate to each other. The Transformer architecture does this by iteratively changing token representations with respect to one another. This has the drawback of requiring computation that grows quadratically with respect to the number of tokens. Furthermore, the scalar attention mechanism used by Transformers requires multiple sets of parameters to operate over different features. The present paper proposes a lighter algorithm for sentence representation with complexity linear in sequence length. This algorithm begins with a presumably erroneous value of a context vector and adjusts this value with respect to the tokens at hand. In order to achieve this, representations of words are built combining their symbolic embedding with a positional encoding into single vectors. The algorithm then iteratively weighs and aggregates these vectors using a second-order attention mechanism, which allows different feature pairs to interact with each other separately. Our models report strong results in several well-known text classification tasks.
