Encoding Dependency Information inside Tree Transformer

Abstract

Representing a sentence in a high dimensional space is fundamental for most natural language processing (NLP) tasks at present. These representations depend on the underlying structures upon which they are built. Two scenarios are possible: one is to view the sentence as a sequence of words and another is to consider its inherent grammatical structure. It is possible to equip the first way with some external grammatical knowledge, but to capture a proper syntax would be close to impossible. Therefore, we investigate the second one by extending the design of an existing dependency tree transformer (DT-Transformer). We propose adding a novel edge encoding mechanism to this prior architecture. Experiments show that in sentence encoding, having access to information about the relationships between “head” words and their “dependent” words and how the heads are influenced by the dependent words achieves better sentence representation. Evaluation on the four tasks shows noteworthy results compared to the existing DT-Transformer, standard Transformer, LSTM-based models, and tree- structured LSTMs. Extensive experimentation with representing the edge embeddings as different distributions (mean and standard deviation), encoding the edges in different ways, and an ablation study to find where to place each module in the architecture and which modules to use in the design is also provided.