Fixed Size Ordinally-Forgetting Encoding and its Applications

In this thesis, we propose the new Fixed-size Ordinally-Forgetting Encoding (FOFE) method, which can almost uniquely encode any variable-length sequence of words into a fixed-size representation. FOFE can model the word order in a sequence using a simple ordinally-forgetting mechanism according to the positions of words. We address two fundamental problems in natural language processing, namely, Language Modeling (LM) and Named Entity Recognition (NER).

We have applied FOFE to FeedForward Neural Network Language Models (FFNN-LMs). Experimental results have shown that without using any recurrent feedbacks, FOFE-FFNN-LMs significantly outperform not only the standard fixed-input FFNN-LMs but also some popular Recurrent Neural Network Language Models (RNN-LMs).

Instead of treating NER as a sequence labeling problem, we propose a new local detection approach, which relies on FOFE to fully encode each sentence fragment and its left/right contexts into a fixed-size representation. This local detection approach has shown many advantages over the traditional sequence labeling methods. Our method has yielded pretty strong performance in all tasks we have examined.