2019/05/21 by Luisa März, März, Luisa, Dietrich Trautmann +3 · 1 citation
Computer Science · #Natural Language Processing Techniques #Topic Modeling #Multimodal Machine Learning Applications
paper · pdf · doi:10.48550/arxiv.1905.08920
The performance of a Part-of-speech (POS) tagger is highly dependent on the\ndomain ofthe processed text, and for many domains there is no or only very\nlittle training data available. This work addresses the problem of POS tagging\nnoisy user-generated text using a neural network. We propose an architecture\nthat trains an out-of-domain model on a large newswire corpus, and transfers\nthose weights by using them as a prior for a model trained on the target domain\n(a data-set of German Tweets) for which there is very little an-notations\navailable. The neural network has two standard bidirectional LSTMs at its core.\nHowever, we find it crucial to also encode a set of task-specific features, and\nto obtain reliable (source-domain and target-domain) word representations.\nExperiments with different regularization techniques such as early stopping,\ndropout and fine-tuning the domain adaptation prior weights are conducted. Our\nbest model uses external weights from the out-of-domain model, as well as\nfeature embeddings, pre-trained word and sub-word embeddings and achieves a\ntagging accuracy of slightly over 90%, improving on the previous state of the\nart for this task.\n