Semi-Supervised Learning of Sequence Models with the Method of Moments
In this talk I will present work presented at EMNLP 2016, about a fast and scalable method for semi-supervised learning of sequence models.
The proposed method is based on anchor words and moment matching techniques to retrieve the hidden assignment in a Hidden Markov structure. We can handle feature-based observations. Unlike other semi-supervised methods, we propose a more efficient approach where no additional decoding passes are necessary on the unlabeled data and no graph needs to be constructed—only one pass is necessary to collect moment statistics.
We demonstrate the effectiveness of this approach on Twitter part-of-speech tagging experiments and show that our method can learn from very few annotated sentences.