Computer Science Graduate Seminar: Unsupervised Training with Applications in Natural Language Processing
Friday, July 12, 2019, 10:00am
Location: Computer Science Center, E3, Room 9222
Speaker: Dipl.-Inform. Malte Nuhn
The state-of-the-art algorithms for various natural language processing tasks require large amounts of labeled training data. At the same time, obtaining labeled data of high quality is often the most costly step in setting up natural language processing systems. Opposed to this, unlabeled data is much cheaper to obtain and available in larger amounts. Currently, only few training algorithms make use of unlabeled data. In practice, training with only unlabeled data is not performed at all. In this thesis, we study how unlabeled data can be used to train a variety of models used in natural language processing. In particular, we study models applicable to solving substitution ciphers, spelling correction, and machine translation. This thesis lays the groundwork for unsupervised training by presenting and analyzing the corresponding models and unsupervised training problems in a consistent manner. We show that the unsupervised training problem that occurs when breaking one-to-one substitution ciphers is equivalent to the quadratic assignment problem (QAP) if a bigram language model is incorporated and therefore NP-hard. Based on this analysis, we present an effective algorithm for unsupervised training for deterministic substitutions. In the case of English one-to-one substitution ciphers, we show that our novel algorithm achieves results close to human performance, as presented in [Shannon 49].
Also, with this algorithm, we present, to the best of our knowledge, the first automatic decipherment of the second part of the Beale ciphers. Further, for the task of spelling correction, we work out the details of the EM algorithm [Dempster & Laird + 77] and experimentally show that the error rates achieved using purely unsupervised training reach those of supervised training. For handling large vocabularies, we introduce a novel model initialization as well as multiple training procedures that significantly speed up training without hurting the performance of the resulting models significantly. By incorporating an alignment model, we further extend this model such that it can be applied to the task of machine translation. We show that the true lexical and alignment model parameters can be learned without any labeled data: We experimentally show that the corresponding likelihood function attains its maximum for the true model parameters if a sufficient amount of unlabeled data is available. Further, for the problem of spelling correction with symbol substitutions and local swaps, we also show experimentally that the performance achieved with purely unsupervised EM training reaches that of supervised training. Finally, using the methods developed in this thesis, we present results on an unsupervised training task for machine translation with a ten times larger vocabulary than that of tasks investigated in previous work.
The computer science lecturers invite interested people to join.