90 likes | 151 Views
ML for NLP With Special Focus on Tagging and Parsing. Kiril Ribarov. Lecture structure. Machine learning in general
E N D
ML for NLPWith Special Focus on Tagging and Parsing Kiril Ribarov Dipartimento di Informatica, Universita di Pisa
Lecture structure • Machine learning in general T.M. Mitchell, Machine Learning (1997, McGraw-Hill):hypothesis, decision trees, ANN, computational learning theory, instance-based learning, genetic algorithms, (Bayesian learning), [case-based, analytical learning] • Natural Language – linguistically • Natural Language – computationally and stochastically
Lecture structure – cont. • The Prague Dependency Treebank – data and tools • Morphological Tagging • Bayesian Learning (HMM, smoothing, EM, Maximum Entropy, related issues as Viterbi search, Lagrange multipliers) • Rule-Based Approach • Perceptron-Based Approach • Tagset structure, tagset size • Morphological contexts
Lecture structure – cont. • Parsing • The problem of parsing, dependency parsing • Statistical parsing • Rule-Based parsing • Language graphs and sentence graphs • Naïve parsing • Rule-Based revisited • Perceptron-Based parsing
Lecture structure – cont. • Parsing by tagging and tagging by parsing • Morphological contexts, tree contexts • G-tags, tagging by g-tags • Alignment of g-tags and m-tags • Some problem definitions
We will include as well • Problems of evaluation and its measurement for tagging and for parsing • Specialties of dependency trees, surface and deep syntax, projectivity and non-projectivity • Current trials on high-quality MT • Ongoing research on valencies
Our aim • To present general ML techniques • To present the Prague Dependency Treebank • To present NLP specific approaches, their modifications, applications (medium: PDT) • To present mistakes and successes • To present the newest ideas developed for automatic dependency acquisition • To raise questions and thus indicate new directions for research in NLP
(Restriction) to Tagging and Parsing • Tagging and parsing as the two most important NLP modules for various application domains. • Tagging and parsing undoubtedly improve: grammar checking, speech processing, information retrieval, machine translation, … • Each of the applications does not necessarily use the same tagging/parsing outputs; modifications are introduced to serve best the specific application • Each of the applications has its specific core modules different than tagging/parsing • Many technicalities in these are approaches are nevertheless similar
Machine Learning in General T.M. Mitchell, Machine Learning (1997, McGraw-Hill): hypothesis, decision trees, ANN, computational learning theory, instance-based learning, genetic algorithms