Natural Language Processing
To provide the students the techniques and tools to devise and devel-op Natural Language Processing (NLP) components and applications. The course will cover the foundations, building blocks and applications of NLP, with an emphasis on the necessary linguistic intuitions as well as a broad coverage of statistical and deep learning models that can be used for language tasks. NLP is an important topic in Artificial Intelli-gence with a wide range of applications, from sentiment analysis to machine translation. Modern NLP is primarily based on statistical meth-ods and machine learning algorithms, where linguistic information is provided by instances of uses of language. For most NLP tasks, state of the art approaches are based on neural models, which will be at the core of this module. However, significant attention will be given to the linguistic principles that underpin the field.
More specifically, students will:
- Gain familiarity with important linguistic concepts involved in language understanding and generation, from morphological analysis to pragmatics
- Gain familiarity with, devise, implement and apply relevant pre-processing steps for natural language processing components and applications
- Critically compare statistical and deep learning approaches for natural language processing
- Map various well established techniques in machine learning to specific problems in natural language processing
- Build, evaluate, critically analyze and improve models using ex-isting machine learning algorithms and frameworks (such as TensorFlow) for a range of natural language processing tasks, including: classification, structured prediction, sequence to se-quence labeling and generation
- Devise, implement and evaluate classifiers for a range of natu-ral language processing tasks.
After the course, students should be able to:
(ILO1) Identify and automatically pre-process texts that can be useful for language processing tasks
(ILO2) Devise and evaluate solutions for a range of natural lan-guage components using existing algorithms, techniques and frameworks, including part-of-speech tagging, language mod-eling, parsing and semantic role labeling
(ILO3) Devise, implement and evaluate algorithms for single and multi-class classification problems
(ILO4) Apply existing statistical and deep learning techniques to language applications such as machine translation.
ILO1 and IL3 will be assessed mainly through the coursework, while ILO2 and ILO4 will be assessed mainly via exam.
1 Introduction to NLP (language challenges, applications, clas-sical vs statistical vs deep learning-based)
2 Basic concepts in Linguistics (including morphology, syntax, semantics, pragmatics)
3 Pre-processing techniques, word meaning (TF-IDF, distribu-tional models, word2vec, glove, etc)
4 Lab/tutorial session on pre-processing and word meaning
5-6 Classification tasks with simple classification models (Naïve Bayes, perceptron): SPAM detection, part-of-speech tagging, word sense disambiguation
7 Classification tasks with CNN models
8 Lab/tutorial session on classification
9 Coursework specification and discussion
10 N-gram language models
11 Neural language models (RNNs, LSTMs, GRUs)
12 Lab/tutorial session on language models
13 Structured prediction - POS tagging with HMM
14 Structured prediction - POS tagging with neural models (RNN)
15 Syntax and parsing
16 Lab/tutorial session on POS tagging
17 Rules-based and probabilistic parsing
18 Neural models for parsing
19 Semantic role labeling
20 Lab/tutorial session on parsing
21-23 Sequence to sequence modelling - machine translation (SMT, NMT, attention)
24 Lab/tutorial session on sequence to sequence modelling
25 Guest lecture on advanced NLP topics
26 Guest lecture on advanced NLP topics
27 Revision lecture
28 Revision lecture
Recommended: COMP60012/COMP70050 Introduction to Machine Learning; COMP70053 Python Programming.
6 weeks of 3 hours of lectures + 1 hour of lab/tutorial (24 hours)
1 invited lecture (2 hours)
End of course wrap up and revision lecture (2 hours)
Week 3: coursework specification
The task will be to develop models to identify and categorise lan-guage into a set of classes. A task from an open competition (SemEval or Kaggle) will be selected, such that students can sub-mit their results directly to the task website and get quantitative feedback (classifier performance) and compare it against other par-ticipants. Students will be able to apply any pre-processing tech-nique, any additional data they find, any algorithm (i.e. CNN clas-sifiers). The course work will be marked on the creativity and breadth of the proposed solution (e.g. comparisons, analyses, etc.), rather than on the performance obtained on the shared task.
Report and link to github code repository will be submitted via CATE. Submissions will subsequently be marked by TAs for creativity of solu-tion (ILOs 1 and 3), code clarity/documentation and written report. The feedback will be given to students before the exam. Informal feedback on the coursework and all other course topics will be given during the lab/tutorial sessions.