Language Technology Projects for Students
Note: Students starting LT projects are advised to
enrol in 433-327 Knowledge Technologies,
433-687 Knowledge Technologies,
or 433-611 Web Search and Text Analysis.
Museum Hypernavigation
Recent technological advances have afforded people any-time, any-place
access to information, while enabling providers to track people's
activities in physical environments. This project will involve
research to integrate the fields of language technology and user
modelling, and to develop a system for synchronised navigation of
hypertext and information rich physical environments. The work will
be done in collaboration with
the Kubadji Project, involving
a team of researchers from CSSE, the Department of Information
Systems, Monash University, and the Melbourne Museum.
Contact:
Tim Baldwin
Discourse-structured, co-indexed named entity recognition
Named entity recognition (NER) is the task of identifying mentions of
people, organisations, places, dates, etc in natural language
data. E.g., we might analyse the sentence:
Wolff, currently a journalist in Argentina, played with Del
Bosque in the final years of the seventies in Real Madrid.
as:
[PERSON Wolff], currently a
journalist in [LOCATION Argentina], played with
[PERSON Del Bosque] in the final years of
the seventies in [ORGANISTION Real Madrid].
Conventionally, NER is treated as a sequential tagging task, with no
regard for the logical structure of the text or discourse
cohesion. However, it is often the case that a NE reoccurs with
regularity throughout a text, with subsequent references being made by
way of an abbreviated form established explicitly at the time of the
first mention, or based on social conventions (e.g. referring to a
person by their surname). This research aims to use basic discourse
models to “track” or co-index an NE throughout a text, and use the
accumulative textual evidence to boost NER performance.
Contact:
Tim Baldwin
Computational Modelling of Prosody in Natural Speech
For many of the world's languages, voice pitch is used to signal lexical and
grammatical contrasts (cf the way we use stress in English to distinguish
CONtest from conTEST).
This project will use methods in digital signal processing and machine learning
to explore the prosodic structure of a language from Papua New Guinea,
in collaboration with researchers in the UK and USA.
(Here's an example to show how voice pitch is used to distinguish words in a PNG language
waímé tree kangaroo,
waíme house rat,
waimé needle).
Contact:
Steven Bird,
Language Engineering in the Field: Preserving 100 Endangered Languages in New Guinea
During 2010, the BOLD:PNG
project is creating an audio archive for a hundred languages
in Papua New Guinea. This project will investigate the workflows
for collecting oral literature, and develop a prototype tool to be
deployed on low-power netbooks for use in remote field locations,
to manage all aspects of the data collection process.
It will combine methods from data modelling, interface design,
and digital signal processing.
Contact:
Steven Bird,
Analyzing Talk
It is easy to create multimedia recordings and share them online using
services like YouTube. However, these sites can only be navigated by
topic. How could a collection be organised so that people could study
the content systematically? This project will involve research into
architectures for collaborative online annotation, and the development
of a system for manipulating time-aligned, interlinear audio
transcriptions. The architecture could be extended to allow for
plugins ranging from simple audio concordancers to sophisticated
knowledge discovery tools, together with illustrative applications to
the analysis of conversation, rhetoric, or dialect. Application to
languages other than English, including endangered languages, will be
possible.
Contact:
Steven Bird,
Projects with the Natural Language Toolkit
The Natural Language Toolkit (NLTK) is an open source library
containing implementations of many natural language processing
algorithms in Python (see nltk.org).
A wide range of projects are described in the
Project section of the NLTK issue tracker,
covering state-of-the-art statistical NLP algorithms, and multilingual processing.
Contact:
Steven Bird,
Greg Restall
Part-of-speech tagging and unknown words
The biggest single source of error in part-of-speech (POS) tagging is
unknown words. Traditionally, unknown word handling has been carried out on
the fly, based on affix-based similarity to known words, or analysis of the
lexical composition of the unknown word. In this research, we will explore
avenues for improving unknown word handling by way of more sophisticated
models of word similarity, word and document context. We will also
investigate methods for effectively trawling existing lexical resources to
form a gold-standard POS lexicon for use in existing POS taggers.
Contact:
Tim Baldwin
Document segmentation/genre classification
Documents are typically logically structured into text segments adhering to
a particular genre of prose. E.g., a newspaper article will typically have a
headline (written in “headlinese”, i.e. in present tense and without
articles: “Government Announces Budget Deficit”), followed by the main
prose of the article. In this project, we seek to develop a classification
of text segment genre (e.g. headline, classified, personal, spoken,
standard written, ...) along broadly syntactic and lexical lines, and
determine the logical structure of a given document according to this
classification. One area of interest is the interaction between document
segmentation/genre classification and text classification.
Contact:
Tim Baldwin
Feature-rich word sense disambiguation
Word sense disambiguation
(WSD) is the task of determining the sense of a given word in context,
based on a fixed sense inventory. Traditionally, WSD research has
focused on using only the immediate lexical context of target words to
perform the disambiguation, with moderate success. Additionally, while
sense inventories are often organised hierarchically based on synonym,
hyponym and hypernym relations, WSD methods have traditionally ignored
this ontological structure and adopted a “bag of senses” style
approach. In this research, we will make use of three novel sources of
information in carrying out supervised WSD: syntactico-semantic (rich
syntactic annotation from a precision grammar, and logical form-style
dependencies), ontological (the hierarchical data provided in the
dictionary definitions) and domain data.
Contact:
Tim Baldwin
Translation similarity toolkit
Translation memories are a powerful aid to human translators, which maintain
an incremental database of past translations and suggest partial or complete
translations for novel inputs based on this translation history. They rely
crucially on the assumption that strings which are similar in one language
will have correspondingly similar translations in a second language. This
project will implement a range of translation similarity methods and analyse
their efficacy in translation tasks involving a range of language pairs, and
also broader-ranging applications relying on some notion of similarity.
Contact:
Tim Baldwin and
Steven Bird.