Parsing low-resource languages and domains

Obiettivo

There are noticeable asymmetries in availability of high-quality natural language processing (NLP). We can adequately summarize English newspapers and translate them into Korean, but we cannot translate Korean newspaper articles into English, and summarizing micro-blogs is much more difficult than summarizing newspaper articles. This is a fundamental problem for modern societies, their development and democracy, as well as perhaps the most important research problem in NLP right now.

Most NLP technologies rely on highly accurate syntactic parsing. Reliable parsing models can be induced from large collections of manually annotated data, but such collections are typically limited to sampled newswire in major languages. Highly accurate parsing is therefore not available for other languages and other domains.

The NLP community is well aware of this problem, but unsupervised techniques that do not rely on manually annotated data cannot be used for real-world applications, where highly accurate parsing is needed, and sample bias correction methods that automatically correct the bias in newswire when parsing, say, micro-blogs, do not yet lead to robust improvements across the board.

The objective of this project is to develop new learning methods for parsing natural language for which no unbiased labeled data exists. In order to do so, we need to fundamentally rethink the unsupervised parsing problem, including how we evaluate unsupervised parsers, but we also need to supplement unsupervised learning techniques with robust methods for automatically correcting sample selection biases in related data. Such methods will be applicable to both cross-domain and cross-language syntactic parsing and will pave the way toward robust and scalable NLP. The societal impact of robust and scalable NLP is unforeseeable and comparable to how efficient information retrieval techniques have revolutionized modern societies.

Campo scientifico

Meccanismo di finanziamento

ERC-SG - ERC Starting Grant

Istituzione ospitante

KOBENHAVNS UNIVERSITET

Contributo UE

€ 1 126 183,20

Indirizzo

NORREGADE 10
1165 Kobenhavn
Danimarca

Regione

Danmark Hovedstaden Byen København

Tipo di attività

Higher or Secondary Education Establishments

Ricercatore principale

Anders Søgaard (Dr.)

Contatto amministrativo

Tine Matthiesen (Mrs.)

Collegamenti

Contatta l’organizzazione Sito web

Costo totale

Nessun dato

Beneficiari (1)