Nettsider med emneord «Machine learning»
The use of lexical semantic information for the task of syntactic parsing has seen varied success. Recently, however, the use of lexical semantic clusters derived from large corpora has been shown to improve parsing performance. It is still unclear, however, how different properties of these clusters affect results. This project aims to investigate the use of different types of clusters during syntactic parsing.
More precisely the idea is to use word clusters as a source for features in a statistical disambiguation model for a dependency parser. Generally, the clusters will group together words with similar distributional properties. The exact nature of these similarity relations, however, will vary depending on the types of context features that are used when performing the clustering. For this project we will basically be doing an extrinsic form of cluster evaluation then; investigating how different clustering parameters in turn affect the performance of a statistical parser.
Classification models are often used to make decisions that affect humans: whether to approve a loan application, extend a job offer, or provide insurance. In such applications, it is desirable for the individuals to not only knowing the results but also have the ability to change the decisions of the model. For example, when a person is denied a loan by a credit scoring model, in addition to know why he/she can not received the loan, it is meaningful for the person to know what he/she can do to influence the decision, i.e. what are the input variables that, if values are changed, can alter the decision of the model. Otherwise, without this information, he/she will be denied the loan as long as the model is deployed, and – more importantly – will lack agency over a decision that affects their livelihood.
Ontologies are developed for representing requirements and specifications that can help domain experts and engineers do various tasks in the real world. Usually, as the ontologies are quite large, it takes a long time for human users to understand the ontologies and it is also quite difficult for machines to do reasoning on such large ontologies. The goal is that we can explore different machine-learning techniques to extract relevant parts of the ontology and reduce reasoning time.
In joint work, SANT and NLPL/EOSC-Nordic have trained and released the first large-scale transformer-based language model for Norwegian: NorBERT!
An important priority for LTG in recent years has been to create NLP resources for the Norwegian language, both in terms of modeling and datasets. This page provides an overview of our existing and ongoing projects to support Norwegian NLP.
Two new papers from SANT were presented at the NoDaLiDa conference.
In this ongoing cross-disciplinary collaboration, researchers in Language Technology (LT) and Political Science (PS) are applying supervised and unsupervised machine learning methods to data from the Norwegian parliament in order to gather knowledge spanning across different dimensions.