Ruprecht-Karls-Universität Heidelberg
Institut für Computerlinguistik

Bilder vom Neuenheimer Feld, Heidelberg und der Universität Heidelberg

Bias

Module Description

Course Module Abbreviation Credit Points
BA-2010[100|75] CS-CL 6 LP
BA-2010[50] BS-CL 6 LP
BA-2010[25] BS-AC, BS-FL 4 LP
BA-2010 AS-CL 8 LP
Master SS-CL, SS-TAC 8 LP
Lecturer Katja Markert
Module Type Proseminar / Hauptseminar
Language English
First Session 29.10.2019 (Note unusual start date)
Time and Place Tuesday, 14:15-15:45
INF 326 / SR 28
End of Commitment Period 21.01.2020

Prerequisite for Participation

For MA Students: none.

For BA Students: Statistical Methods for Computational Linguistics. Also sufficient are "Neural Networks: Architectures and Applications for NLP" or (for PS participants) the lecture "Embeddings"

Assessment

(i) Active Participation (leading discussion, exercises, answering/putting questions)

(ii) Presentation

(iii) Second presentation, term paper or implementation project

Content

Machine Learning algorithms can induce potentially ethically harmful bias in their outputs. In particular, they may violate the "equality of odds" fairness criterion (Hardt et al, NIPS 2016), which states that any ML decision should not be influenced by demographic attributes such as gender, race etc.A biased system is here defined as one that will perform better for some demographics groups than others. Distinct from this bias definition is the definition of a biased model as one that learns stereotypical correlation of concepts (such as often the case in word embeddings that might learn that "programmer" is more similar to "man" than to "woman").

This seminar will investigate NLP work on both cases of bias and consider the following topics:

--- Selection Bias and Bias in frequently used corpora (such as Wikipedia)

--- Measuring, testing and visualising bias

--- Mitigation of Bias in Machine Learning via methods as diverse as changing underlying training data, adversarial training, modifying classification algorithms at the prediction level, debiasing word embeddings, use of knowledge graphs etc.

--- Application scenarios in NLP (such as text classification, coreference resolution) and potentially also adjacent fields (such as image search and collaborative filtering).

This seminar is *not* an opinion mining or sentiment seminar. We will not adress the finding of political opinions such as discovery of left/right-wing newspapers etc. We will concentrate on algorithmic bias in ML.

Module Overview

Agenda

Date Session Materials

Literature

Will be given out at the beginning of term and is normally freely available.

» More Materials

zum Seitenanfang