Knowledge Graphs (KGs) store human knowledge about the world in structured format, e.g., triples of facts or graphs of entities and relations, to be processed by AI systems. In the past decade, extensive research efforts have gone into constructing and utilizing knowledge graphs for tasks in natural language processing, information retrieval, recommender systems, and more. Once constructed, knowledge graphs are often considered as “gold standard” data sources that safeguard the correctness of other systems. Because the biases inherent to KGs may become magnified and spread through such systems, it is crucial that we acknowledge and address various types of bias in knowledge graph construction.
Biases may originate in the very design of the KG, in the source data from which it is created (semi-)automatically, and in the algorithms used to sample, aggregate, and process that data. Causes of bias include systematic errors due to selecting non-random items (selection bias), misremembering certain events (recall bias), and interpreting facts in a way that affirms individuals’ preconceptions (confirmation bias). Biases typically appear subliminally in expressions, utterances, and text in general and can carry over into downstream representations such as embeddings and knowledge graphs.
This workshop – to be held for the first time at AKBC 2020 – addresses the questions: “how do such biases originate?”, “How do we identify them?”, and “What is the appropriate way to handle them, if at all?”. This topic is as-yet unexplored and the goal of our workshop is to start a meaningful, long-lasting dialogue spanning researchers across a wide variety of backgrounds and communities.
KG-BIAS topics of interest include, but are not limited to the following.
Detecting and mitigating gender bias in NLP
Bias in Data and Algorithmic Systems: A "Fish-Eye View" of Problems, Solutions and Stakeholders.
KG-BIAS will be held virtually on June 25, 2020. All times are in Pacific Time.
Mitigating bias in algorithmic processes and systems is a critical issue drawing increasing attention across research communities within the information and computer sciences. Given the complexity of the problem and the involvement of multiple stakeholders – not only developers, but also end-users and third parties – there is a need to understand the landscape of the sources of bias, as well as the solutions being proposed to address them. In this talk, I present insights from a recent survey of 250+ articles across four domains (machine learning, information retrieval, HCI, and RecSys), providing a “fish-eye view” of the field. I will also discuss the particular challenges of data biases, given the heterogeneity of data sources available for learning about our world, drawing on examples of my previous work on image data.
Natural Language Processing (NLP) plays an important role in many applications, including resume filtering, text analysis, and information retrieval. Despite the remarkable accuracy enabled by the advances in machine learning used in many applications, the technique may discover and generalize the societal biases implicit in the data. For example, an automatic resume filtering system may unconsciously select candidates based on their gender and race due to implicit associations between applicant names and job titles, causing the societal disparity discovered by researchers. Various laws and policies have been designed to ensure social equality and diversity. However, there is no such mechanism for a machine learning model for sensitive applications. My research analyzes the potential stereotypes in various machine learning models and develops computational approaches to enhance fairness in a wide range of NLP applications. The broader impact of my research aligns with one the concerns of machine learning community: how can we do AI for (social) good.
Our workshop adheres to all principles and guidelines specified in the ACM Code of Ethics and Professional Conduct.
