Kenneth Church (Northeastern University, USA)

Title: Better Together: Text + Context

We will demo an academic search engine for papers and authors in Semantic Scholar (S2). The UX makes it easy to compare different embeddings and recommendation APIs. Users can choose which embedding/API to use for generating search results. To encourage users to compare and contrast embeddings (vectors for many/most documents in S2), the search result page (SRP) reports similarity scores for each result and four embeddings: Specter, ProNE-s, SciNCL and a GNN (graph neural network). Most of these embeddings are based on deep neural networks (BERT encodings of abstracts, with fine-tuning on nearby documents in the citation graph), but ProNE-s is based on linear algebra (spectral clustering of citation graph). Some embeddings/APIs are better at capturing text (titles and abstract) and others are better at capturing links in the citation graph. Methods that depend too much on English text do not work well on common corner cases such as missing abstracts, duplicate documents and languages other than English. ProNE-s is less exposed to these corner cases because it is based on the citation graph. Experiments will show that ProNE-s finds papers with more citations, and the S2 API finds more recent papers. It is unlikely that one method is better than the others for all use cases. We believe it is useful to encourage users to explore a number of promising options, and vote with their clicks.

Claudia Hauff (Spotify)

Title: On the Challenges of Podcast and Audiobook Search at Spotify

Online music streaming is enjoying ever-growing popularity over the past decades, enabled by the abundance of music content in digital format and online streaming services. In recent years, podcasts (a talk-focused media format) have witnessed a rapid growth among listeners. Podcasts come in many forms and sizes. They range from 20-minute daily meditation sessions, weekly recaps of global news, to interviews with celebrities, and hosts bantering with each other for hours. More and more streaming services are now expanding their catalogs to support both music and podcasts as well as audiobooks on the same platform. This setup requires an effective aggregated search system to assemble information from heterogeneous information sources and content types into one result interface in order to support diverse information needs. In this talk I will go over some of the challenges we encountered.

Ricardo Baeza-Yates (Institute for Experiential AI, USA)

Title: The Limitations of Machine Learning & Us

Lightining Talk followed by a panel!

Machine learning (ML), particularly deep learning, is being used everywhere. However, not always is used well, ethically and scientifically. In this talk we first do a deep dive in the limitations of supervised ML and data, its key component. We cover small data, datification, bias, predictive optimization issues, evaluating success instead of harm, and pseudoscience, among other problems.  The second part is about our own limitations using ML, including different types of human incompetence: cognitive biases, unethical applications, no administrative competence, copyright violations, misinformation, and the impact on mental health. In the final part we discuss regulation on the use of AI and responsible AI principles, that can mitigate the problems outlined above.