Brian Johnson (Apple, USA)

Title: Apple Media Products Search

16 September 2021

Abstract

The Apple Media Products Search Team powers search for App Store, iTunes, Apple Music, Apple TV, Books, and Podcasts across multiple platforms (iOS, iPadOS, macOS, watchOS, tvOS, Web, Windows, Android) . App Store Search is used by millions of users daily serving over 1 billion requests a week across 150 countries and 40 languages. This scale and diversity offers incredible opportunities to improve search and solve challenging problems in information retrieval, ranking , query understanding, machine learning at scale, high performance computing and large scale systems design.

This talk covers three main topics.

Query Understanding

Brian will discuss what is necessary to have a deep understanding of our developers, their inventory, our users, and their intents and interests. Understanding user’s query intent is first step in Search. We need robust and scalable query understanding systems to correctly detect the intent of a user's query. Traditional Natural Language Processing techniques are important in query understanding, as are behavioral data mining, and newer deep learning techniques such as query embedding. Statistical query refinements allow us to learn synonyms, aliases, and nicknames.

Ranking beyond text

With this rich understanding of our users, queries, and inventory , we are working to build next generation search ranking for Apple Media Products. Some of the challenges we face include language processing, shifting inventory, and fraud prevention. When searching the App Store, users are presented with content that matches their query as well as content that matches the intent of their query.

Evaluation

How do we know how good we are? A robust evaluation platform helps us answer this question using well defined IR metrics for precision and recall. The ideas teams develop go through rigorous offline and online evaluations to predict how changes will perform in the wild. This allows for an agile development process. Ongoing loss pattern analysis drives continuous improvement.

Fabian M. Suchanek (Télécom Paris University, Paris, France)

Title: A hitchhiker’s guide to Ontology

18 September 2021 (in presence)

Abstract

Knowledge bases are an important asset in many of today's AI-based applications, including personal assistants and search engines. In this talk, I will give an overview of our recent work in this area.

I will first talk about our main project, the YAGO knowledge base. In this context, I will present different methods of information extraction, in particular also the extraction of beliefs and causal relationships. I will then talk about the incompleteness of knowledge bases. We have developed several techniques to estimate how much data is missing in a knowledge base, as well as rule mining methods to derive that data. I will then present our work on efficient querying of knowledge bases.

Finally, I will talk about applications of knowledge bases in the domain of computational creativity and the digital humanities, as well as about our methods for explainable AI.

Bio

Fabian M. Suchanek is a full professor at the Telecom Paris University in France. Fabian developed inter alia the YAGO knowledge base, one of the largest public general-purpose knowledge bases. This earned him a honorable mention of the SIGMOD dissertation award and the 10 year Test of Time Award of The Web Conference (WWW 2018). His interests include information extraction, automated reasoning, and knowledge bases. Fabian has published around 90 scientific articles, among others at ISWC, VLDB, SIGMOD, WWW, CIKM, ICDE, and SIGIR, and his work has been cited more than 10000 times.

Hinrich Schütze (University of Munich, Germany)

Title: Humans Learn From Task Descriptions and So Should Our Models

17 September 2021 (online)

Abstract

In many types of human learning, task descriptions are a central ingredient. They are usually accompanied by a few examples, but there is very little human learning that is based on examples only. In contrast, the typical learning setup for NLP tasks lacks task descriptions and is supervised with 100s or 1000s of examples. This is even true for so-called few-shot learning, a term often applied to scenarios with tens of thousands of "shots".

Inspired by the GPT models, which also exploit task descriptions, we introduce Pattern-Exploiting Training (PET). PET reformulates task descriptions as cloze questions that can be effectively processed by pretrained language models. In contrast to GPT, PET combines task descriptions with supervised learning. We show that PET learns well from as little as ten training examples and outperforms GPT-3 on GLUE even though it has 99.9% fewer parameters.

Bio

Hinrich Schütze is a full professor at the University of Munich in Germany. He is Chair of Computational Linguistics and Director of the Center for Information and Language Processing. He is known for co-authoring the standard reference book on statistical natural language processing (http://nlp.stanford.edu/fsnlp/). His book "Introduction to Information Retrieval" (http://nlp.stanford.edu/IR-book/information-retrieval-book.html) (co-authored with Chris Manning and Prabhakar Raghavan) was published in 2008 and has been adopted by many IR courses throughout the world. Dr. Schütze obtained his PhD from Stanford University and has worked for a number of Silican Valley companies, including two large search engines and several text mining startups.