- (Poon et al., 2010) ⇒ Hoifung Poon, Janara Christensen, Pedro Domingos, Oren Etzioni, Raphael Hoffmann, Chloe Kiddon, Thomas Lin, Xiao Ling, Mausam, Alan Ritter, Stefan Schoenmackers, Stephen Soderland, DanWeld, FeiWu, Congle Zhang. (2010). “Machine Reading at the University of Washington.” In: Proceedings of the 24th Conference on Artificial Intelligence (AAAI 2010).
Subjectt Headings: Machine Reading
- Machine reading is a long-standing goal of AI and NLP. In recent years, tremendous progress has been made in developing machine learning approaches for many of its subtasks such as parsing, information extraction, and question answering. However, existing end-to-end solutions typically require substantial amount of human efforts (e.g., labeled data and/or manual engineering), and are not well poised for Web-scale knowledge acquisition. In this paper, we propose a unifying approach for machine reading by bootstrapping from the easiest extractable knowledge and conquering the long tail via a self-supervised learning process. This self-supervision is powered by joint inference based on Markov logic, and is made scalable by leveraging hierarchical structures and coarse-to-fine inference. Researchers at the University of Washington have taken the first steps in this direction. Our existing work explores the wide spectrum of this vision and shows its promise.
- Machine reading, or learning by reading, aims to extract knowledge automatically from unstructured text and apply the extracted knowledge to end tasks such as decision making and question answering. It has been a major goal of AI and NLP since their early days. With the advent of the Web, the billions of online text documents contain virtually unlimited amount of knowledge to extract, further increasing the importance and urgency of machine reading.
- In the past, there has been a lot of progress in automating many subtasks of machine reading by machine learning approaches (e.g., components in the traditional NLP pipeline such as POS tagging and syntactic parsing). However, end-to-end solutions are still rare, and existing systems typically require substantial amount of human effort in manual engineering and/or labeling examples. As a result, they often target restricted domains and only extract limited types of knowledge (e.g., a pre-specified relation). Moreover, many machine reading systems train their knowledge extractors once and do not leverage further learning opportunities such as additional text and interaction with end users.
- Ideally, a machine reading system should strive to satisfy the following desiderata:
- End-to-end: the system should input raw text, extract knowledge, and be able to answer questions and support other end tasks;
- High quality: the system should extract knowledge with high accuracy;
- Large-scale: the system should acquire knowledge at Web-scale and be open to arbitrary domains, genres, and languages;
- Maximally autonomous: the system should incur minimal human effort;
- Continuous learning from experience: the system should constantly integrate new information sources (e.g., new text documents) and learn from user questions and feedback (e.g., via performing end tasks) to continuously improve its performance.
- These desiderata raise many intriguing and challenging research questions. Machine reading research at the University of Washington has explored a wide spectrum of solutions to these challenges and has produced a large number of initial systems that demonstrated promising performance. During this expedition, an underlying unifying vision starts to emerge. It becomes apparent that the key to solving machine reading is to:
- 1. Conquer the long tail of textual knowledge via a self-supervised learning process that leverages data redundancy to bootstrap from the head and propagates information down the long tail by joint inference;
- 2. Scale this process to billions ofWeb documents by identifying and leveraging ubiquitous structures that lead to sparsity.
- In Section 2, we present this vision in detail, identify the major dimensions these initial systems have explored, and propose a unifying approach that satisfies all five desiderata. In Section 3, we reivew machine reading research at the University of Washington and show how they form synergistic effort towards solving the machine reading problem. We conclude in Section 4.
3.5 Ontology Induction
- As mentioned in previous subsections, ontologies play an important role in both self-supervision (shrinkage) and large-scale inference (coarse-to-fine inference). A distinctive feature in our unifying approach is to induce probabilistic ontologies, which can be learned from noisy text and support joint inference. Past systems have explored two different approaches to probabilistic ontology induction. One approach is to bootstrap from existing ontological structures and apply self-supervision to correct the erroneous nodes and fill in the missing ones (KOG). Another approach is to integrate ontology induction with hierarchical smoothing, and jointly pursue unsupervised ontology induction, population and knowledge extraction (LOFT).
Facts about "2010 MachineReading"
|Author||Hoifung Poon +, Janara Christensen +, Pedro Domingos +, Oren Etzioni +, Raphael Hoffmann +, Chloe Kiddon +, Thomas Lin +, Xiao Ling +, Mausam +, Alan Ritter +, Stefan Schoenmackers +, Stephen Soderland +, DanWeld +, FeiWu + and Congle Zhang +|
|journal||Proceedings of the 24th Conference on Artificial Intelligence +|
|title||Machine Reading at the University of Washington +|