2008 ExtractingSemNetsFromText

• (Kok & Domingos, 2008) ⇒ Stanley Kok, and Pedro Domingos. (2008). “Extracting Semantic Networks from Text via Relational Clustering.” In: Proceedings of the Nineteenth European Conference on Machine Learning (ECML 2008).

Subject Headings: Markov Logic Networks.

Quotes

Abstract

Extracting knowledge from text has long been a goal of AI. Initial approaches were purely logical and brittle. More recently, the availability of large quantities of text on the Web has led to the development of machine learning approaches. However, to date these have mainly extracted ground facts, as opposed to general knowledge. Other learning approaches can extract logical forms, but require supervision and do not scale. In this paper we present an unsupervised approach to extracting semantic networks from large volumes of text. We use the TextRunner system [1] to extract tuples from text, and then induce general concepts and relations from them by jointly clustering the objects and relational strings in the tuples. Our approach is defined in Markov logic using four simple rules. Experiments on a dataset of two million tuples show that it outperforms three other relational clustering approaches, and extracts meaningful semantic networks.

1 Introduction

A long-standing goal of AI is to build an autonomous agent that can read and understand text. The natural language processing (NLP) community attempted to achieve this goal in the 1970’s and 1980’s by building systems for understanding and answering questions about simple stories [3, 13, 23, 6]. These systems parsed text into a network of predefined concepts, and created a knowledge base from which inferences can be made. However, they required a large amount of manual engineering, only worked on small text sizes, and were not robust enough to perform well on unrestricted naturally occurring text. Gradually, research in this direction petered out.

Interest in the goal has been recently rekindled [16][7] by the abundance of easily accessible Web text, and by the substantial progress over the last few years in machine learning and NLP. The confluence of these three developments led to efforts to extract facts and knowledge bases from the Web [4]. Two recent steps in this direction are a system by Pasca et. al [18] and TextRunner [1]. Both systems extract facts on a large scale from Web corpora in an unsupervised manner. Pasca et. al’s system derives relation-specific extraction patterns from a starting set of seed facts, acquires candidate facts using the patterns, adds high-scoring facts to the seeds, and iterates until some convergence criterion. TextRunner uses a domain-independent approach to extract a large set of relational tuples of the form r(x, y) where [math]\displaystyle{ x }[/math] and [math]\displaystyle{ y }[/math] are strings denoting objects, and [math]\displaystyle{ r }[/math] is a string denoting a relation between the objects. It uses a lightweight noun phrase chunker to identify objects, and heuristically determines the text between objects as relations. These are good first steps, but they still fall short of the goal. While they can quickly acquire a large database of ground facts in an unsupervised manner, they are not able to learn general knowledge that is embedded in the facts.

Another line of recent research takes the opposite approach. Semantic parsing [26, 17, 29] is the task of mapping a natural language sentence into logical form. The logical statements constitute a knowledge base that can be used to perform some task like answering questions. Semantic parsing systems require a training corpus of sentences annotated with their associated logical forms (i.e., they are supervised). These systems are then trained to induce a parser that can convert novel sentences to their logical forms. Even though these systems can create knowledge bases directly, their need for annotated training data prevents them from scaling to large corpora like the Web.

In this paper, we present SNE, a scalable, unsupervised, and domain-independent system that simultaneously extracts high-level relations and concepts, and learns a semantic network [20] from text. It first uses TextRunner to extract ground facts as triples from text, and then extract knowledge from the triples. TextRunner’s triples are noisy, sparse, and contain many co-referent objects and relations. Our system has to overcome these challenges in order to extract meaningful high-level relations and concepts from the triples in an unsupervised manner. It does so with a probabilistic model that clusters objects by the objects that they are related to, and that clusters relations by the objects they relate. This allows information to propagate between clusters of relations and clusters of objects as they are created. Each cluster represents a high-level relation or concept. A concept cluster can be viewed as a node in a graph, and a relation cluster can be viewed as links between the concept clusters that it relates. Together the concept clusters and relation clusters define a simple semantic network. Figure 1 illustrates part of a semantic network that our approach learns. SNE is short for Semantic Network Extractor.

SNE is based on Markov logic [22], and is related to the Multiple Relational Clusterings (MRC) model [12] we recently proposed. SNE is our first step towards creating a system that can extract an arbitrary semantic network directly from text. Ultimately, we want to tightly integrate the information extraction TextRunner component and the knowledge learning SNE component to form a self-contained knowledge extraction system. This tight integration will enable information to flow between both tasks, allowing them to be solved jointly for better performance [14].

We begin by briefly reviewing Markov logic in the next section. Then we describe our model in detail (Section 3). Next we describe related work (Section 4). After that, we report our experiments comparing our model with three Extracting Semantic Networks from Text via Relational Clustering 3 alternative approaches (Section 5). We conclude with a discussion of future work (Section 6).

2 Markov Logic

Markov logic combines first-order logic with Markov networks. In first-order logic [9], formulas are constructed using four types of symbols: constants, variables, functions, and predicates. (In this paper we use only function-free logic.) …

…

---

,

	Author	volume	Date Value	title	type	journal	titleUrl	doi	note	year
2008 ExtractingSemNetsFromText	Stanley Kok Pedro Domingos			Extracting Semantic Networks from Text via Relational Clustering			http://www.cs.washington.edu/homes/pedrod/papers/ecml08.pdf