Parallel Inference on a Linguistic Knowledge Base

This paper presents a possible solution for the text inference problem extracting information unstated in a text, but implied. The inference algorithm consists of a set of highly parallel search methods that when applied to the knowledge base find contexts of sentences that reveal information relevant to the text. Implementation, results and parallelism analysis are discussed. 1 Statement of the problem This paper addresses the issue of parallelism in a class of problems that is largely unexplored, yet of growing importance. Text inference refers to the problem of extracting information that is not stated directly in a text, but is implied. In this paper we present an inference algorithm that operates on a very large linguistic knowledge base. The system is scalable both in size and accuracy and is highly parallel. Humans have a great ability of figuring out correct inferences from text or speech. Perhaps, this is because we have a great deal of world knowledge and can focus our thoughts to filter out irrelevant facts. Consider the text: S1: John hit the ball with the bat. S2: It landed far away. A system is expected to infer that John was playing baseball, that he is the batter, that hitting the ball causes the ball to move and that John must have hit the ball hard since it landed far away. Our approach for finding inferences is essentially to search for the semantic connections in the knowledge base that correspond to the lexical relations between words in the text. The knowledge base search aims to find the local context of each sentence, and furthermore to establish intersections or connections between these local contexts. Thus, for a given text, the algorithm finds a web of knowledge base concepts that semantically relate to the words in the text. Inferences result as an English interpretation of the context web. Massive parallelism is achieved by allowing parallel searches on the knowledge base. 2 The Knowledge Base The knowledge base consists of a special semantic network (N;G;R), where N is the set of nodes representing words or concepts from WordNet 1.5 [Miller 1995], G is the set of WordNet glosses that define the concept nodes and R is the set of semantic and lexical relations. Nodes The knowledge base nodes are English words and the concepts that represent meanings of words. In its current version WordNet 1.5 has 168,217 words organized in 91,591 concepts, thus a total of 259,812 nodes. Glosses Almost all of the WordNet concepts have a gloss expressed in English that defines that concept. For example, the concept fjoint ventureg has an explanatory gloss: (partnership or conglomerate designed to share risk or expertise) . The English text of a gloss encapsulates the most distinctive defining features of its concept. A gloss may be represented as a directed acyclic graphs (DAGs) with nodes as concepts, and linguistic relations as links. Relations The knowledge base relations, that link various nodes, are the building blocks to our solution for text inference. We distinguish here between the WordNet semantic relations between the nodes and the lexical relations recognized in the node glosses. Altogether there are 345,264 WordNet links in the knowledge base. 3 The Text Inference Algorithm Algorithm assumptions and approach The input of the algorithm is a text which has been lexically tagged, semantically disambiguated and spanned by lexical relations. The text is transformed into a forest of text graphs, in which each graph corresponds to a sentence with the nodes being the sentence concepts while the edges are the lexical relations. Figure 1 illustrates the texts graphs corresponding to sentences S1 and S2. The algorithm’s purpose is to establish semantic connections among the concepts recognized in the text. These