Large-scale extraction and use of knowledge from text

Abstract

A large amount of empirically derived world knowledge is essential for many languageprocessing tasks, to create expectations that can help assess plausibility and guide disambiguation. Following Schubert (2002), we present our work on creating a large database of "tuples" from the output of a parser, thus implicitly capturing simple world knowledge expectations, and then utilizing it for two tasks, namely improving parsing and improving the plausibility assessment of paraphrase rules used in textual entailment. With regard to parsing, our goal is to improve the performance of a large-scale phrase structure parser/grammar of English through a method that does not depend on the need to handannotate a large number of sentences beyond those available from existing treebanks. With regard to the plausibility of paraphrase rules, our goal is to use the tuples to assess the plausibility of instantiated rules used in textual entailment, and thus improve the confidence rankings of the entailments that used those rules. In preliminary experiments with these two applications, the use of the tuple data improved performance.

Semantic Knowledge Extraction from Research Documents

“…Peter Clark and Phil Harrison [26] worked on knowledge extraction by making database of "tuples" and thus capturing the simple word knowledge. And then using it in improving parsing and the plausibility assessment of paraphrase rules used in textual entailment.…”

On improving parsing with automatically acquired semantic classes

“…Many other authors have pro-426 cessed text with a morphosyntactic parser and then selected one 427 or more surface patterns to extract a set of candidate semantic 428 classes, that in turn are refined [18,19]. 429 In a different direction, [20] argues that text contains general 430 knowledge in form of assertions and they may be exploited after 431 aggregating big amounts of data, like in KNEXT [20], TextRunner 432 [21] or DART [22]. 433 We use both ideas going one step beyond by combining induced 434 classes with named entity types.…”

PrOntoLearn: Unsupervised Lexico-Semantic Ontology Generation Using Probabilistic Methods

“…theory NELL [3] 24 × 7 learning fixed dynamic ML techniques DART [7] world knowledge × × semi-automated RTE [2], and [13] entailment × × ATP NLU [20] commonsense rules × × semi-supervised Text2Onto [6] ontology learning √ √ semi-supervised LexO [24] complex classes √ × semi-supervised FCA [5] taxonomy √ × FCA OP [4], and [23] ontology population available available semi-/supervised a corpus many be large, it might not contain all the necessary evidence of an event of interest. A corpus contains ambiguous statements about an event that leads to a non-determinism of the state of the event.…”