Bootstrapping Structure using Similarity

TL;DR

This paper introduces a similarity-based algorithm inspired by string edit-distance to bootstrap syntactic structure from unannotated sentences, achieving high precision in structure learning on real corpora.

Contribution

The paper presents a novel similarity-based bootstrapping algorithm for extracting syntactic structure from unannotated text, extending prior methods with a focus on sentence similarity.

Findings

Achieved 86.04% non-crossing brackets precision on ATIS corpus.

Achieved 89.39% non-crossing brackets precision on OVIS corpus.

Demonstrated effectiveness of the method in learning structure from flat sentences.

Abstract

In this paper a new similarity-based learning algorithm, inspired by string edit-distance (Wagner and Fischer, 1974), is applied to the problem of bootstrapping structure from scratch. The algorithm takes a corpus of unannotated sentences as input and returns a corpus of bracketed sentences. The method works on pairs of unstructured sentences or sentences partially bracketed by the algorithm that have one or more words in common. It finds parts of sentences that are interchangeable (i.e. the parts of the sentences that are different in both sentences). These parts are taken as possible constituents of the same type. While this corresponds to the basic bootstrapping step of the algorithm, further structure may be learned from comparison with other (similar) sentences. We used this method for bootstrapping structure from the flat sentences of the Penn Treebank ATIS corpus, and compared the resulting structured sentences to the structured sentences in the ATIS corpus. Similarly, the algorithm was tested on the OVIS corpus. We obtained 86.04 % non-crossing brackets precision on the ATIS corpus and 89.39 % non-crossing brackets precision on the OVIS corpus.