Statistical models for unsupervised prepositional phrase attachment

Abstract

We present several unsupervised statistical models for the prepositional phrase attachment task that approach the accuracy of the best supervised methods for this task. Our unsupervised approach uses a heuristic based on attachment proximity and trains from raw text that is annotated with only part-of-speech tags and morphological base forms, as opposed to attachment information. It is therefore less resource-intensive and more portable than previous corpus-based algorithm proposed for this task. We present results for prepositional phrase attachment in both English and Spanish.

Web-Based Sources for an Annotated Corpus Building and Composite Proper Name Identification

“…They use the one million of words Wall Street Journal corpus. Other authors, on the contrary, do not use the whole corpus for their research but a subcorpus with specific characteristics [12] which requires huge manual efforts. Therefore, we consider the size of tenths of millions of words and Web-based sources.…”

PP Attachment: Where do We Stand?

“…Some other work does take multiple nouns candidates into consideration, but only nouns that are within a certain window preceding the preposition (Ratnaparkhi, 1998;Belinkov et al, 2014) or all the nouns in the sentence (Foth and Menzel, 2006). Using examples from German, de Kok et al (2017) show that these crude approaches are problematic.…”

“…The parser-predicted PP attachments are represented as <preposition, object of the preposition, candidate> triples, which we collect from both ambiguous and unambiguous PP attachment results. Here, unambiguous attachments refer to prepositions that only have one possible attachment site (Ratnaparkhi, 1998).…”

“…These preferences are usually stated in terms of information-theoretical measures, such as point-wise mutual information. Since hand-annotated treebanks usually do not have enough material to obtain reliable bilexical statistics, these statistics were extracted from raw text (Volk, 2001), automatically tagged (Ratnaparkhi, 1998), chunk parsed (Volk, 2002) or parsed (Hindle and Rooth, 1993;Pantel and Lin, 2000;Mirroshandel et al, 2012) corpora, resulting in auxiliary distributions. Since these seminal works in PP attachment, parsers have become faster (Kübler et al, 2009) and more accurate (Chen and Manning, 2014), opening the possibility to obtain better co-occurrence statistics.…”

Improving Prepositional Phrase Attachment Disambiguation Using the Web as Corpus

“…However, resources such as treebanks are not available for many languages and they are difficult to port, so that a less resource-demanding method is desirable. Ratnaparkhi [7] describes a method that requires only a part-ofspeech tagger and morphological information. His method uses raw text to be trained.…”