'Deep' grammatical relations for semantic interpretation

McConville, Mark; Dzikovska, Myroslava O.

doi:10.3115/1608858.1608866

Cited by 1 publication

(1 citation statement)

References 14 publications

(16 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Other related work, in a different vein, is that of [11]. Like us, they extract and normalize verb valence patterns by adopting heuristics to determine the voice of FrameNet's unmarked passives.…”

Section: Related Workmentioning

confidence: 99%

LexPar: A Freely Available English Paraphrase Lexicon Automatically Extracted from FrameNet

Coyne

Rambow

2009

2009 IEEE International Conference on Semantic Computing

View full text Add to dashboard Cite

Abstract-This is a paper about a new resource, namely an English paraphrase dictionary extracted from the FrameNet lexicon and its example data base. I. THE LEXPAR PARAPHRASE DICTIONARYThis paper describes LexPar, a lexical resource for paraphrasing English verbs. Paraphrasing within a language can be useful for various applications, such as machine translation (source-side paraphrasing can increase the likelihood of finding a good translation), multi-document summarization (paraphrasing can help find passages in different documents with the same meaning), information extraction (paraphrasing can help in detecting relevant information from seed search patterns), or dialog systems and other generation applications (paraphrasing can make the output more context-appropriate and less monotone). As a result, there has recently been some interest in detecting paraphrases automatically. The main problem here is is the lack of resources: there are few parallel English-English texts (different translations from the same foreign source are one example), and the range of phenomena in paraphrasing is considerable (complex lexico-syntactic paraphrases), so that it can be difficult to generalize.We take a different approach: we use an existing resource, FrameNet [1], to extract a list of paraphrases for verbs. We associate two verbs if their meanings overlap in a core meaning, even if the mapping from semantic arguments to syntactic arguments is quite different. A typical example is formed by buy and sell: X buys Y from Z and Z sells Y to X are paraphrases because they describe the same underlying situation. In order to exploit such a relation, we need not only know the pair buy, sell , but we also need to know the mapping of the syntactic arguments. In FrameNet, verbs that relate to the same underlying semantics are grouped into a frame, with their syntactic arguments mapped to a set of semantic labels specific to the frame. In theory, this should make the extraction of a paraphrase dictionary simple. In practice, there are three important problems: first, buy and sell are NOT, in fact, in the same frame, but in frames that are related in one of many possible ways; second, verbs in the same frame may not be paraphrases of each other, such as walk and swim; third, the syntactic annotation in FrameNet (which is needed in order to determine the activevoice valence pattern) is not deep, and not entirely reliable.LexPar encompasses both the transformations of the conversive lexical function of Meaning-Text Theory (MTT) [2] and the verb alternations modeled in Beth Levin's verb classes [3] and in VerbNet [4]. In addition, it further generalizes these by including mappings from one set of prepositional phrases to others that reference the same underlying semantic roles (John walked across the field and John walked through the field). We do not include pure synonyms for now which do not involve any syntactic changes (BUY and PURCHASE), but we easily could if this seems useful. Here is an example.

show abstract

Section: Related Workmentioning

confidence: 99%