A Large Resource of Patterns for Verbal Paraphrases
18006Authors
nan
Abstract
Paraphrases play an important role in natural language understanding, especially because there are fluent jumps between hidden paraphrases in a text. For example, even to get to the meaning of a simple dialog like I bought a computer. How much did the computer cost? involves quite a few steps. A computational system may actually have a huge problem in linking the two sentences as their connection is not overtly present in the text. However, it becomes easier if it has access to the following paraphrases: [HUMAN] buy [ARTIFACT] ⇐⇒ [HUMAN] pay [PRICE] for [ARTIFACT] ⇐⇒ [ARTIFACT] cost [HUMAN] [PRICE], and also to the information that I IsA [HUMAN] and computer IsA [ARTIFACT]. In this paper we introduce a resource of such paraphrases that was extracted by investigating large corpora in an unsupervised manner. The resource contains tens of thousands of such pairs and it is available for academic purposes. Keywords: verbal paraphrase, deep learning. 1.
Keywords
Full Content
A Large Resource for Patterns of Verbal Paraphrases Octavian Popescu1, Ngoc Phuoc An Vo2, Vadim Sheinin3 IBM Research, Yorktown Heights, US 1o.popescu@us.ibm.com, 2ngoc.phuoc.an.vo@ibm.com, 3vadims@us.ibm.com Abstract Paraphrases play an important role in natural language understanding, especially because there are fluent jumps between hidden paraphrases in a text. For example, even to get to the meaning of a simple dialog like I bought a computer. How much did the computer cost? involves quite a few steps. A computational system may actually have a huge problem in linking the two sentences as their connection is not overtly present in the text. However, it becomes easier if it has access to the following paraphrases: [HUMAN] buy [ARTIFACT] ⇐⇒ [HUMAN] pay [PRICE] for [ARTIFACT] ⇐⇒ [ARTIFACT] cost [HUMAN] [PRICE], and also to the information that I IsA [HUMAN] and computer IsA [ARTIFACT]. In this paper we introduce a resource of such paraphrases that was extracted by investigating large corpora in an unsupervised manner. The resource contains tens of thousands of such pairs and it is available for academic purposes. Keywords: verbal paraphrase, deep learning. 1. Introduction When two phrases can be interchanged in a text without al- tering the meaning of the whole, we observe a paraphrasing relationship. Paraphrasing is a fundamental property of nat- ural languages, and it is normally recognized as ”saying the same thing with different words”. Proposing a paraphras- ing relation is a hard task for natural language processing (NLP) systems. The task of recognition of paraphrases was proposed in various SemEval competitions (Butnariu et al., 2009; Mihalcea et al., 2010; Specia et al., 2012; Xu et al., 2015) as an independent task or as a part of larger tasks like semantic similarity, textual entailment, etc. The resource we created is instrumental for all these tasks. Some para- phrases are made out of nominal phrases that contain only ostensible nouns and their adjectival determiners, like in a fourteen year old boy ⇐⇒ a teenager. Another type of paraphrases are the ones that involve a verbal constituent, like abandon a kid ⇐⇒ ignore parental obligations. This later type includes nominalizations, that is, even if the phrase has only noun constituents, at least one of them is a noun coming from a verb, like abandoning a kid. One major difference between nominal vs. verbal paraphrases is that the first ones are basically context independent, that is they can be substituted in a text directly, while the sec- ond are context dependent, their replacement in a sentence requires changes in the syntactic and semantic role of their complements and adjuncts. The resource we compiled contains a list of pairs, each member being centered on a verb and its arguments. A pair is a valid verbal paraphrase relation given a certain context that is represented via types associated with each argument. For a given sentence that contains only one verb phrase, we can extract a set of paraphrases. For example, in Figure 1, for the sentence I pay 1,200 for a laptop from Bestbuy, we …
Content truncated to 500 words
Labeled Text
Paraphrases play an important role in natural language understanding, especially because there are fluent jumps between hidden paraphrases in a text. For example, even to get to the meaning of a simple dialog like I bought a computer. How much did the computer cost? involves quite a few steps. A computational system may actually have a huge problem in linking the two sentences as their connection is not overtly present in the text. However, it becomes easier if it has access to the following paraphrases: [HUMAN] buy [ARTIFACT] ⇐⇒ [HUMAN] pay [PRICE] for [ARTIFACT] ⇐⇒ [ARTIFACT] cost [HUMAN] [PRICE], and also to the information that I IsA [HUMAN] and computer IsA [ARTIFACT]. In this paper we introduce a resource of such paraphrases that was extracted by investigating large corpora in an unsupervised manner. The resource contains tens of thousands of such pairs and it is available for academic purposes. Keywords: verbal paraphrase, deep learning. 1.
Publication Details
- Paper ID:
- 18006
- Year:
- 2018
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan