Article Preview
Top1. Introduction
Semantic interpretation is a centric need for different Natural Language Processing applications like Information Extraction, Question Answering, Text Summarization, and Emotion Extraction from Text, Machine Translation and Question Generation. The primary task of Semantic Role Labeling (SRL) is to indicate the semantic relations belonging to a predicate and its associated participants and properties. Such relations are pre-specified in a list of all possible semantic roles for that predicate.
Semantic role labeling is the process, in natural language processing, of detecting the semantic arguments associated with a predicate or verb of a sentence. This detection identifies the arguments classification into specific roles. Semantic arguments include Agent, Patient, Instrument, Locative, Temporal, Manner and Cause Aspects. A semantic role is a relationship having a syntactic predicate, which is often the verb of a sentence (Wu, 2010). For example ’he likes the deer’, the predicate or is ‘likes’. ‘He’ and ‘deer’ have the semantic role labels agent and subject (theme). The crucial fact about semantic roles is that regardless of their syntactic structure, the underlying predicates remain the same. Recognizing and labeling semantic arguments is a key task for answering “Who”, “When”, “What”, “Where”, “Why”, etc.
Researchers have been proposed to either identify semantic roles or to build semantic classifier. Most of the existing models for automatic semantic role labeling are based on a full syntactic parse of the sentence. This parsing defines argument boundaries and accordingly, relevant information for classifiers’ training is needed to disambiguate role labels. Of the earliest models, (Hirst 1988) presented a foundation to deal with the semantic complexities of text, which is suitable basis for both lexical and structural disambiguation. MindNet identified words labelled with semantic relations and structures acquired from semantic rules of natural language (Richardson, Dolan & Vanderwende, 1998).
In 1990s, statistical machine learning approaches have been increasingly developed in the computational linguistics domain. Models was able to learn from complex linguistic knowledge like learning sub-categorization frames (Brent & Berwick, 1991) and classifying verbs according to argument structure properties (Merlo & Stevenson, 2001). Gildea et al. (Gildea & Jurafsky, 2002) presented a system to make predicate-argument structure that reads from parse tree. This structure identified semantic roles through predicate, agent, theme and manner.
Hierarchical SRL proposed by Moschitti, Giuglea, Coppola and Basili (2005) generalized the classical two-level approach (boundary detection and classification) and showed more efficiency and accuracy. A semantic frame, classically pictured in Fillmore (1975), Fillmore (1977), Fillmore (1978), Fillmore (1982) and Fillmore (1985) is a conceptual representation describing an event, relation, or object and the participants in it. FrameNet shows how frame-semantic annotations are created as frame-specific semantic roles generalizations that are not possible with more traditional linguistic approaches. Such conceptual roles are considered frame elements. Frame semantics relates words into semantic frames, where semantic and syntactic properties of predicating words are characterized (Johnson & Fillmore, 2000). A framework for semantic role annotations based on the FrameNet paradigm is introduced by Shen and Lapata (2007) Semantic role assignment is used to extract question answering through an optimization problem in a bipartite graph.