Role-based Association of Verbs, Actions, and Sentiments with Entities in Political Discourse

Abstract

A crucial challenge in measuring how text represents an entity is the need to associate each representative expression with a relevant entity to generate meaningful results. Common solutions to this problem are usually based on proximity methods that require a large corpus to reach reasonable levels of accuracy. We show how such methods for the association between an entity and a representation yield a high percentage of false positives at the expression level and low validity at the document level. We introduce a solution that combines syntactic parsing, semantic role labeling logic, and a machine learning approach-the role-based association method. To test our method, we compared it with prevalent methods of association on the news coverage of two entities of interest-the State of Israel and the Palestinian Authority. We found that the role-based association method is more accurate at the expression and the document levels.

Clause Analysis: Using Syntactic Information to Automatically Extract Source, Subject, and Predicate from Texts with an Application to the 2008–2009 Gaza War

“…Part of the errors were due to nonverbal action patterns such as “the Israeli invasion of Gaza.” Since the system depends on verbal patterns to extract predicates, such nominal patterns are only dealt with by a special pattern relying on a dictionary of actions, and identify the subject correctly only if it is listed in a possessive structure such as in this example. This can possibly be alleviated with either better lexical resources or with a machine learning system such as presented by Fogel-Dror, Sheafer, Shenhav, and Van Atteveldt (2015), but both solutions lack the simplicity and transparency of the clause analysis presented here. A second class of errors was caused by relations between clauses.…”

“…The system is also related to the work described by Van Atteveldt, Kleinnijenhuis, and Ruigrok (2008) who use syntactic information to attribute sentiment expressions to actors, and the study presented by Fogel-Dror, Sheafer, Shenhav, and Van Atteveldt (2015), who use a machine learning approach on top of syntactic relations for sentiment attribution. Compared to these systems, the current system has a different focus: it is aimed at enhancing frequency-based methods to allow framing by incorporating the relations between actors, rather than analyzing the sentiment expressed in a text.…”

A comparison of approaches for imbalanced classification problems in the context of retrieving relevant documents for an analysis

“…Tweets, posts from online platforms, and newspaper articles are types of documents that are often analyzed in social science and whose analysis typically involves some preliminary retrieval step (see, e.g. [ 12 , 14 , 47 , 62 , 84 , 122 , 135 , 143 ]. The entities of interest in social science studies vary widely with regard to their nature and their level of abstraction.…”

“…In many applications, researchers rely on applying human-created sets of keywords and regard those documents as relevant that comprise at least one of the keywords (see, e.g. [ 12 , 14 , 27 , 47 , 57 , 84 , 99 , 122 , 135 ]). Yet, research indicates that humans are not good at generating comprehensive keyword lists and are highly unreliable at the task [ 61 , p. 973–975].…”