Esposito, Fabrizio (2017) Unsupervised Recognition of Motion Verbs Metaphoricity in Atyical Political Dialogues. [Tesi di dottorato]

[img]
Preview
Text
fabrizio_esposito-doctoral_dissertation.pdf

Download (1MB) | Preview
[error in script] [error in script]
Item Type: Tesi di dottorato
Lingua: English
Title: Unsupervised Recognition of Motion Verbs Metaphoricity in Atyical Political Dialogues
Creators:
CreatorsEmail
Esposito, Fabriziofabrizio.esposito3@unina.it
Date: 10 October 2017
Number of Pages: 246
Institution: Università degli Studi di Napoli Federico II
Department: Studi Umanistici
Dottorato: Human mind and gender studies
Ciclo di dottorato: 29
Coordinatore del Corso di dottorato:
nomeemail
Striano, Mauramaura.striano@unina.it
Tutor:
nomeemail
Cutugno, FrancescoUNSPECIFIED
Date: 10 October 2017
Number of Pages: 246
Uncontrolled Keywords: Metaphors, Word Embeddings, Topic Modelling, Motion Verbs, Press Briefings
Settori scientifico-disciplinari del MIUR: Area 10 - Scienze dell'antichità, filologico-letterarie e storico-artistiche > L-LIN/12 - Lingua e traduzione - lingua inglese
Date Deposited: 15 Oct 2017 19:24
Last Modified: 08 Mar 2018 11:21
URI: http://www.fedoa.unina.it/id/eprint/11918
DOI: 10.6093/UNINA/FEDOA/11918

Abstract

This thesis deals with the unsupervised recognition of the novel metaphorical use of lexical items in dialogical naturally-occurring political texts without the recourse to task-specific hand-crafted knowledge. The focus of metaphorical analysis is represented by the class of verbs of motion identified by Beth Levin. These lexical items are investigated in the atypical political genre of the White House Press Briefings due to their role in the communication strategies deployed in public and political discourse. The Computational White House press Briefings (CompWHoB) corpus, a large resource developed as one of the main objectives of the present work, is used for the extraction of the press briefings including the lexical items under analysis. The metaphor recognition of the motion verbs is addressed employing unsupervised techniques which theoretical foundations primarily lie in the Distributional Hypothesis theory, i.e. word embeddings and topic models. Three algorithms are developed for the task, combining the Word2Vec and the Latent Dirichlet Allocation models, and based on two approaches representing their foundational theoretical framework. The first one is defined as "local" and leverages the syntactic relations of the verb of motion with its direct object for the detection of metaphoricity. The second one, termed as "global", drifts away from the use of the syntactic knowledge as feature of the system hence only using the information inferred from the discourse context. The three systems and their corresponding approaches are evaluated against 1220 instances of verbs of motion annotated by human judges according to their metaphoricity. Results show that the global approach performs poorly compared to the other two models also implementing the local approach, leading to the conclusion that a syntax-agnostic system is still far from reaching a significant performance. The evaluation of the local approach yields instead promising results, proving the importance of endowing the machine with syntactic knowledge as also confirmed by a qualitative analysis on the influence of the linguistic properties of metaphorical utterances.

Actions (login required)

View Item View Item