Ionian University
Αρχιτεκτονικές βαθιάς μάθησης στην Γλωσσική Τεχνολογία: οι περιπτώσεις της αυτόματης αξιολόγησης μηχανικής μετάφρασης και της ανίχνευσης ψευδών ειδήσεων
Abstract
dc:descriptionThe automatic evaluation of machine translation (MT) has proven to be a very significant research topic. Most automatic evaluation methods focus on the evaluation of the output of MT as they compute similarity scores that represent translation quality. This work targets on the performance of MT evaluation. We propose and we present a general deep learning (DL) scheme for learning to classify parallel translations, based on different categories of information (string based linguistic features, natural language processing (NLP) metrics, word embeddings and one human (reference) translation). The outputs were generated from a statistical machine translation (SMT) system and a neural machine translation (NMT) system. The model has been applied in two language pairs: English - Greek (EN-EL) and English - Italian (EN-IT). In this thesis, a variety of experiments with different parameter configurations have been conducted. The thesis also explores the linguistic differences that affect evaluation accuracy between different kinds of corpora. A comparative study between a simple embedding layer (mathematically calculated) and pre-trained embeddings is conducted as well. Moreover, an analysis of the impact of feature selection and dimensionality reduction on classification accuracy has been performed.Results showed that using a neural network (NN) model with different input representations produces results that clearly outperform the state-of-the-art for MT evaluation for EN–EL and EN–IT. It has been observed that the proposed algorithm achieved better results on noisy and small datasets. In addition, for a more integrated analysis of the accuracy results, a qualitative linguistic analysis has been carried out in order to address complex linguistic phenomena. The proposed language independent achieves higher classification accuracy compared with models using BLEU score information as well as other classification approaches, such as Random Forest (RF) and Support Vector Machine (SVM). Moreover, a variation of the proposed schema, a light resource learning schema, i.e. a schema that depends on limited resources, has been introduced, which aims to choose the better translation between two MT outputs, based on information regarding the source segments and linguistic string-based features. Instead of relying on high-level experts’ annotations, a novel automatic metric is proposed for determining the better translation, namely the quality estimation (QE) score. This score is based on string-based features derived from both the source and the MT segments. Experimental results have demonstrated a quite good performance for the proposed feed-forward NN, comparable to the existing state of-the-art models for MT evaluation requiring more sophisticated resources.The proposed schema has been applied in fake news detection task as well. In the past decade, an increasing number of social network users is observed, which may leed to exploitation phenomena by malicious users, which distribute and reproduce fake profiles, fake news and propaganda. The proposed schema has used word embeddings and both linguistic and network account features. Furthermore, tweets have been separated into news headers and news text, whereas an extensive experimental setup has performed classification tests in these and the results have showed high accuracy performance in fake news detection. The proposed deep learning architecture outperforms the state-of-the-art classifiers, while using fewer features and embeddings from the tweet text.
Degree
thesis:*- Grantor dc:publisher
- Ionian University
- Year dc:date
- 2021
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Mouratidi, Despoina
- Μουρατίδη, Δέσποινα
Subjects
dc:subject × 18- Μηχανική μάθηση
- Βαθιά μάθηση
- Μηχανική μετάφραση
- Αυτόματη αξιολόγηση μετάφρασης
- Ελλειπή δεδομένα
- Αυτόματη ανίχνευση ψευδών ειδήσεων
- Machine learning
- Deep learning
- Machine translation
- Automatic machine translation evaluation
- Noise dataset
- Automatic fake news detection
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/49889