Companion report to the Ph.D. dissertation: "Generative Probabilistic Alignment Models for Words and Subwords: a Systematic Exploration of the Limits and Potentials of Neural Parametrizations”

Rapport d’accompagnement à la thèse de doctorat: "Modèles d’Alignement Probabilistes Génératifs pour les Mots et Sous-mots: une Exploration Systématique des Limites et Potentialités des Paramétrisations Neuronales”

Anh Khoa Ngo Ho (1)

1 TLP - Traitement du Langage Parlé - LISN

This is a companion document to the Ph.D. dissertation "Generative Probabilistic Alignment Models for Words and Subwords: a Systematic Exploration of the Limits and Potentials of Neural Parametrizations” [Ngo Ho, 2021]. This document contains an exhaustive collection of graphs and tables related the analysis of various aspects of automatic word alignment, such as for instance the aligned/unaligned words, rare/unknown words, function/content words, word order divergences, etc; and for six language pairs: English with French, German, Romanian, Czech, Japanese and Vietnamese. We mostly analyze statistical word alignment models (Giza++ and Fastalign) as well as several variants based on neural models: IBM style word alignment models including context-independent models, contextual models, and character-based models; variants of a fully generative neural model based on variational autoencoders. We also document a deep analysis for Byte-Pair-Encoding, a subword tokenization algorithm. For information regarding these various methods, please refer to the thesis.

Mots clés: Machine translation; Word alignment; Artificial neural network

Domaines: Informatique [cs]

Anh Khoa Ngo Ho. Companion report to the Ph.D. dissertation: "Generative Probabilistic Alignment Models for Words and Subwords: a Systematic Exploration of the Limits and Potentials of Neural Parametrizations”. [Technical Report] Université Paris Saclay; Laboratoire Interdisciplinaire des Sciences du Numérique. 2021. ⟨hal-03153752⟩