Powered by OpenAIRE graph
Found an issue? Give us feedback
addClaim

Latent variable language models

Authors: Tan, Shawn;

Latent variable language models

Abstract

There has been a renewed interest in generative modeling/unsupervised learning for language for downstream natural language understanding tasks. In this thesis, we explore the augmentation of standard language models with latent variables. In the first chapter, we provide a brief introduction of language models, the classical n-gram treatment and the more common Neural Language Models in use today. We also briefly introduce variational autoencoders and the recent work improving upon them. In Chapter 2, we review work that explores the space where latent variable models and language models intersect. We then empirically analyse the effectiveness of a couple of these methods. In particular, we re-implement the models from Bowman et al. (2015) and Yang et al. (2017), and benchmark them against the Penn Treebank dataset with some experiments of our own. In Chapter 3, we discuss an ICML submission: Generating Contradictions, Entailments and Neutral Sentences. In this work, we encode source sentences to a latent distribution space and attempt to manipulate it from there to generate sentences corresponding to the given logical entailment. While our efforts are unsuccessful, we believe that enabling controllable latent variable distributions is an interesting direction to pursue. In Chapter 4, we conclude with a review of the content covered in the thesis, and a higher-level discussion of what possible avenues of future work could resemble.

Dernièrement, il y a eu un renouvellement d'intérêts dans l'application de modèles génératifs en compréhension de la langue. Dans ce mémoire, nous explorons l'ajout de variables latentes dans les modèles de langues traditionnels. Dans le chapitre 1, nous introduisons brièvement les modèles de langues, notamment les modèles n-gram et les modèles de langue neuronaux, couramment utilisés de nos jours. Nous présentons également les auto-encodeurs variationnels ainsi que différents moyens d'améliorer leur performance. Dans le chapitre 2, nous passons en revue les travaux ou des modèles à variables latentes sont appliqués en modélisation de la langue. Nous analysons également l'efficacité de plusieurs de ces méthodes. En particulier, nous analysons les modèles de cite bowman2015generating et cite yang2017improved, et les évaluons entre autres sur Penn Treebank. Dans le chapitre 3, nous présentons un article encore non publié: Generating Contradictions, Entailments and Neutral Sentences. Dans ce travail, nous encodons des phrases sources dans une distribution latente. Nous manipulons par la suite cet espace afin de générer des phrases correspondant à certaines implications logiques. Malgré nos efforts infructueux, nous croyons que l'utilisation de variables latentes contrôlables est une direction intéressante à suivre. Dans le chapitre 4, nous concluons avec un bref survol du mémoire et discutons des travaux futurs possibles.

Keywords

Compréhension du langage naturel, Modèles de langage, latent variables, Apprentissage profond, Traitement du langage naturel, probabilistic models, language models, Modèles génératifs, deep learning, Réseaux de neurones, Apprentissage automatique, neural networks

  • BIP!
    Impact byBIP!
    selected citations
    These citations are derived from selected sources.
    This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
    0
    popularity
    This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network.
    Average
    influence
    This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
    Average
    impulse
    This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network.
    Average
Powered by OpenAIRE graph
Found an issue? Give us feedback
selected citations
These citations are derived from selected sources.
This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
BIP!Citations provided by BIP!
popularity
This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network.
BIP!Popularity provided by BIP!
influence
This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
BIP!Influence provided by BIP!
impulse
This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network.
BIP!Impulse provided by BIP!
0
Average
Average
Average
Upload OA version
Are you the author of this publication? Upload your Open Access version to Zenodo!
It’s fast and easy, just two clicks!