Exploiting word embeddings for modeling bilexical relations
There has been an exponential surge of text data in the recent years. As a consequence, unsupervised methods that make use of this data have been steadily growing in the field of natural language processing (NLP). Word embeddings are low-dimensional vectors obtained using unsupervised techniques on...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2017 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/112211 |
| Acceso en línea: | https://hdl.handle.net/2117/112211 https://dx.doi.org/10.5821/dissertation-2117-112211 |
| Access Level: | acceso abierto |
| Palabra clave: | Àrees temàtiques de la UPC::Informàtica |
| Sumario: | There has been an exponential surge of text data in the recent years. As a consequence, unsupervised methods that make use of this data have been steadily growing in the field of natural language processing (NLP). Word embeddings are low-dimensional vectors obtained using unsupervised techniques on the large unlabelled corpora, where words from the vocabulary are mapped to vectors of real numbers. Word embeddings aim to capture syntactic and semantic properties of words. In NLP, many tasks involve computing the compatibility between lexical items under some linguistic relation. We call this type of relation a bilexical relation. Our thesis defines statistical models for bilexical relations that centrally make use of word embeddings. Our principle aim is that the word embeddings will favor generalization to words not seen during the training of the model. The thesis is structured in four parts. In the first part of this thesis, we present a bilinear model over word embeddings that leverages a small supervised dataset for a binary linguistic relation. Our learning algorithm exploits low-rank bilinear forms and induces a low-dimensional embedding tailored for a target linguistic relation. This results in compressed task-specific embeddings. In the second part of our thesis, we extend our bilinear model to a ternary setting and propose a framework for resolving prepositional phrase attachment ambiguity using word embeddings. Our models perform competitively with state-of-the-art models. In addition, our method obtains significant improvements on out-of-domain tests by simply using word-embeddings induced from source and target domains. In the third part of this thesis, we further extend the bilinear models for expanding vocabulary in the context of statistical phrase-based machine translation. Our model obtains a probabilistic list of possible translations of target language words, given a word in the source language. We do this by projecting pre-trained embeddings into a common subspace using a log-bilinear model. We empirically notice a significant improvement on an out-of-domain test set. In the final part of our thesis, we propose a non-linear model that maps initial word embeddings to task-tuned word embeddings, in the context of a neural network dependency parser. We demonstrate its use for improved dependency parsing, especially for sentences with unseen words. We also show downstream improvements on a sentiment analysis task. |
|---|