Martín Quesada Zaragoza


2022

pdf
Limitations and Challenges of Unsupervised Cross-lingual Pre-training
Martín Quesada Zaragoza | Francisco Casacuberta
Proceedings of the 15th biennial conference of the Association for Machine Translation in the Americas (Volume 1: Research Track)

Cross-lingual alignment methods for monolingual language representations have received notable attention in recent years. However, their use in machine translation pre-training remains scarce. This work tries to shed light on the effects of some of the factors that play a role in cross-lingual pre-training, both for cross-lingual mappings and their integration in supervised neural models. The results show that unsupervised cross-lingual methods are effective at inducing alignment even for distant languages and they benefit noticeably from subword information. However, we find that their effectiveness as pre-training models in machine translation is severely limited due to their cross-lingual signal being easily distorted by the principal network during training. Moreover, the learned bilingual projection is too restrictive to allow said network to learn properly when the embedding weights are frozen.

2020

pdf
Translating Knowledge Representations with Monolingual Word Embeddings: the Case of a Thesaurus on Corporate Non-Financial Reporting
Martín Quesada Zaragoza | Lianet Sepúlveda Torres | Jérôme Basdevant
Proceedings of the 6th International Workshop on Computational Terminology

A common method of structuring information extracted from textual data is using a knowledge model (e.g. a thesaurus) to organise the information semantically. Creating and managing a knowledge model is already a costly task in terms of human effort, not to mention making it multilingual. Multilingual knowledge modelling is a common problem for both transnational organisations and organisations providing text analytics that want to analyse information in more than one language. Many organisations tend to develop their language resources first in one language (often English). When it comes to analysing data sources in other languages, either a lot of effort has to be invested in recreating the same knowledge base in a different language or the data itself has to be translated into the language of the knowledge model. In this paper, we propose an unsupervised method to automatically induce a given thesaurus into another language using only comparable monolingual corpora. The aim of this proposal is to employ cross-lingual word embeddings to map the set of topics in an already-existing English thesaurus into Spanish. With this in mind, we describe different approaches to generate the Spanish thesaurus terms and offer an extrinsic evaluation by using the obtained thesaurus, which covers non-financial topics in a multi-label document classification task, and we compare the results across these approaches.