ESTUDIO III: Relación entre rendimiento neuromuscular y
6.1.3 Materiales y medición
and strongly divergent languages, a sizeable set of high precision translation can still be extracted.
Moreover, Haghighi et al were the first to use a probabilistic model, which has produced interesting results across a variety of language pairs and data conditions. Unfortunately, they did not compare their context-based model with the standard approach in the same particular setting. Instead, they com-pared their model with a basic spelling-based model that uses edit distance.
In another setting, they compared their combination model with the standard approach. Given the diverse settings existing in today’s prevailing lexicon landscape, further studies are warranted to examine the impacts of transla-tion systems based on comparable, specific settings (i.e., comparing ‘an apple’
with ‘an apple’) to ensure findings could be meaningfully interpreted. Thus, the lack of focus on this particular setting has paved the way for the researcher to conduct several experiments based on the CCA approach.
2.7 Summary and conclusion
This chapter discusses the brief history of bilingual lexicon extraction, in par-ticular by focusing on issues related to the most used linguistic resources in the field. Different types of corpora have been elaborated that deal with the use of the resources for different learning tasks. This chapter also discusses the basic concepts that are relevant to the focus of the undertaken study. In addition, a survey on a number of previous work available in this research field is also presented to highlight prevailing conditions which present the many challenges in this research area.
In the early years, learning from parallel corpora was the primary task in bilin-gual lexicon extraction. Work relying on parallel texts have already recorded significant success. The parallel corpora are mostly constructed manually or semi-automatically, especially by field linguists. These laborious tasks carried
2.7 Summary and conclusion
out by the researchers are inevitable as people do not naturally produce the same text in multiple languages, thus rendering parallel corpora to be not readily available. Fortunately, non-parallel corpora or monolingual corpora are easily accessible, which make them a viable alternative to parallel corpora.
However, the performance of a system using monolingual corpora is generally lower than a system using parallel corpora in cases where the information available is not extensive. Alternatively, extra information such as external bilingual lexicons may improve the former’s performance, but this required additional information can be quite scarce on many occasions.
Other methods with good performance could be identified, but they too have limitation namely being less effective in dealing with the orthographic repre-sentation of a word. Overcoming this limitation would entail seeking a more natural means that could find word pairs that are spelled identically across the languages. However, such approach would not help extend the lexicon to a greater height as a pair of languages does not have many words of simi-lar spelling between them, unless both languages share a simisimi-lar history and posses a similar cultural perspective, notably found in loanwords.
To address these problems, many researchers have attempted to improve the learning from non-parallel corpora by using a context-based approach. This approach could represent a method that is very feasible by exploiting the common patterns observed between word pairs of different languages involv-ing co-occurrence information. In general, the three main steps used in the context-based approaches are as follows:
• Define the context in which a word occurs in the source and target languages.
• Translate as many source context words into the target language using an external bilingual lexicon.
2.7 Summary and conclusion
• Compute the translation for each unknown source word using the simi-larity metric, which uses the target word with the most similar context.
Many of the models that have been proposed thus far are different in many ways, but their main difference lies in the co-occurrence information they ac-quire from the context. For example, Fung (1995) suggests using the context heterogeneity, the information about the number of different tokens preceding and following a word in a context, whereas many others have suggested using highly associated context words. In addition, some researchers recommend using different existing association measures to find highly associated con-text words (Rapp, 1999; Chiao and Zweigenbaum, 2002; Koehn and Knight, 2002; Shezaf and Rappoport, 2010). Likewise, there are other researchers who suggest several improved versions of the measures (Andrade et al., 2010), or advocate several new measures (Prochasson et al., 2009).
Likewise, there are several researchers who have used other context-based techniques, which have achieved encouraging results, namely using confidence weighting for each context term (Fung, 1998), re-scoring translation candidates for a source word (Chiao and Zweigenbaum, 2002), and eliminating incorrect translations (Shezaf and Rappoport, 2010). More interestingly, a purely un-supervised work has also been explored by a few researchers (Diab and Finch, 2000).
All the studies carried out by these researchers have their own specific goal.
Some of these studies aimed to translate general words (Rapp, 1999; Koehn and Knight, 2002), while some others focused on domain specific terms (Chiao and Zweigenbaum, 2002). In addition, many of these studies tried to translate single-word terms (Haghighi et al., 2008; Diab and Finch, 2000; Otero and Campos, 2005) and others attempted to handle multi-word terms.
As a conclusion, many techniques are available in bilingual lexicon extraction.
Some of the existing ideas are interestingly simple, whilst others require rather
2.7 Summary and conclusion
complicated techniques. Therefore, which one is the most effective technique is left open to debate –underscoring the imperative in furthering research on exploring new, novel methods. Currently, the context-based approach remains popular in extracting a bilingual lexicon from non-parallel corpora, with most of the methods resorting to the use of an external bilingual lexicon to improve performance.
In this chapter, the main discussion is mainly focused on the methods that mainly weigh the co-occurrences based on context windows. In addition, there is another different way to weigh the co-occurrence by replacing the traditional window-based with a syntax-based co-occurrence counting approach. An ex-ample of this approach based on Otero and Campos (2008)’s work is elabo-rated by a survey carried by the researcher. Another syntax-based work that warrant analyses are the dependency-based counting (Garera et al., 2009), and the positional-based counting approach using POS equivalents (Tanaka and Matsuo, 1999; Tanaka, 2002) among others. These models are effective and efficient to produce precise lexicon translations as Otero and Campos (2008) argue that “syntactic contexts are considered to be less ambiguous and more sense sensitive than contexts defined as windows of size N ”. Therefore, higher precision should be expected with a syntax-based method compared to a window-based method. However, this contention will be hard to defend as studies that have been performed were mainly based on the availability of extensive external resources and tools, such as lemmatizers and POS taggers.
In actual applications, other additional resources such as pivot lexicons or thesaurus have to be made available. Currently, the literature is quite replete with studies that are mostly based on word co-occurrence in certain window, bag-of-work vectors and/or minimal resources requirement. However, research that focuses on minimally supervised work is quite wanting, where robustness and versatility in learning texts from minimal resources are expected to be the forefront of future systems.
2.7 Summary and conclusion
Based on the preceding discussion pertaining to previous studies, the au-thor observes the following important aspects that entail further investigation.
First, extracting bilingual lexicon from non-parallel corpora represents a new focus in future research as existing lexicon landscape is expected to change based on the fast changing communications taking place in the world. Sec-ond, conditions of minimal resources do exist, but they have not been dealt with thoroughly thus far. Third, further compounding these two situations, advanced linguistic tools, such as lemmatizers and POS taggers, which have features prominently in many studies, may not always be available to other researchers. Based on these prevailing constraints, this doctoral study aims to examine the impacts of minimally supervised learning using minimal resources, including latent data, which will represent an important breakthrough by us-ing new, innovative approach in lexicon translation efforts.
In the next chapter, we present our initial experiments’ results using that approaches, but first, we review some related approaches or decisions that have been taken by other researchers in their work in order to perform each stage in a general bilingual lexicon work. The sub tasks include, for instance, the resource acquisition and the selection of the lists of vocabularies.
Chapter 3
A General Framework, Related Approaches and Initial Experiments
3.1 Introduction
Chapter 2 described several previous studies that are quite notable in this re-search field of bilingual lexicon extraction. Most previous studies were based on the context vector approach. Hence, their methods have not so much differ-ences apart from different measures they choose or different external resources they acquire, although there were previous studies that provide interesting techniques in their work; such as giving higher priority to some specialised words in the context of a word (Prochasson et al., 2009) and back-tracing potential matching (Haghighi et al., 2008); among others. From descriptions provided in the previous chapter, we have identified some important compo-nents in a bilingual lexicon extraction task to be presented in this chapter.
This chapter presents a general framework, which was formed by several im-portant components; including, corpora, initial bilingual lexicons, context win-dows and association measures. For each component, related approaches that have been taken by previous studies are discussed in this chapter. Based on the