Example-based machine translation using the marker hypothesis
Gough, Nano (2005) Example-based machine translation using the marker hypothesis. PhD thesis, Dublin City University.
Full text available as:
The development of large-scale rules and grammars for a Rule-Based Machine Translation (RBMT) system is labour-intensive, error-prone and expensive. Current research in Machine Translation (MT) tends to focus on the development of corpus-based systems which can overcome the problem of knowledge acquisition.
Corpus-Based Machine Translation (CBMT) can take the form of Statistical Machine Translation (SMT) or Example-Based Machine Translation (EBMT). Despite the benefits of EBMT, SMT is currently the dominant paradigm and many systems classified as example-based integrate additional rule-based and statistical techniques. The benefits of an EBMT system which does not require extensive linguistic resources and can produce reasonably intelligible and accurate translations cannot be overlooked. We show that our linguistics-lite EBMT system can outperform an SMT system trained on the same data.
The work reported in this thesis describes the development of a linguistics-lite EBMT system which does not have recourse to extensive linguistic resources. We apply the Marker Hypothesis (Green, 1979) — a psycholinguistic theory which states that all natural languages are ‘marked’ for complex syntactic structure at surface form by a closed set of specific lexemes and morphemes. We use this technique in different environments to segment aligned (English, French) phrases and sentences. We then apply an alignment algorithm which can deduce smaller aligned chunks and words. Following a process similar to (Block, 2000), we generalise these alignments by replacing certain function words with an associated tag. In so doing, we cluster on marker words and add flexibility to our matching process. In a post hoc stage we treat the World Wide Web as a large corpus and validate and correct instances of determiner-noun and noun-verb boundary friction.
We have applied our marker-based EBMT system to different bitexts and have explored its applicability in various environments. We have developed a phrase-based EBMT system (Gough et al., 2002; Way and Gough, 2003). We show that despite the perceived low quality of on-line MT systems, our EBMT system can produce good quality translations when such systems are used to seed its memories.
(Carl, 2003a; Schaler et al., 2003) suggest that EBMT is more suited to controlled translation than RBMT as it has been known to overcome the ‘knowledge acquisition bottleneck’. To this end, we developed the first controlled EBMT system (Gough and Way, 2003; Way and Gough, 2004). Given the lack of controlled bitexts, we used an on-line MT system Logomedia to translate a set of controlled English sentences, We performed experiments using controlled analysis and generation and assessed the performance of our system at each stage. We made a number of improvements to our sub-sentential alignment algorithm and following some minimal adjustments to our system, we show that our controlled EBMT system can outperform an RBMT system.
We applied the Marker Hypothesis to a more scalable data set. We trained our system on 203,529 sentences extracted from a Sun Microsystems Translation Memory. We thus reduced problems of data-sparseness and limited our dependence on Logomedia. We show that scaling up data in a marker-based EBMT system improves the quality of our translations. We also report on the benefits of extracting lexical equivalences from the corpus using Mutual Information.
Archive Staff Only: edit this record