Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

This report provides a new state of affairs with regard to language technology for Dutch, alanguage with approx. 25 million speakers. Language technology for Dutch is highly developed and the importance and status of Dutch is confirmed by other measurements, such asthe number of online sales, which is growing strongly, and so is the presence of Dutch online

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Academic collection --- Conferences - Meetings --- Computational linguistics --- Netherlands --- Congresses

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

In this paper, the technique of domain adaptation is applied to improve the statistical machine translation of a particular baseline engine. More specifically, cross-entropy and class based data selection are used to select the most appropriate data from an out-of-domain (OOD) corpus given an in-domain (ID) corpus. Both methods consist of selecting a list of k OOD sentences ranked according to how similar they are to the ID corpus. The descending amount of k best lines, with k each time divided by two, will, hypothetically, suit the ID data increasingly better and will each time discard ‘noise’ from the OOD data which lies furthest from the ID corpus. These different sized sets will be analysed to see if a certain threshold, or optimal k amount of sentences can be observed. In the MT engine, the data selected ‘pseudo in-domain data’ will be combined with the ID data in the tuning step, using log-linear interpolation or a back-off model. Finally, the translation evaluation scores will be compared to see which method yields the most optimisation with respect to the original baseline MT engine. This technique will be applied on the one hand to adapt the domain of an OOD corpus (Europarlv7) to a large legal ID corpus (Belgian Official Gazette) and on the other hand to adapt the domain of an OOD corpus (Digital Corpus of the European Parliament) to an ID corpus which has a similar (European) background and domain (Acquis Communautaire), both to see whether this technique still stands – theoretically – under these conditions.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

In recent years, increasing attention has been paid to the potential of using pictographs to open up the online world, so that users with intellectual disabilities can benefit from the same tools for remote communication (email, instant messaging, social media) which define so much of what it means to be a socially active member of society nowadays. This thesis describes the development of an automatic translation system that aims to enable language-impaired, intellectually disabled individuals to compose written messages simply by selecting a sequence of pictographic images. By way of contrast with existing approaches for pictograph-to-text translation, the system that we develop here, Depicto, takes a 100% rule-based approach. That is, all stages in the translation process make use of linguistic rules, as opposed to statistical data. Such an approach makes it possible to encode elegant generalizations about the pictographic input, which has advantages for the expressivity and consistency of translation. Rule-based approaches are generally also costly, however. Thus, aside from the obvious objective of testing whether this approach can actually be realized, we also explore how development can be made more feasible. In addition, we set two design criteria: first, the system must be sensitive to the needs of its users; second, it must be possible to extend to other target languages. (Currently, the system translates to Dutch.) In Chapter 2, we introduce all third-party resources used by the system. In Chapter 3, we show how the pictographic symbol set Sclera can be analysed as a natural language and how this language can be modelled by a constraint-based grammar. This grammar is written in an implemented variant of the HPSG framework. In the first half of Chapter 4, we see how the semantic structure of analysed pictographic sequences is translated so as to be compatible with the input expected by the target language grammar. This happens in the second module in the Depicto chain. In the second half of Chapter 4, we describe a basic grammar model of Dutch and show how this is used to generate well-formed sentences based on this translated semantic structure. Next, in the first half of Chapter 5, we evaluate the system as a whole. We argue that both its precision, i.e., ability to produce well-formed output, and performance are high, but are forced to concede that its coverage is limited. We also show how the Depicto system fares when pitted against a fundamentally statistical translation system. All in all, we conclude that Depicto is able to translate (a subset of) pictographic sequences into well-formed natural language sentences. By adopting an explicitly modular design, we try to make the system as appealing as possible to developers dealing with other target languages. At the same time, extending Depicto's analysis module remains a costly task. Whether such costliness is an adequate trade-off for the high quality of the system's output will be determined by future work. As a translation system, Depicto succeeds. However, as an assistive writing tool, it currently falls short. Its analysis module imposes rather stringent word order constraints on its input, and its output is an unfiltered list of all possible translation hypotheses. In future work, we will focus on minimizing the first limitation and removing the second altogether. This will most likely involve exciting experiments in hybridization.