539 research outputs found
Sort by
TectoMT – a deep-linguistic core of the combined Chimera MT system
Get PDFChimera is a machine translation system that combines the TectoMT deep-linguistic core with phrase-based MT system Moses. For English–Czech pair it also uses the Depfix post-correction system. All the components run on Unix/Linux platform and are open source (available from Perl repository CPAN and the LINDAT/CLARIN repository). The main website is https://ufal.mff.cuni.cz/tectomt. The development is currently supported by the QTLeap 7th FP project (http://qtleap.eu)
Adapting SMT Query Translation Reranker to New Languages in Cross-Lingual Information Retrieval
Get PDFWe investigate adaptation of a supervised machine learning model for reranking of query translations to new languages in the context of cross-lingual information retrieval. The model is trained to rerank multiple translations produced by a statistical machine translation system and optimize retrieval quality. The model features do not depend on the source language and thus allow the model to be trained on query translations coming from multiple languages. In this paper, we explore how this affects the final retrieval quality. The experiments are conducted on medical-domain test collection in English and multilingual queries (in Czech, German, French) from the CLEF eHealth Lab series 2013--2015.
We adapt our method to allow reranking of query translations for four new languages (Spanish, Hungarian, Polish, Swedish). The baseline approach, where a single model is trained for each source language on query translations from that language, is compared with a model co-trained on translations from the three original languages
Using MT-ComparEval
Get PDFThe paper showcases the MT-ComparEval tool for qualitative evaluation of machine translation (MT). MT-ComparEval is an opensource
tool that has been designed in order to help MT developers by providing a graphical user interface that allows the comparison
and evaluation of different MT engines/experiments and settings
Tools and Guidelines for Principled Machine Translation Development
Get PDFThis work addresses the need to aid Machine Translation (MT) development cycles with a complete workflow of MT evaluation methods. Our aim is to assess, compare and improve MT system variants. We hereby report on novel tools and practices that support various measures, developed in order to support a principled and informed approach of MT developmen
QTLeap WSD/NED Corpora: Semantic Annotation of Parallel Corpora in Six Languages
No full textThis work presents parallel corpora automatically annotated with several NLP tools, including lemma and part-of-speech tagging,
named-entity recognition and classification, named-entity disambiguation, word-sense disambiguation, and coreference
Pronoun Prediction with Linguistic Features and Example Weighing
Get PDFWe present a system submitted to the WMT16 shared task in cross-lingual pronoun prediction, in particular, to the English-to-German and German-to-English sub-tasks. The system is based on a linear classifier making use of features both from the target language model and from linguistically analyzed source and target texts. Furthermore, we apply example weighing in classifier learning, which proved to be beneficial for recall in less frequent pronoun classes. Compared to other shared task participants, our best English-to-German system is able to rank just below the top performing submissions
Task3 Patient-Centred Information Retrieval: Team CUNI
No full textIn this paper we present our participation as the team of
the Charles University at Task3 Patient-Centred Information Retrieval. In the monolingual task and its subtasks, we submitted two runs: one is based on language model approach and the second one is based on vector space model. For the multilingual task, Khresmoi translator, a Statistical Machine Translation (SMT) system, is used to translate the queries into English and get the n-best-list. For the baseline system, we take 1-best-list translation and use it for the retrieval, while for other runs, we use a machine learning model to rerank the n-best-list translations and predict the translation that gives the best CLIR performance in terms of P@10. We present set of features to train the model, these features are generated from the SMT verbose output, different resources like UMLS Metathesaurus, MetaMap, document collection and from the Wikipedia
articles. Experiments on previous CLEF eHealth IR tasks test set show significant improvement brought by the reranker over the baseline system
Machine Translation of Medical Text in the KConnect Project
No full textThis paper presents the work on Machine Translation (MT) that has been conducted within the KConnect project, funded under the H2020 programme and focused on development and commercialization of cloud-based services for multilingual Semantic Annotation, Semantic Search and Machine Translation of electronic health records and medical publications. We first present the main goal and role of MT in the project and then briefly describe the main methods and components developed in the project, including training data acquisition, methods of domain adaptation, deployment of MT as cloud-based web-service, and the training toolkit allowing easy adaptation of the MT service to new languages
A Pilot Eye-Tracking Study of WMT-Style Ranking Evaluation
No full textThe shared translation task of the Workshop of Statistical Machine Translation (WMT) is one of the key annual events of the field. Participating
machine translation systems in WMT translation task are manually evaluated by relatively ranking five candidate translations
of a given sentence. This style of evaluation has been used since 2007 with some discussion on interpreting the collected judgements but
virtually no insight into what the annotators are actually doing. The scoring task is relatively cognitively demanding and many scoring
strategies are possible, influencing the reliability of the final judgements. In this paper, we describe our first steps towards explaining the
scoring task: we run the scoring under an eye-tracker and monitor what the annotators do. At the current stage, our results are more of a
proof-of-concept, testing the feasibility of eye tracking for the analysis of such a complex MT evaluation setup
Planting Trees in the Desert: Delexicalized Tagging and Parsing Combined
Get PDFVarious unsupervised and semi-supervised
methods have been proposed to tag and parse
an unseen language. We explore delexicalized
parsing, proposed by (Zeman and Resnik,
2008), and delexicalized tagging, proposed
by (Yu et al., 2016). For both approaches
we provide a detailed evaluation on Universal
Dependencies data (Nivre et al., 2016), a de-facto standard for multi-lingual morphosyntactic processing (while the previous work used other datasets). Our results confirm that in separation, each of the two delexicalized techniques has some limited potential when no annotation of the target language is available. However, if used in combination, their errors multiply beyond acceptable limits. We demonstrate that even the tiniest bit of expert annotation in the target language may contain significant potential and should be used if available