539 research outputs found
Sort by
CUNI Neural Experiments
No full textWe presented the current state of development and experiments with neural MT at Charles University within the project QT21
The QT21 Combined Machine Translation System for English to Latvian
No full textThis paper describes the joint submis-
sion of the QT21 projects for the
English
→
Latvian translation task of the
EMNLP 2017 Second Conference on Ma-
chine Translation
(WMT 2017). The sub-
mission is a system combination which
combines seven different statistical ma-
chine translation systems provided by the
different groups.
The systems are combined using either
RWTH’s system combination approach,
or
USFD’s
consensus-based
system-
selection approach. The final submission
shows an improvement of 0.5 B
LEU
compared to the best single system on
newstest2017
UFAL Medical Corpus 1.0
No full textUFAL Medical Corpus is a collection of parallel corpora assembled for the purposes of the EU projects KConnect, Khresmoi and HimL aiming at more reliable machine translation of medical texts
LanideNN: Multilingual Language Identification on Character Window
No full textIn language identification, a common first step in natural language processing, we want to automatically determine the language of some input text. Monolingual language identification assumes that the given document is written in one language. In multilingual language identification, the document is usually in two or three languages and we just want their names. We aim one step further and propose a method for textual language identification where languages can change arbitrarily and the goal is to identify the spans of each of the languages. Our method is based on Bidirectional Recurrent Neural Networks and it performs well in monolingual and multilingual language identification tasks on six datasets covering 131 languages. The method keeps the accuracy also for short documents and across domains, so it is ideal for off-the-shelf use without preparation of training data
Deep Architectures for Neural Machine Translation
No full textIt has been shown that increasing model depth improves the quality of neural machine translation.
However, different architectural variants to increase model depth have been proposed, and so far, there has been no thorough comparative study.
In this work, we describe and evaluate several existing approaches to introduce depth in neural machine translation.
Additionally, we explore novel architectural variants, including deep transition RNNs, and we vary how attention is used in the deep decoder.
We introduce a novel "BiDeep" RNN architecture that combines deep transition RNNs and stacked RNNs.
Our evaluation is carried out on the English to German WMT news translation dataset, using a single-GPU machine for both training and inference.
We find that several of our proposed architectures improve upon existing approaches in terms of speed and translation quality.
We obtain best improvements with a BiDeep RNN of combined depth 8, obtaining an average improvement of 1.5 BLEU over a strong shallow baseline.
We release our code for ease of adoption
Findings of the WMT 2017 Biomedical Translation Shared Task
No full textAutomatic translation of documents is
an important task in many domains, in-
cluding the biological and clinical do-
mains. The second edition of the Biomed-
ical Translation task in the Conference of
Machine Translation focused on the au-
tomatic translation of biomedical-related
documents between English and various
European languages. This year, we ad-
dressed ten languages: Czech, German,
English, French, Hungarian, Polish, Por-
tuguese, Spanish, Romanian and Swedish.
Test sets included both scientific publica-
tions (from the Scielo and EDP Sciences
databases) and health-related news (from
the Cochrane and UK National Health Ser-
vice web sites). Seven teams participated
in the task, submitting a total of 82 runs.
Herein we describe the test sets, participat-
ing systems and results of both the auto-
matic and manual evaluation of the trans-
lations
An Exploration of Word Embedding Initialization in Deep-Learning Tasks
No full textWord embeddings are the interface between
the world of discrete units of text
processing and the continuous, differentiable
world of neural networks. In this
work, we examine various random and
pretrained initialization methods for embeddings
used in deep networks and their
effect on the performance on four NLP
tasks with both recurrent and convolutional
architectures. We confirm that pretrained
embeddings are a little better than
random initialization, especially considering
the speed of learning. On the other
hand, we do not see any significant difference
between various methods of random
initialization, as long as the variance
is kept reasonably low. High-variance initialization
prevents the network to use the
space of embeddings and forces it to use
other free parameters to accomplish the
task. We support this hypothesis by observing
the performance in learning lexical
relations and by the fact that the network
can learn to perform reasonably in its task
even with fixed random embeddings
The IR Task at the CLEF eHealth Evaluation Lab 2016: User-centred Health Information Retrieval
No full textThis paper details the collection, systems and evaluation methods used in the IR Task of the CLEF 2016 eHealth Evaluation Lab. This task investigates the effectiveness of web search engines in providing access to medical information for common people that have no or little medical knowledge. The task aims to foster advances in the development of search technologies for consumer health search by providing resources and evaluation methods to test and validate search systems. The problem considered in this year’s task was to retrieve web pages to support the information needs of health consumers that are faced by a medical condition and that want to seek relevant health information online through a search engine. As part of the evaluation exercise, we gathered 300 queries users posed with respect to 50 search task scenarios. The scenarios were developed from real cases of people seeking health information through posting requests of help on a web forum. The presence of query variations for a single scenario helped us capturing the variable quality at which queries are posed. Queries were created in English and then translated into other languages. A total of 49 runs by 10 different teams were submitted for the English query topics; 2 teams submitted 29 runs for the multilingual topics
Edinburgh’s Statistical Machine Translation Systems for WMT16
Get PDFThis paper describes the University of Edinburgh’s
phrase-based and syntax-based
submissions to the shared translation tasks
of the ACL 2016 First Conference on Machine
Translation (WMT16). We submitted
five phrase-based and five syntaxbased
systems for the news task, plus one
phrase-based system for the biomedical
task
Incorporation of a valency lexicon into a TectoMT pipeline
No full textIn this paper, we focus on the incorporation of a valency lexicon into TectoMT system for Czech-Russian language pair. We demonstrate valency errors in MT output and describe how the introduction of a lexicon influenced the translation results. Though there was no impact on BLEU score, the manual inspection of concrete cases showed some improvement