Articles

We present an extensive experimental study of a Statistical Machine Translation system, Moses (Koehn et al., 2007), from the point of view of its learning capabilities. Very ac- curate learning curves are obtained, by us- ing high-performance computing, and extrap- olations are provided of the projected perfor- mance of the …...

In this article, we present MetaMorpho, a rule based machine translation system that was used to create MorphoLogic’s submission to the WMT08 shared Hungarian to English translation task. The architecture of Meta- Morpho does not fit easily into traditional categories of rule based systems: the building blocks of its grammar …...

Minimum error rate training (MERT) is a widely used learning procedure for statistical machine translation models. We contrast three search strategies for MERT: Powell’s method, the variant of coordinate descent found in the Moses MERT utility, and a novel stochastic method. It is shown that the stochastic method obtains test …...

In this paper a new discriminative word align- ment method is presented. This approach models directly the alignment matrix by a con- ditional random field (CRF) and so no restric- tions to the alignments have to be made. Fur- thermore, it is easy to add features and so all available …...

We describe two methods to improve SMT accuracy using shallow syntax information. First, we use chunks to refine the set of word alignments typically used as a starting point in SMT systems. Second, we extend an N-gram- based SMT system with chunk tags to better account for long-distance reorderings. Exper- …...

This paper describes our statistical machine translation systems based on the Moses toolkit for the WMT08 shared task. We address the Europarl and News conditions for the follow- ing language pairs: English with French, Ger- man and Spanish. For Europarl, n-best rescor- ing is performed using an enhanced n-gram or …...

We explore the augmentation of statistical ma- chine translation models with features of the context of each phrase to be translated. This work extends several existing threads of re- search in statistical MT, including the use of context in example-based machine translation (Carl and Way, 2003) and the incorporation of …...

The treatment of ‘spurious’ words of source language is an important problem but often ignored in the discussion on phrase-based SMT. This paper explains why it is impor- tant and why it is not a trivial problem, and proposes three models to handle spurious source words. Experiments show that any …...

This paper reports on the participation of the TALP Research Center of the UPC (Universitat Politècnica de Catalunya) to the ACL WMT 2008 evaluation campaign. This year’s system is the evolution of the one we em- ployed for the 2007 campaign. Main updates and extensions involve linguistically motivated word re- …...

Trasl´an makes full use of MT during our translation workflow, where the raw output from our Machine Translation (MT) system is passed onto human translators who perform post-editing (if necessary) to arrive at the final translation. Within Trasl´an we have found that using MT has enabled us to increase the …...

Our work in this area started as a re- search project but when L2F joined TecnoVoz, a Portuguese national consortium including Academia and Industry partners, our focus shifted to real-time professional solutions. The integration of our domain-independent Spoken Dialogue System (SDS) framework into commercial products led to a major reengineering …...

In this paper we describe a natural language generation system which takes as its input a set of assertions encoded as a semantic graph and outputs a data structure connecting the se- mantic graph to a text which expresses those assertions, encoded as a TAG syntactic tree. The scope of …...

Training word alignment models on large cor- pora is a very time-consuming processes. This paper describes two parallel implementations of GIZA++ that accelerate this word align- ment process. One of the implementations runs on computer clusters, the other runs on multi-processor system using multi-threading technology. Results show a near-linear speed- …...

In this paper, we describe our efforts to build on WORDNET resources, using WORDNET lexical data, the data format that it comes with and WORDNET’s software infrastructure in order to generate a biomedical extension of WORDNET, the BIOWORDNET. We began our efforts on the assumption that the soft- ware resources …...

It is a widely accepted belief in natural lan- guage processing research that naturally oc- curring data is the best (and perhaps the only appropriate) data for testing text mining sys- tems. This paper compares code coverage us- ing a suite of functional tests and using a large corpus and …...

This paper describes the structure of a test suite for evaluation of clinical question an- swering systems; presents several manually compiled resources found useful for test suite generation; and describes the adaptation of these resources for evaluation of a clinical question answering system. 1...

Natural language processing modules such as part-of-speech taggers, named-entity recog- nizers and syntactic parsers are commonly evaluated in isolation, under the assumption that artificial evaluation metrics for individual parts are predictive of practical performance of more complex language technology sys- tems that perform practical tasks. Although this is an important …...

Experiments in natural language processing and machine learning typically involve run- ning a complicated network of programs to create, process, and evaluate data. Re- searchers often write one or more UNIX shell scripts to “glue” together these various pieces, but such scripts are suboptimal for several rea- sons. Without significant …...

Type checking defines and constrains system output and intermediate representations. We report on the advantages of introducing multi- ple levels of type checking in deep parsing sys- tems, even with untyped formalisms. 1...

Computing precision and recall metrics for named entity tagging and resolution involves classifying text spans as true positives, false positives, or false negatives. There are many factors that make this classification complicated for real world systems. We describe an evaluation system that attempts to control this complexity through a set …...