Marine Carpuat

Word Sense Disambiguation vs. Statistical Machine Translation

We directly investigate a subject of much recent debate: do word sense disambiguation models help statistical machine translation quality? We present empirical results casting doubt on this common, but unproved, assumption. Using a state-of-the-art Chinese word sense disambiguation model to choose translation candidates for a typical IBM statistical MT system, we find that word sense disambiguation does not yield significantly better translation quality than the statistical machine translation system alone. Error analysis suggests several key factors behind this surprising finding, including inherent limitations of current statistical MT architectures.

A stacked, voted, stacked model for named entity recognition

This paper investigates stacking and voting methods for combining strong classifiers like boosting, SVM, and TBL, on the named-entity recognition task. We demonstrate several effective approaches, culminating in a model that achieves error rate reductions on the development and test sets of 63.6% and 55.0% (English) and 47.0% and 51.7% (German) over the CoNLL-2003 standard baseline respectively, and 19.7% over a strong AdaBoost baseline model from CoNLL-2002.

Boosting for named entity recognition

This paper presents a system that applies boosting to the task of named-entity identification. The CoNLL-2002 shared task, for which the system is designed, is language-independent named-entity recognition. Using a set of features which are easily obtainable for almost any language, the presented system uses boosting to combine a set of weak classifiers into a final system that performs significantly better than that of an off-the-shelf maximum entropy classifier.

Improving Arabic-to-English Statistical Machine Translation by Reordering Post-Verbal Subjects for Alignment

We study challenges raised by the order of Arabic verbs and their subjects in statistical machine translation (SMT). We show that the boundaries of post-verbal subjects (VS) are hard to detect accurately, even with a state-of-the-art Arabic dependency parser. In addition, VS constructions have highly ambiguous reordering patterns when translated to English, and these patterns are very different for matrix (main clause) VS and non-matrix (subordinate clause) VS. Based on this analysis, we propose a novel method for leveraging VS information in SMT: we reorder VS constructions into pre-verbal (SV) order for word alignment. Unlike previous approaches to source-side reordering, phrase extraction and decoding are performed using the original Arabic word order. This strategy significantly improves BLEU and TER scores, even on a strong large-scale baseline. Limiting reordering to matrix VS yields further improvements.

One Translation Per Discourse

We revisit the one sense per discourse hypothesis of Gale et al. in the context of machine translation. Since a given sense can be lexicalized differently in translation, do we observe one translation per discourse? Analysis of manual translations reveals that the hypothesis still holds when using translations in parallel text as sense annotation, thus confirming that translational differences represent useful sense distinctions. Analysis of Statistical Machine Translation (SMT) output showed that despite ignoring document structure, the one translation per discourse hypothesis is strongly supported in part because of the low variability in SMT lexical choice. More interestingly, cases where the hypothesis does not hold can reveal lexical choice errors. A preliminary study showed that enforcing the one translation per discourse constraint in SMT can potentially improve translation quality, and that SMT systems might benefit from translating sentences within their entire document context.

Using N-best lists for named entity recognition from Chinese speech

We present the first known result for named entity recognition (NER) in realistic large-vocabulary spoken Chinese. We establish this result by applying a maximum entropy model, currently the single best known approach for textual Chinese NER, to the recognition output of the BBN LVCSR system on Chinese Broadcast News utterances. Our results support the claim that transferring NER approaches from text to spoken language is a significantly more difficult task for Chinese than for English. We propose re-segmenting the ASR hypotheses as well as applying post-classification to improve the performance. Finally, we introduce a method of using n-best hypotheses that yields a small but nevertheless useful improvement NER accuracy. We use acoustic, phonetic, language model, NER and other scores as confidence measure. Experimental results show an average of 6.7% relative improvement in precision and 1.7% relative improvement in F-measure.

A Kernel PCA Method for Superior Word Sense Disambiguation

We introduce a new method for disambiguating word senses that exploits a nonlinear Kernel Principal Component Analysis (KPCA) technique to achieve accuracy superior to the best published individual models. We present empirical results demonstrating significantly better accuracy compared to the state-of-the-art achieved by either naive Bayes or maximum entropy models, on Senseval-2 data. We also contrast against another type of kernel method, the support vector machine (SVM) model, and show that our KPCA-based model outperforms the SVM-based model. It is hoped that these highly encouraging first results on KPCA for natural language processing tasks will inspire further development of these directions.

Identifying concepts across languages: a first step towards a corpus-based approach to automatic ontology alignment

The growing importance of multilingual information retrieval and machine translation has made multilingual ontologies an extremely valuable resource. Since the construction of an ontology from scratch is a very expensive and time consuming undertaking, it is attractive to consider ways of automatically aligning monolingual ontologies, which already exist for many of the worlds major languages.

The NRC System for Discriminating Similar Languages

We describe the system built by the National Research Council Canada for the ”Discriminating between similar languages” (DSL) shared task. Our system uses various statistical classifiers and makes predictions based on a two-stage process: we first predict the language group, then discriminate between languages or variants within the group. Language groups are predicted using a generative classifier with 99.99% accuracy on the five target groups. Within each group (except English), we use a voting combination of discriminative classifiers trained on a variety of feature spaces, achieving an average accuracy of 95.71%, with per-group accuracy between 90.95% and 100% depending on the group. This approach turns out to reach the best performance among all systems submitted to the open and closed tasks.

SemEval-2016 Task~10: Detecting Minimal Semantic Units and their Meanings (DiMSUM)

This task combines the labeling of multiword expressions and supersenses (coarse-grained classes) in an explicit, yet broad-coverage paradigm for lexical semantics. Nine systems participated; the best scored 57.7% F1 in a multi-domain evaluation setting, indicating that the task remains largely unresolved. An error analysis reveals that a large number of instances in the data set are either hard cases, which no systems get right, or easy cases, which all systems correctly solve.