Sentence Splitting Research Articles

The Bulgarian National Corpus: Theory and Practice in Corpus Design

The paper discusses several key concepts related to the development of corpora and reconsiders them in light of recent developments in NLP. On the basis of an overview of present-day corpora, we conclude that the dominant practices of corpus design do not utilise adequately the technologies and, as a result, fail to meet the demands of corpus linguistics, computational lexicology and computational linguistics alike.We proceed to lay out a data-driven approach to corpus design, which integrates the best practices of traditional corpus linguistics with the potential of the latest technologies allowing fast collection, automatic metadata description and annotation of large amounts of data. Thus, the gist of the approach we propose is that corpus design should be centred on amassing large amounts of mono- and multilingual texts and on providing them with a detailed metadata description and high-quality multi-level annotation.We go on to illustrate this concept with a description of the compilation, structuring, documentation, and annotation of the Bulgarian National Corpus (BulNC). At present it consists of a Bulgarian part of 979.6 million words, constituting the corpus kernel, and 33 Bulgarian-X language corpora, totalling 972.3 million words, 1.95 billion words altogether. The BulNC is supplied with a comprehensive metadata description, which allows us to organise the texts according to different principles. The Bulgarian part of the BulNC is automatically processed (tokenised and sentence split) and annotated at several levels: morphosyntactic tagging, lemmatisation, word-sense annotation, annotation of noun phrases and named entities. Some levels of annotation are also applied to the Bulgarian-English parallel corpus with the prospect of expanding multilingual annotation both in terms of linguistic levels and the number of languages for which it is available. We conclude with a brief evaluation of the quality of the corpus and an outline of its applications in NLP and linguistic research.

A corpus of full-text journal articles is a robust evaluation tool for revealing differences in performance of biomedical natural language processing tools

BackgroundWe introduce the linguistic annotation of a corpus of 97 full-text biomedical publications, known as the Colorado Richly Annotated Full Text (CRAFT) corpus. We further assess the performance of existing tools for performing sentence splitting, tokenization, syntactic parsing, and named entity recognition on this corpus.ResultsMany biomedical natural language processing systems demonstrated large differences between their previously published results and their performance on the CRAFT corpus when tested with the publicly available models or rule sets. Trainable systems differed widely with respect to their ability to build high-performing models based on this data.ConclusionsThe finding that some systems were able to train high-performing models based on this corpus is additional evidence, beyond high inter-annotator agreement, that the quality of the CRAFT corpus is high. The overall poor performance of various systems indicates that considerable work needs to be done to enable natural language processing systems to work well when the input is full-text journal articles. The CRAFT corpus provides a valuable resource to the biomedical natural language processing community for evaluation and training of new models for biomedical full text publications.

Comparing and combining chunkers of biomedical text

Text chunking is an essential pre-processing step in information extraction systems. No comparative studies of chunking systems, including sentence splitting, tokenization and part-of-speech tagging, are available for the biomedical domain. We compared the usability (ease of integration, speed, trainability) and performance of six state-of-the-art chunkers for the biomedical domain, and combined the chunker results in order to improve chunking performance.We investigated six frequently used chunkers: GATE chunker, Genia Tagger, Lingpipe, MetaMap, OpenNLP, and Yamcha. All chunkers were integrated into the Unstructured Information Management Architecture framework. The GENIA Treebank corpus was used for training and testing. Performance was assessed for noun-phrase and verb-phrase chunking.For both noun-phrase chunking and verb-phrase chunking, OpenNLP performed best (F-scores 89.7% and 95.7%, respectively), but differences with Genia Tagger and Yamcha were small. With respect to usability, Lingpipe and OpenNLP scored best. When combining the results of the chunkers by a simple voting scheme, the F-score of the combined system improved by 3.1 percentage point for noun phrases and 0.6 percentage point for verb phrases as compared to the best single chunker. Changing the voting threshold offered a simple way to obtain a system with high precision (and moderate recall) or high recall (and moderate precision).This study is the first to compare the performance of the whole chunking pipeline, and to combine different existing chunking systems. Several chunkers showed good performance, but OpenNLP scored best both in performance and usability. The combination of chunker results by a simple voting scheme can further improve performance and allows for different precision-recall settings.

Feature generation and representations for protein–protein interaction classification

Automatic detecting protein–protein interaction (PPI) relevant articles is a crucial step for large-scale biological database curation. The previous work adopted POS tagging, shallow parsing and sentence splitting techniques, but they achieved worse performance than the simple bag-of-words representation. In this paper, we generated and investigated multiple types of feature representations in order to further improve the performance of PPI text classification task. Besides the traditional domain-independent bag-of-words approach and the term weighting methods, we also explored other domain-dependent features, i.e. protein–protein interaction trigger keywords, protein named entities and the advanced ways of incorporating Natural Language Processing (NLP) output. The integration of these multiple features has been evaluated on the BioCreAtIvE II corpus. The experimental results showed that both the advanced way of using NLP output and the integration of bag-of-words and NLP output improved the performance of text classification. Specifically, in comparison with the best performance achieved in the BioCreAtIvE II IAS, the feature-level and classifier-level integration of multiple features improved the performance of classification 2.71% and 3.95%, respectively.

Sentence splitting and discourse structure in translations

The typical informational density in German non-fictional prose often triggers sentence splitting in translations into Norwegian. Information given as VP- or NP-adjunctions in the monosentential source versions is extracted and rendered as independent sentences or conjuncts in coordinative structures in the target version. Typically, subordination in the source version signals that the information should not be seen as the main focus of the sentence. Hence, an important means of guiding the reader with regard to discourse function is lost in the translated version. This article aims at showing how certain choices of target version form can be explained by the goal of preserving the discourse structure of the original. A recurring strategy is the use of coordinated clauses or VPs in the target version. Coordination — at least in narratives licensing consequential readings between the first and second conjunct — is exploited as a device of ’downgrading’ the information of the first conjunct to a background role, preserving the second conjunct as the main focus equivalent to the original. Thus, coordination seems to compensate for the more restricted use of syntactic subordination in Norwegian.

Relation extraction and the influence of automatic named-entity recognition

We present an approach for extracting relations between named entities from natural language documents. The approach is based solely on shallow linguistic processing, such as tokenization, sentence splitting, part-of-speech tagging, and lemmatization. It uses a combination of kernel functions to integrate two different information sources: (i) the whole sentence where the relation appears, and (ii) the local contexts around the interacting entities. We present the results of experiments on extracting five different types of relations from a dataset of newswire documents and show that each information source provides a useful contribution to the recognition task. Usually the combined kernel significantly increases the precision with respect to the basic kernels, sometimes at the cost of a slightly lower recall. Moreover, we performed a set of experiments to assess the influence of the accuracy of named-entity recognition on the performance of the relation-extraction algorithm. Such experiments were performed using both the correct named entities (i.e., those manually annotated in the corpus) and the noisy named entities (i.e., those produced by a machine learning-based named-entity recognizer). The results show that our approach significantly improves the previous results obtained on the same dataset.

Analysing the English-Xhosa parallel corpus of technical texts with Paraconc: a case study of term formation processes

The aim of this article is to show how corpus-based research, using computer tools can contribute to the development of technical terminology for Xhosa by translators. Paraconc has been used in analysing a parallel corpus of English-Xhosa texts. Paraconc is a parallel concordancer which makes it easy to analyse translated texts. This software program was developed by Michael Barlow (1995, 2003). He designed this software for linguists and translators who wish to work with translated texts. The alignment process, which is very important for the successful operation of this software, creates links between the source text and the target text. Some screenshots which show alignment of texts at sentence level, merging and splitting of sentences, word frequency lists, parallel concordances, etc. are cited. The author further demonstrates how parallel concordancing, and to a lesser extent, sentence alignment in Paraconc, have been used for identifying terms in the English-Xhosa Parallel Corpus.

The effects of display size and text splitting on reading lengthy text from screen

Abstract The present paper reports on an experimental investigation of reader performance and preferences with a screen-presented journal article. The effects of display size (20 lines and 60 lines) and sentence splitting on readers' manipulation, comprehension and subjective impressions are assessed. The results indicate that neither variable significantly affects comprehension but adjusted manipulation levels are significantly higher in the small window condition. Splitting sentences across screens also caused readers to return to the previous page to reread text significantly more. Subjective data reveal a preference for larger screens and high awareness of text format. Implications for future work are discussed.