A Parallel Corpus of Theses and Dissertations Abstracts

Felipe Soares, Gabrielli Harumi Yamashita, Michel Jose Anzanello

Abstract

In Brazil, the governmental body responsible for overseeing and coordinating post-graduate programs, CAPES, keeps records of all theses and dissertations presented in the country. Information regarding such documents can be accessed online in the Theses and Dissertations Catalog (TDC), which contains abstracts in Portuguese and English, and additional metadata. Thus, this database can be a potential source of parallel corpora for the Portuguese and English languages. In this article, we present the development of a parallel corpus from TDC, which is made available by CAPES under the open data initiative. Approximately 240,000 documents were collected and aligned using the Hunalign tool. We demonstrate the capability of our developed corpus by training Statistical Machine Translation (SMT) and Neural Machine Translation (NMT) models for both language directions, followed by a comparison with Google Translate (GT). Both translation models presented better BLEU scores than GT, with NMT system being the most accurate one. Sentence alignment was also manually evaluated, presenting an average of 82.30% correctly aligned sentences. Our parallel corpus is freely available in TMX format, with complementary information regarding document metadata.

Introduction

Modern Machine Translation (MT) systems, whether based on Statistical (SMT) or Neural (NMT) architectures, require massive amounts of parallel data to achieve high accuracy. While general-purpose corpora are available, they often lack the specialized terminology and formal register required for scientific domains. This paper addresses the shortage of high-quality, large-scale parallel data for the Portuguese-English language pair in the scientific domain by leveraging the CAPES Theses and Dissertations Catalog (TDC).

Methodology

The development process involved a multi-stage pipeline designed to transform raw database records into a clean, aligned corpus. Data was retrieved from the CAPES open data portal covering the years 2013–2016. Key steps included:

  1. Filtering & Cleaning: Removing entries with missing abstracts and performing language verification to ensure correct field mapping.
  2. Preprocessing: Addressing the challenges of all-uppercase source data through case folding and specialized sentence segmentation.
  3. Alignment: Utilizing LF Aligner (a wrapper for Hunalign) to match Portuguese sentences with their English translations using length heuristics and bilingual dictionaries.
  4. Model Training: Training Moses-based SMT and OpenNMT-based neural models to evaluate the corpus's utility against general-purpose benchmarks.

The CAPES TDC Parallel Corpus

The primary contribution of this work is the CAPES TDC Parallel Corpus, which consists of 239,018 documents and approximately 1.29 million aligned sentence pairs. It encompasses roughly 34 million tokens per language across nine distinct knowledge areas, as detailed below:

Table 1: Distribution of documents, sentences, and tokens across different knowledge areas in the CAPES TDC Parallel Corpus.

The inclusion of metadata such as university, year, and keywords in an SQLite database format adds multi-dimensional value, enabling research beyond translation, such as sociolinguistic studies or cross-disciplinary analysis.

Evaluation and Results

The effectiveness of the corpus was evaluated using the BLEU score metric. Models trained on this domain-specific data were compared against Google Translate (GT), a general-purpose state-of-the-art system.

Table 2: BLEU scores for SMT (Moses), NMT (OpenNMT), and Google Translate across development and test partitions.

The results demonstrate that domain specificity is crucial. The NMT model (OpenNMT) achieved the highest scores, notably outperforming Google Translate in the English → Portuguese direction (43.22 vs. 38.92). Manual evaluation of alignment quality showed that 82.30% of sentence pairs were correctly aligned, with most errors stemming from complex segmentation issues in long abstracts.

Conclusion

The CAPES TDC Parallel Corpus provides a significant new resource for Portuguese-English NLP. By outperforming general-purpose models in the scientific domain, this work underscores the importance of specialized datasets. Future work will focus on improving segmentation and exploring tasks like automatic document classification and cross-language text mining.

Summary updated

Abstract

In Brazil, the governmental body responsible for overseeing and coordinating post-graduate programs, CAPES, keeps records of all theses and dissertations presented in the country. Information regarding such documents can be accessed online in the Theses and Dissertations Catalog (TDC), which contains abstracts in Portuguese and English, and additional metadata. Thus, this database can be a potential source of parallel corpora for the Portuguese and English languages. In this article, we present the development of a parallel corpus from TDC, which is made available by CAPES under the open data initiative. Approximately 240,000 documents were collected and aligned using the Hunalign tool. We demonstrate the capability of our developed corpus by training Statistical Machine Translation (SMT) and Neural Machine Translation (NMT) models for both language directions, followed by a comparison with Google Translate (GT). Both translation models presented better BLEU scores than GT, with NMT system being the most accurate one. Sentence alignment was also manually evaluated, presenting an average of 82.30% correctly aligned sentences. Our parallel corpus is freely available in TMX format, with complementary information regarding document metadata
BibTeX
@inproceedings{soares2018parallel,
  author = {Felipe Soares and Gabrielli Harumi Yamashita and Michel Jose Anzanello},
  title = {{A Parallel Corpus of Theses and Dissertations Abstracts}},
  booktitle = {PROPOR},
  year = {2018},
  url = {https://openreview.net/forum?id=XpmfN6nHAtz},
}

Loading related content...

Figure