On the Use of Word Alignments to Enhance Bitext Compression

Abstract

The amount of information that is stored in digital form in more than one language is growing very fast as a consequence of the globalization. Furthermore, there are countries and supra-national entities whose legislation enforces the translation (and storage) of all the official texts into all their official languages.Two texts that are mutual translations are usually referred to as a bilingual parallel corpus or, in short, as a bitext. Compressing independently the two texts of a bitext is far form efficient, since the information conveyed by both texts, the meaning, is similar. We take advantage of this fact to devise a bitext compression algorithm that compresses and stores the two texts that form a bitext simultaneously.In our approach, a single model is used to represent both bitext components. For this purpose, we define a biword as a pair made of two words, each one from a different text, that are mutual translations in the bitext. This new concept allows one to represent with a single symbol two words with high mutual information.The algorithm consists of a simple processing pipeline with two stages. The first one (preprocessing) performs a text (sentence and word) alignment in which no preexisting resources are used; it takes as input the bitext and outputs its biword-based representation. The second stage (compression) implements a customization of the mPPM model in which biwords are used as symbols and a limited length dictionary is used to obtain two-byte codewords.We carried out an exhaustive experimentation with seven language pairs (es-ca, es-gl, es-en, es-fr, es-it, es-pt and fr-en) and bitexts of different sizes. For comparison purposes we compressed the concatenation of the two texts of each bitext using general-purpose compressors. Large improvements are obtained for bitexts of closely-related languages (up to 82.73% for es-gl) due to the monotonicity of the word alignments obtained in the preprocessing stage. For less-related language pairs smaller improvements, between 11.12% (es-pt) and 2.48% (fr-en), are achieved because some non-monotonic word alignments are discarded to ensure that the original bitext can be fully recovered when decompressing. Multiword units may be used in the biwords in order to reduce the number of discarded word alignments. * Funded by Spanish projects TIN2006-15071-C03-01, TIN2006-15071-C03-02 and VA012B08. Miguel A. Martínez-Prieto is granted by JCyL and ESF. We thank Mikel L. Forcada for inspiration.

A Two-Level Structure for Compressing Aligned Bitexts

“…Our strategy, called Two-Level Compressor for Aligned Bitexts (2LCAB), is based on two main ideas: (i) the use of biwords [9], pairs of aligned words, as the basis of the model, that is, as the symbols to compress, and (ii) the use of a two level structure for the representation of the vocabularies, where the vocabulary of biwords, at the second level, is represented in compressed form using the vocabularies of the first level. Figure 2 shows a conceptual description of this scheme.…”

Generalized Biwords for Bitext Compression and Translation Spotting

“…In contrast to PPM, some text-compression methods use words rather than characters as input tokens (Moffat, 1989;Moffat & Isal, 2005). Analogously, Martínez-Prieto, Adiego, Sánchez-Martínez, de la Fuente, and Carrasco (2009), and his colleagues (2009, 2010) propose the use of biwords -pairs of words, each one from a different text, with a high Figure 1: Processing pipeline of a biword-based bitext compression approach.…”

Humans Outperform Machines at the Bilingual Shannon Game

“…The Hutter Prize [12], a competition to compress a 100 m-word extract of English Wikipedia, was designed to futher encourage research in text compression. Bilingual and multilingual text compression is a less-studied field [1,[13][14][15][16][17][18]. These papers provide different algorithms for compressing text in multilingual format, but they do not demonstrate how humans perform on this task.…”