Knowledge-lean projection of coreference chains across languages

Abstract

Common technologies for automatic coreference resolution require either a language-specific rule set or large collections of manually annotated data, which is typically limited to newswire texts in major languages. This makes it difficult to develop coreference resolvers for a large number of the so-called low-resourced languages. We apply a direct projection algorithm on a multi-genre and multilingual corpus (English, German, Russian) to automatically produce coreference annotations for two target languages without exploiting any linguistic knowledge of the languages. Our evaluation of the projected annotations shows promising results, and the error analysis reveals structural differences of referring expressions and coreference chains for the three languages, which can now be targeted with more linguistically-informed projection algorithms.

Multi-source annotation projection of coreference chains: assessing strategies and testing opportunities

“…Then, we run the algorithm in the two settings described above. Note that the projection results for setting 1 are slightly lower as compared to the results reported in (Grishina and Stede, 2015 Table 3: Results for Russian alignments, since we were not interested in maximizing Precision at the cost of low Recall. Our goal was to obtain balanced scores to base our experiments upon.…”

“…We re-implement a simple direct projection algorithm as described in (Postolache et al, 2006) and (Grishina and Stede, 2015), and we run it for the English-German, English-Russian, German-Russian and Russian-German language pairs, since we are not interested in projecting into English. The direct projection is illustrated in Fig.1 where coreference mentions God, his and his work are transferred to the German side via word alignments.…”

“…For our experiments, we have chosen a trilingual parallel annotated coreference corpus of English, German and Russian from (Grishina and Stede, 2015). This corpus was annotated with coreference chains according to the guidelines described in (Grishina and Stede, 2016) Since the corpus was already aligned bilingually for two language pairs -EnglishGerman and English-Russian -we first align the German-Russian corpus at the sentence level using LF Aligner 3 and then select parallel sentences present in all the three languages.…”

“…This corpus was annotated with coreference chains according to the guidelines described in (Grishina and Stede, 2016) Since the corpus was already aligned bilingually for two language pairs -EnglishGerman and English-Russian -we first align the German-Russian corpus at the sentence level using LF Aligner 3 and then select parallel sentences present in all the three languages. This method reduces the average number of sentences per language by 5% and the average number of coreference chains per language by 6% (as compared to the corpus statistics published by Grishina and Stede (2015)). Then we re-run GIZA++ word aligner (Och and Ney, 2003) on the resulting sentences for all the language combinations with German and Russian as targets.…”

Discovery of Discourse-Related Language Contrasts through Alignment Discrepancies in English-German Translation

“…The method that we use to extract transformation patterns is similar to coreference annotation projection applied by Postolache et al (2006) and by Grishina & Stede (2015). Both studies use data manually annotated for coreference relations.…”