Natural language processing has made rapid progress over the last decades. Yet, computational linguistic resources and tools are restricted to a handful of languages. It seems unrealistic to develop high-quality resources for all languages using traditional methods. Especially the creation of grammars and syntactic treebanks is an expensive process. Various methods aim at overcoming this shortage of NLP resources. Our approach targets the induction of linguistic annotations in a cross-lingual setting: Using a bilingual corpus, existing analysis tools are applied to the resource-rich language side of the bitext. The resulting annotations are projected to the second, resource-poor language using automatic word alignments as a bridge. The projection approach for resource induction is built on the assumption that the linguistic analysis of a sentence carries over to its translation in an aligned parallel corpus. While this assumption does not hold uniformly, the projected annotations can be used to train NLP tools for the target language. This has been shown for PoS tagging [1], NP-bracketing [1], dependency analysis [2, 3], word sense disambiguation [4], extraction of semantic roles [5] and temporal labelling [6].
Within the ParGram project [7], computational grammars for English, French, German, Norwegian, Japanese, Urdu and other languages are written according to the framework of Lexical Functional Grammar. Manual development of large-scale LFG grammars is an expensive process that may be sped up by automation techniques. One strand of work that targets the automatic induction of LFG grammars is the induction from existing syntactic treebanks [8]. However, this method rests on the availability of high-quality treebanks. To overcome the need of treebank creation, we investigate the cross-lingual projection approach to induce syntactically annotated corpora for new languages. Given the considerable divergence of constituent structures across languages, the grammar architecture of LFG, with its strong lexicon component and multiple levels of representations seems especially suited for a cross-lingual grammar induction task. F-structures are largely invariant across languages, and are thus especially suited to serve as a pivot for cross-lingual annotation projection. Following this insight, we pursue cross-lingual projection of grammatical functions (GFs) to induce an f-structure bank for Polish. We project English f-structures to aligned Polish texts, to yield an f-structure bank that may be used to train a dependency parser for Polish. A full-fledged LFG grammar for Polish may be obtained by mapping the induced f-structures to c-structures for Polish.
Our approach follows [2] and adapts their method of projecting dependencies (from English to Spanish/ Chinese) to the LFG framework. The experiment is conducted on the JRC-Acquis Multilingual Parallel Corpus [9], a large collection of European Union legislative texts that - unlike Europarl - includes texts in Polish. We create a subcorpus aligned on the sentence level containing 257,144 sentence pairs. Word alignment is provided by the statistical machine translation system MOSES [10], based on statistics captured from the entire E-P JRC-Acquis corpus. The English side is parsed using the English Pargram LFG grammar, with selection of the most probable analysis. To a certain degree English is an isolating language that makes use of function and non-content words. Polish, by contrast, is a highly inflecting language and needs in general fewer or as many words as English to express the same content. In order to decide whether alignment of one Polish word with one or more English words conforms to general translational mappings, we use unidirectional Polish-English word alignment as a basis for projection.
In the projection step, GFs relating two English words in the source language LFG parses are transferred to the corresponding Polish sentence via the word alignment links. The projected GFs may be incorrect, due to (i) errors in the source annotations obtained from automatic LFG parsing, (ii) poor accuracy of word alignment, or (iii) true mismatches of functional structure between English and Polish. The annotation and alignment errors radically impair the quality of the projected GFs. Those shortages can be overcome by applying correction rules similar to [2] that locally transform the induced Polish f-structures. We defined two post-projection correction rules that are motivated by general linguistic properties of Polish such as (i) absence of articles that could correspond to specifiers (SPEC-DET) 'the' or 'a/an' in English, and (ii) possessive relations expressed by genitive marking of NP as opposed to an of-PP in English. Further correction rules may be formulated, taking into account morpho-syntactic information concerning case, number, tense etc.
We evaluate the quality of the automatically induced f-structures against a gold standard of 50 Polish f-structures from the test corpus which were manually corrected (11.98 t/s for English, 9.76 t/s for Polish). We calculated (i) precision, recall and f-score for exact match of projected GFs, distinguishing direct projection and projection with post-modification using the two correction rules mentioned above, and (ii) accuracy of the projected dependencies taking into account automatically derived (noisy) in contrast to hand-corrected (optimal) word alignments, to establish an upper bound. As expected, direct projection of GFs is noisy (49.98% f-score). Application of language-specific transformation rules considerably improves the accuracy of projected GFs (63.5% f-score). The quality of word alignment is a crucial factor for the quality of projection: projection based on corrected word alignments enhances the quality of the induced f-structures by 12 percentage points (pp) f-score for direct projection and by 19.45 pp f-score for projection with transformation rules. In line with [2], these results clearly indicate that direct projection of GFs is significantly outperformed by projection using post-projection transformations, both for automatic alignment (13.52 pp f-score improvement), and for manually corrected word alignments, yielding an upper bound (20.97 pp f-score improvement). The upper bound (projection based on perfect word alignment) in conjunction with two correction rules achieves an accuracy of 82.95% f-score. According to [9], the quality of word alignment can still be improved using a factored alignment model using part-of-speech, lemma and morphological information. Thus, if both language sides can be enriched with word level information, we can expect an increase of the quality of word alignment, and thus projection quality, within the ranges of the upper bound, possibly enhanced by further post-transformation rules. Compared to the results of unlabeled dependency projection for Spanish/Chinese in [2], we gain higher f-score (on average by 16.4 pp). Compared to [3], precision of direct projection is lower by 18/32 pp. [3] relies on one-to-one alignment links (intersection) only, which increases precision but decreases recall. As [3] does not report recall figures, we cannot compare the results. As our aim is to induce f-structures as complete as possible, it is worth noting that we obtain balanced precision and recall values. [11] represents an LFG-based approach, like ours. We observe that combining argument information from two languages (English and German) enhances precision but degrades recall. Considering f-score, our projection of GFs via automatic word alignment outperforms the projection by [11] by 16.7 pp. However, due to the different languages and corpora involved, a strict comparison of the approaches is difficult.
In future work, we will explore extended training sets, improved word alignments, and use of morpho-syntactic information to further improve the projection quality. The resulting Polish f-structure treebank may be used to train a dependency parser (in [2], a dependency parser for Spanish/Chinese could be trained to yield 72.1%/52.4% labeling f-score, using projected dependency f-scores of 72.1%/53.9% as input for training). We will further explore induction of a full-fledged LFG grammar for Polish, by adding a module that learns f-to-c-structure mappings for Polish.
References