📄 A Statistical Parser for Czech

Summary

The paper asks how well a statistical parser built for English carries over to Czech. Czech has rich inflection and fairly free word order, and other Slavic languages, Ukrainian among them, share both traits. The authors train on the Prague Dependency Treebank (PDT), which is annotated with dependencies rather than phrase structure. Their parser is lexicalized and adapts Model 1 of (Collins 97). Because that model expects lexicalized trees, they convert each dependency structure into a tree. Several trees can encode the same dependencies, so the conversion choices define the model's independence assumptions. A simple baseline uses flat trees, "XP" labels and coarse POS categories, and reaches about 72% dependency accuracy. Linguistically motivated changes to how relative clauses, coordination and punctuation are represented, together with changes to the model's parameterization, raise accuracy to about 80%. On news and business text the figure is roughly 82%. The same parser scores 91% on English WSJ text. The work shows that existing English methods can come reasonably close on a morphologically rich, free-order language.

Key points

• The PDT has about 480,000 words of general news, business and science text. It is split into a 19k-sentence training set plus development and evaluation sets of about 3,500 sentences each.
• The Czech morphological analyzer can assign more than 3,000 tags. The PDT includes both manual and machine-assigned tags and lemmas.
• Accuracy is measured as the share of correct dependency links. Each sentence has one link per token, counting an artificial root node.
• On the development set, accuracy rose from 71.9% (baseline) to 79.3% (full model). Final results come from a held-out evaluation set that nobody saw during development, with about 80% overall.
• Newswire text scored over 2% above the average. The Science section was clearly harder, likely because of longer sentences and more varied vocabulary.

Interesting ideas

• I1. Choose the conversion deliberately. When dependency data feeds a phrase-structure parser, decide on purpose how flat the trees are, how nodes are labeled and which tagset is used. Each choice changes what the model treats as independent, which matters whenever data must be adapted to a tool built for another formalism.
• I2. Relabel coordination by a conjunct. The PDT makes the conjunction the head of a coordinated phrase. Labeling that phrase after its second conjunct, rather than the conjunction, recovers the phrase's real category. It also stops the model from happily joining two unlike phrases, which helps any treebank with conjunction-headed coordination.
• I3. Back off instead of losing sentences. When the full system finds no parse, the authors score the output of a simpler variant without bigrams. Sentences that still fail get all their words attached to the root. This keeps rare search failures from distorting the evaluation and suits any comparison of systems with uneven coverage.
• I4. Report accuracy by genre. Breaking results down by text type showed that science articles are much harder than news. This kind of reporting is useful when a corpus mixes domains.
• I5. Separate syntactic and semantic relations. As future work, the authors propose splitting word-word relations into tag-tag syntactic dependencies and lemma-lemma semantic dependencies, using the internal structure of rich tags. This looks promising against data sparsity in highly inflected languages such as Ukrainian.