English. This paper describes the automatic procedure we developed to convert an Italian dependency treebank into a different format. We defined about 4,250 formal rules for rewriting dependencies and token tags as well as an algorithm for treebank rewriting able to avoid rule interference. At the end of this process a large portion of the whole treebank was automatically converted, with very few errors, leaving only a small amount of work to be done manually.
The availability of large annotated language resources is a prerequisite for the development of reliable automatic annotation tools using machine learning techniques.
Automatic tools able to enrich real texts with sentence syntactic structures are central instruments in Natural Language Processing (NLP) pipelines for a reliable annotation of text corpora. Modern NLP parsers heavily depend on complex training phases performed by examining manually annotated treebanks. Data sparsity, especially for low-resourced languages, seriously affect parsers performances, forcing scholars to annotate more and more data.
Since 2012 the state-of-the-art for Italian
treebanks were not so satisfactory: three different
projects and institutions produced three treebanks
using different background theories, different
formats and also different syntactic structures. They
were the Italian Syntactic Semantic Treebank
ISST
Size (approx.) tokens sentences Type
TUT
VIT
ISST and TUT were used as gold standards in
various evaluation campaigns
This paper describes the latest effort for the Italian treebank merging: the conversion, harmonisation and integration of the written sections of VIT, not previously included into ISST, with the other two resources for reaching a global amount of about 600,000 tokens and 23,000 sentences syntactically annotated. For practical issues we decided to convert VIT into the MIDT format and then use the set of already designed automatic procedures and checking programs to transform it into the final UD format.
There are other notable works aimed at treebank
conversion in various languages, for example we
can cite
The Venice Italian Treebank was created by the
Laboratory of Computational Linguistics of the
Department of Language Sciences, University of
Venice
At a later time, one of the authors converted
the treebank from phrase-structure to dependency
structures
The Merged Italian Dependency Treebank was
created as a first attempt to merge two existing
Italian resources, namely the TUT and a special
version of the ISST treebank named ISST-TANL
The main part of the VIT conversion process
was completely automatic. Using the Semgrex
package3
Semgrex represents nodes in a dependency graph as a (non-recursive) attribute-value matrix. It then uses regular expressions for subsets of attribute values. For example, fword:amo;tag:/N.*/g refers to any node that has a value ‘amo’ for the attribute ‘word’ and a ‘tag’ starting with ‘N’, while ‘fg’ refers to any node in the graph. The most important part of Semgrex is that it allows you to specify relations between nodes or group of nodes. For example, ‘fg=1 <subj fg=2’ finds all the pairs of nodes connected by a directed ‘subj’ relation. Logical connectives can be used to form more complex patterns and node naming (the ‘=’ assignments) can help retrieve matched nodes from the patterns.
Unfortunately Semgrex is simply a query language and, in its original form, cannot be used to rewrite dependency (sub)graphs. In order to extend the possibility of Semgrex, we then modified the original application to manage pairs of patterns: the first is used to search into the treebank for the required subgraphs, and the second is used to specify how the retrieved subsgraphs have to be rewritten. For example the pattern pair ftag:detg=1 >arg ftag:noung=2 --> ftag:ARTg=1 <DET ftag:NNg=2, what we called a ‘Semgrex rule’, changes the direction of the dependency and, at the same time, changes the words tags and relation label. The starting and final patterns have to contain the same number of nodes and dependency edges. Node naming has been the fundamental trick to introduce such extension allowing for node matching between patterns. 4.2
For converting VIT into MIDT format, we manually defined about 4,050 Semgrex rules each capturing a specific syntactic configuration in VIT and transforming it into the MIDT schema and about 150 rules for residual tag rewriting. We spent about six months for writing the entire set of rules.
We have defined a set of new rewriting operations on a general dependency treebank: DEL REL(graphID, depID, headID): deletes a dependency edge between two graph nodes; INS REL(graphID, depID, headID, label): inserts a new labelled dependency edge between two graph nodes; REN TAG(graphID, nodeID, tag): replace the tag of a specific graph node.
The conversion task has been implemented as a three-steps process: first of all, each Semgrex rule is always applied to the original treebank producing a set of matching subgraphs that have to be rewritten; for each match, a set of specific operations for rewriting the subgraph corresponding to the processed matching are generated and stored; last, the whole set of operations produced processing the entire set of Semgrex rules, each applied to the original treebank, is sorted by graphID, duplicates are removed and every operation is applied graph by graph respecting the following order: first dependency deletions, second dependency insertions and lastly tag renaming.
This way of processing the original treebank and transforming it into the new format should guarantee that we do not experience rule interference during the conversion, because the generation of the rewriting operations due to the Semgrex rules application is decoupled from the real treebank rewriting. 5
The set of rules manually written for converting VIT dependency structures can be subdivided into two macro-classes: (a) rules that do not modify the structures and (b) rules that need to modify the dependencies, both in term of edge direction and in term of different structuring between the involved nodes.
Regarding the rules that do not modify the dependency structures, they simply rename the dependency label using a 1:1 or an N:1 look-up table, as VIT, with respect to MIDT, typically involves more specific dependency types. Figure 1 outlines some simple examples of such kind of conversions.
There are, of course, other kind of operations on subgraphs that require also the rewriting of the dependency structure. A good example concerns relative clauses in which the role of the relative pronoun and, as a consequence, the connections of the edge expressing the noun modification are completely different in the two formalisms. Figure 2 shows one example of this kind of rewriting.
Cases of coordination presented several problems: in VIT the head of the coordinated structure is linked to the connective and then the two (or possibly more) coordinated structures can be linked with a wide range of different dependency types (e.g. between phrases - sn, sa, savv, sq, sp, predicative complements - acomp, ncomp, adjuncts - adj, adjt, adjm, adjv, subjects - subj, objects - obj, etc.) leading to a large number of different combinations. Moreover, each dependency combination has to be further specified by the different token tags. MIDT represents coordinate structures in a different way: the connective and the second conjunct are both linked to the first conjunct that is connected to the head of the coordinated structure.
Figure 3 shows one example: the first formal rule represents an abstract rule pattern that has to be filled with all the real tag combinations found in VIT, generating a huge number of different rules, one of them outlined by the second complete formal rule. This process generated more than 2,800 different rules for handling all the coordinated structures in VIT.
There is also a need for a third kind of rules for rewriting single PoS-tags that might have remained unchanged during the main conversion process.
One further point deserves some discussion. In VIT, articulated prepositions are represented as two different tokens both linked with a common head: the preposition is tagged part/partd/partda and usually connected to the head with some kind of modification relations and the article is always tagged art and linked to the head with a det relation. In MIDT articulated prepositions are represented by a single token. As we said before, our process does not allow re-tokenisation rules. Given that MIDT is only an intermediate format and the goal is to convert VIT into the UD standard that requires two tokens for this phenomenon, we decided to avoid any re-tokenisation and to convert such structures linking the preposition to the head and the article to the preposition by introducing a new, dummy, relation label ‘REL EA’. 6
Applying all the 4,250 Semgrex rules, we obtained a converted treebank in which 228,534 out of 280,641 dependency relation were automatically converted, giving a global coverage of 81.4%.
To test the effectiveness of the conversion procedure and the conversion rules we randomly selected 100 sentences (2582 dependency relations to be converted) from the treebank and manually checked every newly created dependency relation, both in term of the connected nodes and the assigned label.
We obtained the following results: among the 2008 relations that have been automatically converted we found 125 wrongly converted dependency relations. So, on this sample, we obtained a coverage of 2008/2582 = 77.8%, slightly less than on the whole treebank, with a conversion error rate = 125/2008 = 6.2%. 7
This paper presents the procedure we developed to convert VIT, one Italian treebank, into a different format. Most of the described conversion procedure rely on an automatic algorithm based on formal rules that is able to automatically convert the 81.4% of the treebank. This procedure can be, in principle, adaptable to any conversion between different dependency treebank formats.
The formal rules has been manually defined by using a well known dependency search procedure, Semgrex from StanfordNLP group, properly extended to handle rewriting rules and the final result was manually evaluated to test the effectiveness of the written rules obtaining a very small error rate.
To the best of our knowledge, there is no general
purpose tool available to automatise this task for
dependency graphs. We can find some powerful
converters in literature but they are usually tied to
specific pair of tagsets (often tailored to the Penn
treebank)
Even if the described procedure can convert a large part of the treebank automatically with a very small quantity of errors, the conversion certainly needs a careful manual analysis to complete the task and check the new treebank for remaining mistakes. The VIT treebank contains a lot of specific and peculiar dependency subgraph for representing phenomena in a very detailed way. Trying to capture all these different variations into formal rules can result in a very large rule set mostly composed of rule that handle single cases. We stopped the production of new rules when this situation arose.
We wish to thank Rodolfo Delmonte and Maria Simi for their precious suggestions and explanations for the analysis of linguistic phenomena and for defining the conversion process.