English. The paper presents an extension of the Italian Universal Dependencies Treebank with an “enhanced” representation level (e-IUDT), aimed at simplifying the information extraction process. The modules developed to semi-automatically build e-IUDT were delexicalized to perform cross-language enhancements: preliminary experiments in this direction led to promising results.
Marneffe and Manning, 2008): the so-called
“basic” representation where a close parallelism
to the source text is maintained (i.e. where each
word of the original sentence is present as a
node), and the so-called “collapsed and
propagated” representation which was conceived with
a specific view to information extraction tasks.
Within the current version of UD, the “collapsed
and propagated” representation has evolved into
the graph-based enhanced representation
proposed by
This paper describes the approach we used for developing and validating the enhanced version of the Italian UD Treebank and reports the first results of transfer experiments to English.
Enhanced dependencies were proposed as a way to simplify the process of information extraction. Enhancements, for the most part, result in additional links added to the dependency tree, motivated by inferences, which remain however anchored at the surface representation level. The result of enhancing a dependency tree is a graph, possibly with cycles, but not necessarily a super graph (since some of the original arcs may be discarded).
The current UD guidelines are quite conservative, i.e. they suggest practically feasible enhancements only. Despite this, enhancements cannot always be achieved automatically, and the task is challenging enough to be interesting. According to the guidelines enhanced graphs may contain some or all of the following enhancements, described with particular emphasis on Italian: 1. Added subject relations in control and raising constructions; 2. Shared heads and dependents in coordination; 3. Co-reference in relative clause constructions; 4. Modifier specialization by means of case markers; 5. Null nodes for elided predicates.
In the case of control and raising constructions, the subject of the subordinated non-finite clause is added. Consider the following examples, with controlled and raised subjects marked in bold: 1) Subject control: La mamma ha promesso a Maria di comprare il pane ‘The mother promised Maria to buy the bread’ 2) Object control: La mamma ha convinto Maria a comprare il pane ‘The mother convinced Maria to buy the bread’ 3) Oblique control: La mamma ha chiesto a Maria di comprare il pane ‘The mother asked Maria to buy the bread’ 4) Subject raising: La mamma sembra apprezzare il pane integrale ‘The mother seems to like whole bread’ Control and raising predicates are superficially very similar, with a main difference: whereas Raising predicates have a ‘non-thematic’ argument, all arguments of Control predicates are ‘thematic’. Such a distinction is neutralized in the enhanced UD representation. In both cases, however, the selection of the controlled/raised argument is lexically-driven.
Coordination is another major source of potential enhancements, as information shared among conjuncts is typically attached only to the first conjunct and could be propagated to the other conjuncts, where this is applicable. In propagating information, it is useful to distinguish two cases, a) b)
The book store buys and sells used books. according to whether dependents of the first conjunct are propagated or the head of the first conjunct is propagated instead. Figure 2 shows Italian examples for each case.
The book store sells books and magazines
In basic UD, relative pronouns are normally attached to the main predicate of the relative clause, typically as nominal subjects (nsubj) or direct objects (obj). In the corresponding enhanced graph, the relative pronoun is linked to its antecedent with the ref relation and its dependency to the head of the relative clause is transferred to the antecedent itself, as exemplified in Figure 3 where it can be observed that the resulting enhanced representation contains a cycle.
The book that I read
Adding case information to the relation name of non-core dependents serves the purpose of disambiguating their semantic role. This information is expressed in terms of the preposition or the subordinating conjunction introducing noncore dependents. In particular: nmod and obl relation labels, respectively marking nominal and oblique modifiers introduced by prepositions, are augmented with language specific case information; acl and advcl labels, corresponding respectively to noun modifying clauses and adverbial clauses, are augmented with markers introducing them. A similar type of specialization also applies to the conj dependency label linking conjuncts in coordinated structures, which is specialized with respect to the conjunction type (e, o, oppure …), as identified by the lemma of the cc dependency (i.e. the relation between a conjunct and a preceding coordinating conjunction).
After having dinner he went home
Special null nodes are added in clauses to stand for a predicate which is elided; other cases of ellipsis are not being dealt with in the current UD guidelines due to major difficulties in their reconstruction. This type of enhancement occurs when the basic (i.e. pre-enhancement) tree contains an orphan relation which in the enhanced graph is removed and replaced by the reconstructed explicit syntactic structure. A new null node is added in place of the missing predicate and dependencies are redirected. Figure 5 shows an example of predicate elision, along with the enhanced version which introduces a new node (labeled as E6.1) obtained as a copy of the token ‘chiamava’.
In intimacy she was calling him captain and he [calling her] boss. This is the most problematic among the foreseen UD enhancements, due to several reasons such as: correct insertion points are difficult to anticipate; phraseological verbs and verbs with clitics (either in pronominal form or with clitic complements, see example in Figure 5) would require copying a variable number of tokens (the verb and the object with a shift in gender in the case at hand), which is not always easy to be identified; the appropriate syntactic role of the dependents of the added (i.e. recovered) predicate must be inferred by proper alignment with the dependents of the originally explicit predicate. Moreover, the proposed UD treatment requires a major change in the treebank format with the addition of new tokens with special labeling and numbering. Therefore, the introduction of null nodes calls for an ad hoc treatment and introduces a complexity in the processing of the treebank which is not fully justified if the aim is only to address the cases of predicate elision, for the fact that this is a rare phenomenon in treebanks. Other cases of elision, such as subject elision, are much more meaningful for Italian.
Besides the standard enhancements foreseen for
UD illustrated above, we are currently evaluating
cases that could be treated as such for Italian, and
could possibly be relevant for other languages as
well. These include:
• case information, which could also be added
for some core relations such as ccomp.
Consider as an example the following sentences:
Non so se verrà domani ‘I don’t know
whether (he) will come tomorrow’ vs Non so
quando arriverà ‘I don’t know when (he)
will arrive’. Without enhancing the ccomp
relation, the semantics of the subordinated
clause (conditional vs temporal) remains
underspecified;
• null nodes for elided subjects: Italian is a
pro-drop language and the omission of
explicit subjects occurs quite frequently in
actual language usage; according to
The typology of representation enhancements could also be further extended to neutralize diathesis alternations, as proposed by Candito et al. (2017) for French. In what follows, we focus on the standard UD enhancements, excluding the treatment of predicate elision for which more careful investigation and detailed guidelines are required. UD enhanced representation cannot be generated through a completely automatic process: this is a task that entails a global vision of the tree to be completed and often requires additional linguistic knowledge concerning e.g. raising/control properties and/or selectional preferences of predicates. To build the enhanced Italian UD Treebank (henceforth, e-IUDT), we followed a threestep approach, articulated as follows: 1. Guessing: by making use of heuristics, a script suggests target nodes whose representation might be enhanced, e.g. the best extra subject candidate(s) in raising/control constructions, or the heads/dependents to be propagated in coordinated constructions. During this step, additional annotations are produced in the representation of involved tokens. For example, the annotation ExtraSubjOf = j added to token i is an indication that i is an additional subject headed by j. In other cases, the additional annotation indicates a label to be used for specializing a given relation or whether a conjunct should be propagated. Table 1 summarizes the additional annotations used; 2. Revising: the human annotator is called to validate the proposed changes, automatically generated during the previous step; 3. Enhancing: validated additional annotations are used to automatically generate the enhanced UD representation. Enhancements are not limited to retyping or addition of dependencies; in some cases, they involve the reshaping of the dependency graph, and for this reason an automatic transformation reduces the chances of occasional errors.
The heuristics behind the guessing step make use of lexical resources extracted from the corpus itself: this is the case, for example, of lexical information on raising/control properties of predicates, guiding the identification of extra-subject candidates.
Following the three-step strategy sketched above, we built a gold standard e-IUDT resource on top of the development data set of the Italian UD treebank (Release 2.2), constituted by 11,908 tokens. In Table 2, the first two columns (headed by “IT DEV (GOLD)”) summarize the enhancements contained in the developed resource, which involve 21,75% of the words. Most of them are represented by the specialization of modifiers and conjoining relations, immediately followed by head propagation, relative clauses and extra-subjects. Interestingly enough, it can be noticed that the distribution of enhancements remains quite similar across different subsets of the same language (e.g. the development vs test sets for Italian), whether manually revised (dev) or not (test), or for another language, English. 4
Different cross-lingual techniques have been
developed for adding enhanced dependencies to
existing UD treebanks, both rule-based
An example follows where, without lexical information, the identification of extra subjects fails. Consider the sentence I carri armati … andavano a Budapest … a spegnere i fuochi ‘The tanks ... went to Budapest ... to extinguish the fires’. In UD, the obl relation covers both lexically realized indirect objects and other oblique complements: however, without distinguishing between the two it is impossible to recover the extra subject of the infinitive clause. A suggestion could be to introduce a specialization of the obl relation for identifying indirect objects. Dependency specialization turned out to be a challenging conversion case when applied to the English UD treebank: problems encountered were somehow unexpected, being mostly due to a different strategy for annotating multi-word case markers, not always compliant with the general UD annotation guidelines. This explains the lower results reported in Table 3 for English with respect to Italian. 5
We extended the Italian UD Treebank with an enhanced representation level: Italian is now among the few languages within UD with a gold enhanced Treebank which will be part of Release v2.3. The modules used to semi-automatically build e-IUDT were delexicalized to carry out cross-language enhancements: preliminary results for both Italian and English are promising. The contribution also includes better and more detailed specifications to the constantly inprogress guidelines. Current developments include: from a mono-lingual perspective, extension of the typology of enhancements; from the multi-lingual perspective, testing and extending the enhancement component successfully used with English for other languages.