The native TUT format

TUT adopts a representation format based on the dependency paradigm centred upon the notion of predicate-argument structure, as described with reference to major Italian linguistic phenomena in the Linguistic notes. The choice of this paradigm that describes syntactic structures using dependency relations between pairs of words, depends on the partial configurationality of the reference language, i.e. Italian is a free word order language.
In TUT the dependency relations are annotated by following the Augmented Relational Structure (ARS) where each relation is implemented as a feature structure that can include values for a morpho-syntactic, a functional-syntactic and a syntactic-semantic component.

The need for a description of grammatical relations more detailed and more proximate to semantics has determined the development of a rich and flexible grammatical relation system for TUT, i.e. around 250 relations annotated at variable degree of specification according to a hierarchical organization. When the annotator cannot select a specific relation to label the dependency edge linking two words, he/she can select a more generic relation from the higher levels of this taxonmy.
To represent some phenomenon involving discontinuity and deletions as well as pro-drop subject, TUT has been enriched with a trace-filler notation (see section 1. Traces and co-indexing in Linguistic notes for examples and further details).
See at the following example:

Each line contains all the information concerning a single node-word X:
the position of X within the linear order of the sentence,
the morphological features of X (in round brackets),
the position of the head-word Y from which X depends, and the name of the relation linking X to Y (both in square brackets).
According to the ARS, the name of each relation may include three components separated by hyphens:
MORPHOSYNTACTIC - FUNCTIONALSYNTACTIC - SEMANTIC (the symbol + is used as a separator between 2 parts of a single component).

The converted formats

Parallel treebanks may serve as a suitable infrastructure for the comparison of parsers from different linguistic frameworks, thus contributing in the investigation of the causes of the irreproducibility of state-of-the-art results on annotations other than Penn Wall Street Journal and languages other than English.
A conversion tool, called TUTtoPENN converter, has been applied to the native TUT in order to generate a Penn-like annotation (called TUT-Penn). As a side effect, has been developed two other formats that show intermediate layers of variation/similarity with respect to the TUT and Penn in terms of both richness of functional-syntactic information (i.e. amount and specificity of grammatical relations) and type of linguistic framework (i.e. constituency versus dependency, or minimal versus maximal projection). For more information and downloading of the conversion tool see the TUTtoPENN converter web page.
The following image shows the cascade of formats (a rich selection of examples in parallel formats is available in Parallel annotations in TUT formats):

• a) native TUT, which is the original dependency-based annotation format developed for the treebank
• b) Constituency-TUT, which is a constituency-based Xbar-like format. In the trees of this format, each word corresponds to a single terminal node, which allways projects in a chain of non-terminal nodes, i.e. phrases, according to the grammatical category of the word itself. For instance, the grammatical category projection chain for a Verb includes, in the order, Vbar (i.e. the intermediate projection) and VP (i.e. the maximal projection), while for a Noun includes Nbar and NP. All the projections can be multiplicated when needed to link together all the constituents of the tree. For instance, in the figure, one more Nbar is needed to link the PP 'di Berisha' with the NP 'governo'. By contrast with the Penn format, Constituency-TUT implements trees which are less flat, i.e. each word corresponds to a single terminal node of the tree, and applies a maximal projection strategy, i.e. all grammatical categories project (at least one) intermediate and (one) maximal projections. Moreover, Constituency-TUT annotates the same functional-syntactic relations of TUT, here annotated on constituents rather than on edges like in TUT.
• c) Augmented Penn, which annotates the dependencies of TUT (when possible on constituents), thus showing a repertory of functional relations larger than standard Penn, but draws trees structurally identical to those of Penn. Like Penn, it implements a representation more flat that Constituency-TUT, since terminal nodes can be used to represent more words, and applies a minimal projection strategy that does not uses intermediate projections in the trees.
• d) TUT-Penn, which includes a few functional relations and implements a minimal projection strategy.

Moreover, from the Constituency-TUT has been derived a treebank for Italian based on Combinatory Categorial Grammar called CCG-TUT (see the CCG-TUT web page for more details). In the process of conversion that generates the CCG-TUT, TUT dependency trees are mapped onto constituency trees (i.e. a binarized ConsTUT trees enriched with labels that distinguish heads, arguments and modifiers) and then mapped into CCG derivations.

TUT corpora and data

Currently the treebank includes five sections:

• The first one is composed of 700 sentences coming from various newspapers. Most of them, however, concern the situation in Albania in the nineties. This section is called NEWS.
• The second section contains sentences that include the Italian verb "vedere" (to see; 350 sentences) and chiamare (to call; 50 sentences). This section is called VEDCH.
• The third section includes 1,100 sentences extracted from the Italian Civil Code of law. This is the CODICECIVILE section.
• 201 sentences concern European directives, from the Italian section of the JRC-Aquis multilingual parallel corpus, and have been put in the section EUDIR.
• The Wiki section includes 459 sentences, randomly chosen from the Italian version of Wikipedia. This is the WIKI section, which has been developed in 2010 to extend the previously released treebank.

Note that the sections NEWS and VEDCH include isolated sentences, while the other four sections involve full paragraphs (Articles of the Code, European Directives and Wikipedia articles respectively).
Finally, it is possible to download 200 English sentences annotated in the TUT standard format, i.e. the ENGLISH section. This could enable people not acquainted with Italian to have a feeling of how the annotation schema is organizad. For peculiar features (and extensions) of the TUT scheme for representing English structures see the section 16. Applying TUT on English in the Linguistic notes.

The following table shows the size of the sections in terms of sentences and tokens in TUT native and CoNLL (where available) formats:

section	sentences	tokens in TUT format	tokens in CoNLL format
CODICECIVILE	1,100	30,669	28,048
NEWS	700	19,134	18,044
VEDCH	400	14,390	12,508
EUDIR	201	7,955	7,426
WIKI	459	15,766	14,746
ENGLISH	200	5,940

TUT old versions

The following table shows the links for downloading the treebank in the native and in constituency-based formats. See the CCG-TUT web page for downloading the treebank in CCG format and for more information about it.

Raw text*	Release	TUT FORMAT	TUT in CoNLL FORMAT	Cons-TUT FORMAT	TUT-PENN FORMAT
civil law newspaper JRC-Passage-Evalita	Evalita**	civil law and newspaper			civil law and newspaper
	1.1	civillaw newspaper
	2.1**	civillaw newspaper JRC-Passage-Evalita		civillaw newspaper JRC-Passage-Evalita	civillaw newspaper
wikipedia	2.2	wikipedia	wikipedia
English	E0.1	English

*Texts originally collected include characters accented by using different strategies (e.g. "à" and "a'") and types of encoding (from ISO-Latin to UTF-8). As regards the characters, some word can therefore occur in the treebank in various writing forms (e.g. "città" and "citta'") not all corresponding to those of the Italian dictionary (like the lemmatized form that TUT includes for each word, which is e.g. "CITTÀ" both for "città" and "citta'"). As regards the encoding, from 2007, all the released annotation of TUT are instead in standard UTF-8.
**The releases Evalita (development set for Evalita 2007) and 2.1 (development set for Evalita 2009) include the data in original TUT and CoNLL format both.