PoSTWITA-UD: an Italian Twitter Treebank in Universal Dependencies

PoSTWITA-UD: an Italian Twitter Treebank in Universal Dependencies Manuela Sanguinetti⋆, Cristina Bosco⋆, Alberto Lavelli•, Alessandro Mazzei⋆, Oronzo Antonelli‡, Fabio Tamburini⋄ ⋆Department of Computer Science, University of Turin, Italy •FBK-ict, Trento, Italy ‡DISI / ⋄FICLIT, University of Bologna, Italy {msanguin, bosco, mazzei}@di.unito.it, lavelli@fbk.eu, antonelli.oronzo@gmail.com, fabio.tamburini@unibo.it Abstract Due to the spread of social media-based applications and the challenges posed by the treatment of social media texts in NLP tools, tailored approaches and ad hoc resources are required to provide the proper coverage of specific linguistic phenomena. Various attempts to produce this kind of specialized resources and tools are described in literature. However, most of these attempts mainly focus on PoS-tagged corpora and only a few of them deal with syntactic annotation. This is particularly true for the Italian language, for which such a resource is currently missing. We thus propose the development of PoSTWITA-UD, a collection of tweets annotated according to a well-known dependency-based annotation format: the Universal Dependencies. The goal of this work is manifold, and it mainly consists in creating a resource that, especially for Italian, can be exploited for the training of NLP systems so as to enhance their performance on social media texts. In this paper we focus on the current state of the resource. Keywords: social media language, Twitter, Italian, Universal Dependencies 1. Introduction The increasing reliance on the popularity of the Internet and social media in every-day life has led, among other things, to the proliferation of the so-called user-generated contents. As one of the most popular social media, Twitter is among the main providers of this type of contents, the usage of which in scientific research ranges from data analysis, sen- timent analysis and opinion mining, to language technolo- gies. Often, though, Twitter user-generated contents are not edited and/or revised for grammatical accuracy (Lynn et al., 2015). Furthermore, the limited number of characters for each tweet can stimulate creativity and encourage an in- novative and non-standard usage of language conventions. As regards NLP, dealing with this kind of linguistic data presents a series of challenges, which are reflected in the lower output quality of various automatic tools and at dif- ferent linguistic levels (see, e.g., Gimpel et al.(2011), Foster et al. (2011) and Ritter et al. (2011)). These considerations highlight the need for properly an- notated resources to provide adequate coverage of such a phenomenon. This is especially true considering the (rel- atively) little progress made in this field: for the Italian language in particular, at the time of writing, the only linguistically-annotated social media corpora we are aware of are those of Bosco et al. (2016) (i.e. the PoSTWITA corpus) and Rei et al. (2016), both annotated at PoS level only. The contribution of the work hereby presented aims at fill- ing this gap, by creating a treebank of Italian non-canonical texts retrieved from Twitter. The treebank has been built using as starting data the PoSTWITA corpus (mentioned above), and it has been syntactically annotated in compli- ance with the Universal Dependencies format. Several …

Content truncated to 500 words