Search

We automatically predict properties of wines on the basis of smell and flavor descriptions from experts' wine reviews. We show wine experts are capable of describing their smell and flavor experiences in wine reviews in a sufficiently consistent manner, such that we can use their descriptions to predict properties of a wine based solely on language. The experimental results show promising F-scores when using lexical and semantic information to predict the color, grape variety, country of origin, and price of a wine. This demonstrates, contrary to popular opinion, that wine experts' reviews really are informative.

This paper introduces the first efforts towards a common ground for modality annotation for Portuguese. We take into account two existing schemes for European and Brazilian Portuguese, already implemented to written texts, and to spontaneous speech data, respectively. We compare the two schemes, discuss their strengths and weaknesses, and, then, introduce our unifying proposal, pointing out the issues which seem to be already pacified and points that should be considered when the scheme starts to be implemented.

We present an experiment in the automatic tagging of modality in Portuguese. As we are currently lacking a suitable resource with detailed modal information for Portuguese, we experiment with small sample of 160.000 tokens, manually annotated according to the modality scheme that we previously developed for European Portuguese (Hendrickx et al., 2012). We consider modality as the expression of the speaker (or subject)’s attitude towards the proposition and our modality scheme accounts for seven major modal values, and nine sub values. This experiment focuses on three modal verbs, poder ‘may/can’, dever ‘shall/might’ and conseguir ‘manage to/ succeed in/ be able to’, which may all have more than one modal value. We first report on the task of correctly detecting the modal uses of poder and dever, since these two verbs may have non modal meanings. For the identification of the modal value of each occurrence of those three verbs, we applied a machine learning approach that takes into consideration all the features available from a syntactic parser’s output. We obtained the best performance using SVM with a string kernel and the system improved the baseline for all three verbs, with a maximum F-score of 76.2.

The social web has brought forth various genres of interpersonal communication (computer-mediated communication, henceforth: cmc) such as chats, discussion forums, wiki talk pages, Twitter, comment and discussion threads on weblogs and social network sites. These genres display linguistic and structural peculiarities which differ both from speech and from written text. Projects that want to build and exchange cmc corpora would greatly benefit from a standard that allows the user to annotate these peculiarities in TEI. From the perspective of several corpus projects which aim at building and annotating cmc corpora for several European languages, this panel will discuss how the models provided by the TEI encoding framework may be adapted to the special requirements of cmc genres. The basis of the discussion is a customized TEI schema presented at the TEI conference held in Würzburg 2011 (Beißwenger et al. 2012)1. The panel papers will elaborate on basic features that a TEI standard for cmc resources should include and outline open issues with which further work will have to deal. The overall goal of the panel is to stimulate the discussion within the TEI community about how a standard for the representation of cmc in TEI should look like and what might be a practical and reasonable way to go about creating such a standard. In order to push the development of a general standard for the representation of cmc genres and cmc discourse forward, the papers in the panel will present problem overviews for basic issues in representing cmc features in TEI P5 and outline perspectives as well as first suggestions for the treament of these challenges through modifications and expansions of the encoding framework. Starting from these suggestions, the group is planning to work out feature requests and load them onto the TEI projects page on sourceforge.net.

We present the process of constructing a corpus of spoken and written material for Santome, a Portuguese-related creole language spoken on the island of S. Tomé in the Gulf of Guinea (Africa). Since the language lacks an official status, we faced the typical difficulties, such as language variation, lack of standard spelling, lack of basic language instruments, and only a limited data set. The corpus comprises data from the second half of the 19th century until the present. For the corpus compilation we followed corpus linguistics standards and used UTF-8 character encoding and XML to encode meta information. We discuss how we normalized all material to one spelling, how we dealt with cases of language variation, and what type of meta data is used. We also present a POS-tag set developed for the Santome language that will be used to annotate the data with linguistic information.

We present a study of constructions of the type in European Portuguese. We analyse these constructions as complex predicates where both the light verb and the noun share an important role in the predication and we focus on the aspectual combinatory properties of the two elements of the complex predicate. We propose that light verbs inherit the feature specification of the corresponding main verbs, but that the light verbs are underspecified regarding (some of) the event structure features.

This paper discusses a corpus-based investigation of the distribution of the third-person neuter singular pronoun in Dutch (“het”). We labeled all pronominal occurrences of “het” in a large corpus of documents. On the basis of the annotated corpora, we developed an automatic classification system using machine learning techniques to distinguish between the different uses of the neuter pronoun. Although our annotation reveals a completely different distribution of the different uses of the pronoun in Dutch and English, we show that the learning method used for English can be successfully ported to Dutch.

Este trabalho apresenta o processo de construção de um corpus de material oral e escrito do forro (santome), um crioulo de base lexical portuguesa falado na ilha de São Tomé. O corpus compreende dados da segunda metade do século XIX até ao presente. Abordamos as dificuldades típicas de línguas não oficiais que são predominantemente orais, tais como a normalização ortográfica e um conjunto de dados mais restrito. Para a compilação do corpus seguimos padrões linguísticos de corpora e para codificar os metadados utilizámos a codificação de carateres UTF-8 e XML. Definimos um conjunto de metadados e apresentamos as etiquetas desenvolvido para a anotação dos dados com informação linguística. This paper presents the process of building a corpus of spoken and written material of forro (santome), a Portuguese-related creole spoken on the island of S. Tomé (Gulf of Guinea, Africa). The corpus comprises data from the second half of the 19th century until the present. We address the usual difficulties related to non-official languages that are predominantly spoken, such as orthographic normalization and a relatively small data set. For the corpus compilation we followed corpus linguistics standards and used UTF-8 character encoding and XML to encode meta information. We also present a POS-tag set developed for forro that will be used to annotate the data with linguistic information.

We investigate what distinguishes reported dreams from other personal narratives. The continuity hypothesis, stemming from psychological dream analysis work, states that most dreams refer to a person's daily life and personal concerns, similar to other personal narratives such as diary entries. Differences between the two texts may reveal the linguistic markers of dream text, which could be the basis for new dream analysis work and for the automatic detection of dream descriptions. We used three text analytics methods: text classification, topic modeling, and text coherence analysis, and applied these methods to a balanced set of texts representing dreams, diary entries, and other personal stories. We observed that dream texts could be distinguished from other personal narratives nearly perfectly, mostly based on the presence of uncertainty markers and descriptions of scenes. Important markers for non-dream narratives are specific time expressions and conversational expressions. Dream texts also exhibit a lower discourse coherence than other personal narratives.