Dina Wonsever

Contextual Rules for Text Analysis

In this paper we describe a rule-based formalism for the analysis and labelling of texts segments. The rules are contextual rewriting rules with a restricted form of negation. They allow to underspecify text segments not considered relevant to a given task and to base decisions upon context. A parser for these rules is presented and consistence and completeness issues are discussed. Some results of an implementation of this parser with a set of rules oriented to the segmentation of texts in propositions are shown.

Combining Rules and CRF Learning for Opinion Source Identification in Spanish Texts

In this work we present a system for the automatic annotation of opinions in Spanish texts. We focus mainly in the definition of a TFS-style model for the predicates of opinion and their arguments, in the creation of a lexicon of opinion predicates and in two additional variants for identifying the source of opinions. The original system extracts opinions and all its elements (predicate, source, topic and message) based on hand-coded rules, the first variant uses a CRF model for learning the source, assuming that the predicate is already tagged, and the second variant is a combined version, with the result of source recognition via the rule-based system being added as an additional attribute for training the CRF model. We found that this hybrid system performs better than each of the systems evaluated separately. This work involved the construction of several resources for Spanish: a lexicon of opinion predicates, a 13,000 word corpus with whole opinion annotations and a 40,000 word corpus with annotations of opinion predicates and sources.

Coreference resolution between sources of opinions in Spanish texts

This paper presents a system that identifies coreferent elements in Spanish texts, with the purpose of contributing to an opinion extraction project. In particular, the system looks for the actual source when it is not identified in the opinion and solves coreferences between opinion sources in digital media texts in Spanish. The developed system takes as input texts where the opinions have already been identified, and uses syntactic and semantic information to identify the relationships between the different entities, in order to create coreference chains between the sources of the opinions in the text. The algorithm uses a score method to select the correct antecedent between the candidates. It achieved a precision of 82.8% and a recall of 85.6%.

Building a supertagger for Spanish HPSG

Abstract We created a supertagger for the Spanish language aimed at disambiguating the HPSG lexical frames for the verbs, nouns and adjectives in a sentence. The supertagger uses a maximum entropy model and achieves an accuracy of 84.16% over the verb classes, 86.60% over the noun classes and 91.30% over the adjective classes on the test set. The tagset contains 92 verb classes, 27 noun classes and 13 adjective classes extracted from a Spanish HPSG-compatible annotated corpus that was created by automatically transforming the AnCora Spanish corpus. The tags include information about the arguments structure, their syntactic categories and semantic roles. These are important pieces of HPSG style feature structures.

Supertagging for a Statistical HPSG Parser for Spanish

We created a supertagger for the Spanish language aimed at disambiguating the HPSG lexical frames for the verbs in a sentence. The supertagger uses a CRF model and achieves an accuracy of 83.58i¾ź% for the verb classes on the test set. The tagset contains 92 verb classes, extracted from a Spanish HPSG-compatible annotated corpus that was created by automatically transforming the Ancora Spanish corpus. The verb tags include information about the arguments structure and syntactic categories of the arguments, so they can be easily translated into HPSG lexical entries.

Opinion Search in Spanish Written Press

We describe a press reading tool that focuses on sayings or opinions in the news about topics and sources of the reader’s choice. This tool offers a graphical interface through which several sources can be linked given a certain topic and which allows readers to visualize opinions in a timeline and browse through them. This is done over a corpus that includes three Uruguayan written press media and integrates their current and previous editions. The opinion recognizer is mainly rule-based, with rules written in a formalism called Contextual Rules and machine learning methods and regular expressions rules were used during the different stages of the system. The accuracy of the system was evaluated with respect to the information retrieved, showing 76% precision.

The Influence of Syntactic Information on Hedge Scope Detection

In this paper we elaborate over the use of sequential supervised learning methods on the task of hedge cue scope detection. We address the task using a learning methodology that proposes the use of an iterative, error-based approach to improve classification performance. We analyze how the incorporation of syntactic constituent information to the learning and post-processing steps produces a performance improvement of almost twelve points in terms of F-score over previously unseen data.