Three Issues in Cross-Language Frame Information Transfer

International Conference RANLP 2009 - Borovets, Bulgaria, pages 441–448 Three issues in cross-language frame information transfer Sara Tonelli, Emanuele Pianta FBK-Irst Via Sommarive 18 I-38100 Povo (Trento), Italy {satonelli,pianta}@fbk.eu Abstract In this paper we address the task of transfer- ring FrameNet annotations from an English cor- pus to an aligned Italian corpus. Experiments were carried out on an English-Italian bitext ex- tracted from the Europarl corpus and on a set of selected sentences from the English FrameNet corpus that have been manually translated into Italian. Our research activity is aimed at an- swering the following three questions: (1) What is the best annotation transfer algorithm for the English-Italian couple? (2) What kind of parallel corpus is best suitable to the annotation transfer task? (3) How should the annotation transfer be evaluated, given the final aim of the transfer? Keywords Frame semantics, cross-language annotation transfer, automatic development of lexical resources. 1 Introduction In recent years, the creation of annotated lexical re- sources has become crucial to the development of text processing systems, especially to train supervised learning systems and evaluate unsupervised or hand- crafted systems. The FrameNet database [8] clearly exemplifies this trend. This resource contains more than 135,000 annotated sentences pointing to more than 10,000 lexical units, with a rich repository of se- mantic roles (the frame elements) and almost 900 situ- ation descriptions (the frames). FrameNet has proved to be useful in a number of NLP tasks, from textual entailment [3] to question answering [16], and the de- velopment of systems for frame recognition has become a topic of great interest for the NLP community, with a devoted task at the last SemEval workshop1. Given the success of the English FrameNet initia- tive, many researchers have focused on the develop- ment of FrameNet-like resources for other languages through manual annotation, for example [4] for Ger- man and [17] for Spanish. Manual annotation guaran- tees high accuracy but requires trained annotators and is expensive and time-consuming. For this reason, a second approach has been investigated, which is based on the automatic projection of frame information from 1 http://framenet.icsi.berkeley.edu/semeval/FSSE.html English texts into a new language using bilingual par- allel corpora and possibly carrying out automatic an- notation of frame information on the English side. If no parallel corpora are available, manually translating an annotated English corpus and automatically trans- ferring the annotations may represent a reliable al- ternative to hand-labeling a new corpus from scratch, given that translators are more easily available than linguistic annotators, particularly for complex tasks like frame annotation. In this paper we explore the possibility to develop a FrameNet database for Italian using transfer methodologies, and discuss the advan- tages of using a manually translated parallel corpus for the transfer task. Besides, we present and discuss two existing evaluation frameworks and propose a new evaluation approach. More specifically we try to an- swer the 3 following questions: (1) What is the best annotation transfer algorithm for the English-Italian couple? (2) What kind of parallel corpus is best …

Content truncated to 500 words