Recognizing Textual Entailment in Non-english Text via Automatic Translation into English

We show that a task that typically involves rather deep semantic processing of text—being recognizing textual entailment our case study—can be successfully solved without any tools at all specific for the language of the texts on which the task is performed. Instead, we automatically translate the text into English using a standard machine translation system, and then perform all linguistic processing, including syntactic and semantic levels, using only English language linguistic tools. In this case study we use Italian annotated data. Textual entailment is a relation between two texts. To detect it, we use various measures, which allow us to make entailment decision in the two-way classification task (YES / NO). We set up various heuristics and measures for evaluating the entailment between two texts based on lexical relations. To make entailment judgments, the system applies named entity recognition module, chunking, partof-speech tagging, n-grams, and text similarity modules to both texts, all those modules being for English and not for Italian. Rules have been developed to perform the two-way entailment classification. Our system makes entailment judgments basing on the entailment scores for the text pairs. The system was evaluated on Italian textual entailment data sets: we trained our system on Italian development datasets using the WEKA machine learning toolset and tested it on Italian test data sets. The accuracy of our system on the development corpus is 0.525 and on the test corpus is 0.66, which is a good result given that no Italian-specific linguistic information was used.