1 SCUs are sub-sentential content units, not bigger than a clause, taken from a corpus of manually-made summaries. SCUs are used in the evaluation of Summarization tasks
1
T ENTAILS H IF, TYPICALLY, A HUMAN READING T WOULD INFER THAT H IS MOST LIKELY TRUE This definition of entailment is based on (and assumes) common human understanding of language as well as background knowledge; in fact, for textual entailment to hold we require that: TEXT AND KNOWLEDGE ENTAIL H, BUT KNOWLEDGE ALONE CANNOT ENTAIL H This means that H may be entailed by incorporating some prior knowledge that would enable its inference from T, but it should not be entailed by that knowledge alone. In other words, it is not allowed to validate H's truth regardless of T. The example below presents a hypothesis referring to a given topic and the entailing sentences found in the corpus of 10 documents about the same topic: Seven submariners were onboard the AS-28. The Russian military was racing against time early Friday to rescue a small submarine that had become trapped on the seabed with seven crew aboard, the Ria-Novosti news agency reported. All seven aboard the AS28 mini-submarine appeared to be in satisfactory condition, naval spokesman Capt. Igor Dygalo said. It was carrying six sailors and a representative of the company that manufactured it. The seven men on board were said to have as little as 24 hours of air. There are seven crew members aboard the vessel, stranded on the ocean floor at a depth of around 190 meters (623 feet) in a bay off the coast of the Kamchatka peninsula in Russia's Far East region. As it can be seen from the above example, the major difference with respect to the main task is that in the framework of the main exercise, where isolated T/H pairs are given, both Text and Hypothesis are artificially created in a way that they do not contain references to information outside the H/T pair, and hence the context necessary to judge the entailment relation is given by T. Only language and world knowledge are needed within the main task, while reference knowledge is typically not required. In contrast, in the Entailment Search task both Text and Hypothesis are to be interpreted in the context of the corpus, as they rely on explicit and implicit references to entities, events, dates, places, situations, etc. pertaining to the topic. As specifically regards the reference to the time in which a sentence was written - which impacts particularly the interpretation of tenses - it must be taken into account that while T (a sentence in a document) is naturally anchored to the publication date of the document containing it, H is anchored (by construction) to the time at which a summary of all the topic documents was written. The anchor time of the hypotheses related to a given topic can be conventionally fixed as the day after the publication of the last topic document. See, for instance, the following example: European Union officials were in Turkey to observe the trial against Orhan Pamuk. EU officials are now in Turkey to observe Pamuk's trial. 2

上一页下一页