Textual entailment
Under Textual Entailment (textual Close, short TE ) is understood in the Computational Linguistics ( Natural language processing ) modeling inference relations in the field of natural language. These inferential relationships are expressed in the form of binary relationships between two textual units (e.g. sentences), which are referred to as text or hypothesis .
Textual entailment deliberately defines the concept of inference at the level of natural language and through human judgment. Textual reasoning also includes inference relationships that are not mandatory, but that a person would draw because they are very plausible . A hypothesis follows from a text if a reader of the text would assume with a high probability that the hypothesis is true.
| example | |
| Text (T) | As a result of climate change, more and more ice is melting in the Arctic. |
| Hypothesis (H) | The Arctic is getting warmer. |
Here there is a textual inferential relationship between text and hypothesis. Assuming that the text (T) is true, the hypothesis (H) is extremely plausible. It does not matter that there could be circumstances in which T is true and H is false (e.g. if the ice melts from exposure to the sun, but the water becomes colder on average).
Advantages in formal semantics
Classical research in the field of formal semantics is generally representation-centered. Appropriate formal languages for meaning representation are developed, and relationships between textual units are defined through relationships of their formal representations.
However, textual entailment is not limited to specific strategies. Any algorithm that takes two textual units T and H as input and decides whether H follows from T can be used as a textal entailment method. The spectrum ranges from very flat, low-knowledge processes to full logic-based semantic construction. Textual entailment therefore also represents an interesting scenario for the evaluation of semantic processing methods.
The great attraction of textual entailment for machine language processing lies in the fact that the semantic processing of many applications can at least for the most part be traced back to the decision of textual inference relationships. An example is evaluating the output of machine translation systems against a reference translation. The output is complete and correct if it is followed by the reference and vice versa.
Natural language variability
One of the characteristics of natural language is that there are many different ways of saying what you want to say. Several meanings can be contained in a single text and the same meanings can be expressed by different texts. This variability in semantic expression can be seen as a dual problem of ambiguity language.
The interpretation of a text would, in theory, require a thorough semantic interpretation on a logic-based representation of its meanings. As a practical solution for processing natural languages, textual entailment tries not to go into depth and to interpret it in a rather "simple" way.
- Prime Minister gives speech on the development of the economy.
- Federal Chancellor addresses the economic situation.
- Merkel talks about economic prospects.
Examples
Textual entailment can be represented in three different relationships:
| species | text | hypothesis |
|---|---|---|
| Positive TE (text is hypothesized): | Help the needy, God will reward you. | Giving the money to poor people has positive consequences. |
| Negative TE (text against hypothesis): | Help the needy, God will reward you. | Giving the money to poor people has no consequences. |
| No TE (text not connected nor reflected): | Help the needy, God will reward you. | Giving the money to poor people will make you a better person. |
Practical applications
Software systems that use language understanding often also cover types of textual entailment. For example, cognitive search engines (or meaningful search engines) use textual entailments; this will also find a search hit in which only one of its textual entailments matches the search query.
literature
- Roger Chaffin: The concept of a semantic relation. In: Adrienne Lehrer a. a. (Ed.): Frames, Fields and contrasts. New essays in semantic and lexical organization. Erlbaum, Hillsdale, NJ 1992, ISBN 0-8058-1089-7 , pp. 253-288.
- Hermann Helbig : The semantic structure of natural language. Knowledge representation with MultiNet. Springer, Heidelberg 2001, ISBN 3-540-67784-4 .
- Ido Dagan, Dan Roth, Mark Sammons: Recognizing Textual Entailment: Models and Applications (= Synthesis Lectures on Human Language Technologies). Morgan & Claypool Publishers, 2013, ISBN 978-1-59829-834-5 .
Web links
- Peter Kolb - HS Textual Entailment - Universität Potsdam - ling.uni-potsdam.de, accessed on February 1, 2015
- VL Introduction to Computational Linguistics, Representations and Processes in Language Processing - coli.uni-saarland.de, accessed on February 1, 2015
Individual evidence
- ^ Department of Computer Science and Engineering, University of North Texas, USA: Textual Entailment as a Directional Relation (PDF; 226 KB), 2007.
- ^ Ion Androutsopoulos, Prodromos Malakasiotis: Department of Informatics Athens University of Economics and Business - A Survey of Paraphrasing and Textual Entailment Methods (PDF; 648 KB), 2010.
- ↑ Aljoscha Burchardt: Saarbrücken University - Modeling Textual Entailment with Role-Semantic Information (PDF; 4.98 MB), 2008.
- ^ Günter Neumann: Textual Entailment - Methods and Applications (PDF; 572 KB), 2010.
- ↑ Peter Kolb: University of Potsdam - HS Textual Entailment - Overview (PDF; 134 KB), 2008.
- ↑ Ido Dagan, Oren Glickman, Bernardo Magnini: The PASCAL Recognizing Textual Entailment Challenge (PDF; 295 KB), 2005.