Textual Entailment

Have you wondered how a Question-Answering system could detect different variants of the same question and still provide you the same answer?

One of the key challenges faced by NLP systems is the concern of semantic variability. Think about it for a moment. How often can the same expression be expressed in different ways in human languages? Person A might express the same meaning with a text that comprises completely different words than a sentence formed by another Person B.

The secret sauce lies in recognizing textual entailment.

Textual Entailment

Textual Entailment helps in identifying the relationship between two sentences, or in other words, the semantic variability of the expression. The generic task of Recognizing Textual Entailment (RTE) can be specified, given two texts, whether the meaning of one text can be inferred or entailed from the other.

In the RTE domain, we treat one of the given text content as text or premise and the other as hypothesis. The task of the RTE would be to verify the hypothesis entails the premise.

Let us consider the following pair of premises and hypotheses.

"Nikita knows how to swim". "Nikita cleared her swimming test".

In the above example, the inference relation in Entailment is unidirectional, ie, from premise to hypothesis. The relation could be bidirectional as well. Let us consider another example.

"Nikita undertook her swimming test today and cleared it". "Nikita knows how to swim".

RTE finds different applications in various fields. Let us examine a few.

Question Answering System

The Question Answering system aims to provide answers for questions posed in natural languages. As one would expect, questions would have similar meanings despite the difference in the semantics. Similarly, the answer itself could be formulated using different words and expressions compared to the question.

Considering these challenges, the role of RTE in a Question Answering System is to identify texts that entail the expected answer.

Information Extraction

Information Extraction systems focus on identifying relations among the elements in the text. Consider the following examples,

The house is owned by Aami Aami bought a new house

In the given examples, the entities involved are Aami and house. The relation is the person Aami "owns" the house. In the context of Information Extraction, the role of RTE is to differentiate text variants that express the same target relations.

Information Retrieval

While Information Retrieval may sound similar to Information Extraction, the roles executed by Information Retrieval systems are different. The role of the information retrieval systems is to retrieve documents in response to the given search query.

For Information Retrieval Systems, the RTE task ensures a relevant document is retrieved despite the presence or absence of query tokens.

Document Summarization

Document Summarization enables to avoid redundancy in text segment compared to another text segment in the same document. The task of RTE is to omit a redundant sentence or expression from the summary that could be entailed from another expression in the summary.


The different approaches for recognizing textual entailment can be classified into two broad categories based on the representation of the involved premise and hypothesis.

In the first approach, the premise and hypothesis can be represented using lexical methods. The lexical approach works on the strings comparisons. While the approach does consider word overlapping and subsequence mapping, it ignores the semantic relationships between the words.

The alternative approach is using syntactic methods. The syntactical methods focus on representing the involved text fragments using directed graphs. The directed graphs approach allows the text fragments to maintain the relationship. The entailment relationship can be determined by comparing the graph representation of premise and hypothesis.

The entailment relationship is often represented using a numeric value. This enables to portray the degree of entailment, instead of having a true/false value. Running deep learning models on a large corpus enables to discover a lot of features that could be useful in the entailment task.

Last updated