Create a quick coherent summary of a desired lengthStatistical document summarization technologyNTT Communication Science Laboratories

We have proposed a method to represent a document as a discourse tree that describes the dependency relationships, causal relationships, transitive relationships, etc., between sentences and to extract a summary of the rooted subtree that maximizes the sum of sentence importance based on the given length constraints. The summary is a rooted subtree of a discourse tree, so it maintains coherence. In addition, since the extraction of the optimal rooted subtree is formulated as a computationally intractable problem called the "tree knapsack problem," we have also devised an algorithm to solve efficiently.

In the past, the automatic summarization methods would extract a set of sentences to maximize the sum of sentence importance, and thus the generated summaries ignore the relationship between sentences. That is, summaries were incoherent. Such summaries can be difficult to read as text, or be misleading the reader, and so on.

• A method for generating coherent summaries
• A representation of a discourse tree based on the dependency relationships between sentences
• Efficient algorithm for finding an optimal rooted subtree

• Summarizing the news, especially documents with a clear logical structure, such as explanations and editorials

In this technology, documents are represented as discourse trees that we have developed independently. The discourse tree is a tree that describes the dependency relationships between sentences, such as causal relationships and transitive relationships, and places the most important sentence in the document at the root. Then, the rooted subtree that maximizes the sum of the sentence importance determined based on a word importance database is extracted from this tree as a summary. These summaries have the advantage of preserving the dependency relationships between sentences, thus maintaining coherence. The search for the optimal root tree is formulated as a computationally intractable combinatorial optimization problem called the "tree knapsack problem." To solve this problem efficiently, we also proposed a search algorithm using a data structure called ZDD (Zero-suppressed binary decision diagram). In comparison to conventional methods, which summarize the set of sentences maximize the sum of the importance of the sentences while ignoring the relationships between them, we confirmed that both the comprehensiveness of information and the coherence of the sentences are significantly higher.

NTT Communication Science Laboratories - Innovative Communication Laboratory