In the modern world, the relevance of using the methods of automatic summarization and annotation is increasing. Currently, there is a problem of information overload. Abstracts and annotations make it possible to understand the main content of the document and determine the need to refer to the original source. Automatic summarization and annotation helps a person to efficiently process large amounts of information.

An abstract is a coherent text that briefly expresses not only the central theme or subject of a document, but also the purpose, the methods used, the main results of the described research or development. Abstracts are usually prepared for scientific and technical documents (scientific books, articles, patents for inventions, etc.). The abstract focuses the reader’s attention on new information and determines the appropriateness of its reference to the original document. It helps a person navigate information flows, quickly select the most valuable and useful information for himself. The process of writing an abstract is called abstracting.

An annotation is a summary of the content of a document that gives a general idea of its topic. Thus, if the abstract briefly acquaints the reader with the essence of the content presented in the document (facts, methods, experiments, etc.), then the abstract performs only a signal function, informing that an article or book has been published on a certain topic. The process of composing an annotation is called annotation.

Abstracts and annotations are secondary documents (primary, or original, documents are books, articles, patents, etc.). In each secondary document, two components of information can be distinguished: substantive and documentary. The first component contains information from the primary source (what the book is about, article is about). The second component is information about the primary document itself (type of document: book, article, etc.; type: printed, handwritten; year of publication; place of publication, etc.). Later we will focus only on the first component of the secondary document.

Abstracting and annotating text is a complex intellectual activity. It is time consuming to compile abstracts or annotations by a person. This leads to the fact that the latest information (especially foreign) reaches scientists, teachers, engineers and other specialists very slowly, which, in turn, leads to repetition in different countries and within one country of the same research, a later application of the latest techniques, technologies, processes. To somehow avoid this, modern computers are used to compile abstracts and annotations.

Compiling an abstract or annotation of text using a computer is called automatic summarization or annotation.

According to the method of constructing the text, the methods of automatic summarization and annotation are divided into two groups: extracting and generating. When using extraction methods, the most important fragments (sentences, paragraphs) are extracted from the source text. In this case, these fragments are not processed, but extracted in the order and form in which they are given in the text. Among the extraction methods, we will consider the following: methods based on machine learning and methods based on graph theory.

The advantage of machine learning-based methods is the convenience of testing a range of important indicators. In his work K.F. Wong examines a combination of various signs of the importance of a sentence: superficial signs (position of the sentence), content (frequency of words). Two algorithms have been developed: a supervised learning algorithm and a partial learning algorithm. Evaluation of the results showed that the best is a combination of superficial and substantive features. In methods based on graph theory, the text is represented as a graph, the nodes of which represent fragments of text (words, sentences, paragraphs), and the edges denote relations between nodes, for example, semantic relations. In the work of L. Plaza, a summarization method is presented, based on the representation of text into concepts with the subsequent transformation of the document and sentences into a graph. The method uses additional resources - the UMLS biomedical thesaurus and the MetaMap program for converting text into concepts from the UMLS thesaurus. The method consists of the following steps: presentation of the document in the form of a graph, clustering of concepts, and selection of proposals. Generative summarization and annotation methods are based on linguistic rules for natural language processing or artificial intelligence techniques.

Generating methods are capable of creating new text not explicitly presented in the text of the original document. The authors of the work describe the creation of annotations for numerical data. Changes in the input data (gas turbine sensor data) are determined, they are represented in symbolic form, the necessary changes are determined, and a text describing these changes is generated. The system requires a data analysis component and a text generation module. To accomplish these tasks, the authors carried out a knowledge collection procedure: a survey of experts on the description of numerical data, the development of an ontology of examples of data description.

The paper investigates the problem of automatic generation of the structure of annotations. The authors note that predicates and predicate phrases have a communicative function - warning the reader about the content of an annotated document by explicit indication (“mentions”, “represents”). The developed algorithm receives as input a set of extracted fragments of sentences and determines how to combine the fragments into an annotation.

At each step, the algorithm selects the most appropriate predicate (phrase) from a predetermined dictionary to be inserted at the beginning of the current fragment. Evaluation of the results showed that the developed algorithm can predict the structure of annotations in more than 60% of cases.

Thus, we can conclude that modern approaches to automatic summarization and annotation differ in the variety of methods used. Materials for an abstract and annotation can be not only texts, but also numerical data.

References

1. Plaza, L., Diaz, A., & Gervas, P. (2008, August). Concept-graph based biomedical automatic summarization using ontologies. In Coling 2008: Proceedings of the 3^rd textgraphs workshop on graph-based algorithms for natural language processing (pp. 53-56).
2. Wong, K. F., Wu, M., & Li, W. (2008, August). Extractive summarization using supervised and semi-supervised learning. In Proceedings of the 22^nd international conference on computational linguistics (Coling 2008) (pp. 985-992).