Artificial Intelligence, Intelligence Systems, Neural Networks

Article # 13_2018

31 p.

PDF

Igor Trofimov, Elena Suleymanova, Natalia Vlasova, Alexey Podobryaev
Disambiguation between eventive and non-eventive meaning of nouns

Event extraction is an advanced form of text mining having numerous applications. One of the challenges faced by event extraction systems is the problem of automatic distinguishing between eventive and non-eventive use of ambiguous event nominals. The proposed disambiguation method relies on an automatically generated training set. In order to learn the difference between eventive and non-eventive reading of a target ambiguous nominal, the classifier is trained on two sets of automatically labelled examples featuring unambiguous distributionally similar lexical substitutes for either reading. The method was evaluated on a small sample of 6 ambiguous event-denoting nouns and performed fairly well (77,38% average accuracy, although with more than 20% variation for individual nouns). Suggestions for future work include development of a more advanced distributional model and research towards automated selection of unambiguous substitutes.

Key words: word sense disambiguation, automatic training set generation, distributional semantic model, event, event nominal, event-related information extraction.

article citation

DOI

Article # 26_2018

13 p.

PDF

Alexander Smirnov, Dmitry Stepanov
The use of convolutional neural networks for recognition of the type of premises using special features of the premises

The article proposes a method for recognizing a convolutional neural network (CNN) of the type / class of internal elements of a building layout using specific features of these elements, such as borders or edges. An algorithm for pre-processing of images is proposed to highlight the borders / edges in the image. Also a database is created with images of various types of interior elements of the building layout, such as a corridor, a door (doorway), corner structures and stairs. The article also discusses the own structure of the SNA, and presents data on the accuracy of recognition of various types of premises. The developed method is proposed to use for the primary navigation of mobile robots. (In Russian).

Key words:  convolutional neural network, premises recognition, image filtering, block-parallel processing, feature extraction.

article citation

DOI

Article # 33_2018

18 p.

PDF

Igor Trofimov, Elena Suleymanova
A dependency-based distributional semantic model for identifying taxonomic similarity

Are dependency-based distributional semantic models worth the computational cost and the linguistic resources they require? As our evaluation study suggests, the answer should be "yes" if the task in hand involves distinguishing between feature-based similarity and pure association. After extensive parameter tuning, window-based models still fall behind dependencybased ones when evaluated on our Russian-language similarity/association dataset. (In Russian).

Key words:  distributional semantic model, dependency-based DSM, taxonomic similarity, feature-based similarity, word2vec, skipgram, RuSim1000.

article citation

DOI