This paper describes research aiming to refine the tags that are assigned automatically to images by an industrial auto-tagging system (Imagga). The present annotation contains English keyword tags proposed by the original auto-tagging algorithms which assign to each image a set of keywords corresponding to shapes and colors that are recognized in the image. After that the system extends the original metadata by adding the keywords of similar images found in external resources. Thus, on the average, 40 tags with various origin and relevance ranking are assigned automatically to each image. In this paper we propose at first to refine the keywords using Natural Language Processing (NLP) post-editing tools: to recognize the morphological variants, derivative tags, synonyms and phrases. Further we propose to use conceptual similarity metrics and to identify tags that might be viewed as semantically redundant. The experiments described here are performed on a subset of images selected from ImageNet.

Olha Kanishcheva, Tetiana Kovalova, Maria Shvedova, Ruprecht von Waldenfels. Proceedings of the Second Ukrainian Natural Language Processing Workshop (UNLP). 2023.

Today, the research activities of scientists and research organizations is estimated through the informetrical methods (bibliometrics) with absolute and relative indicators based on the number of publications and their citations in scientometric databases. The leading authors and research organizations are given grants and awards, become international laureates, get financial support or other dividends. The authors argue that the method of research activities evaluation influences the authors’ behavior and increases plagiarism, rigged citation and academic dishonesty.The fraud practices misrepresent the estimations of research activities quality and the current status of a discipline and publication activities. The problem is analyzed; simultaneous use of bibliometric methods and most advanced scientific methods of modern methods of defining writing styles is suggested as a solution. The authors emphasize that traditionally libraries are meant to educate, to popularize knowledge and to inform. Today, they have to go into non-traditional functions, i.e. to support linguistic research and to build linguistic corpora. This would prevent manipulations when estimating publication activities with bibliometric methods and would promote academic honesty in academic practices.

Basically, the task of Content Based Image Retrieval (CBIR) is, given an image from an user, to find the most similar images among images in the database. It is not an easy task due to the fact that the definition of \"similarity\" is very different between human and computer. This difference is called \"semantic gap\". To filling this gap, many methods try to take into account the user feedback, i.e., running the system in cycle, at each loop to ask the feedback from users on the output, then using it to refine the searching procedure. However, focusing too much on user feedback may be harmful as it shadows the importance of feature design on this task, especially in earlier loops when less information from users are available. Many researchers investigate feature extraction for CBIR since years so it is natural to raise a question: \"Is the feature design is at the top of its performance?\". Recently, the study on Searching Images with MPEG-7 (& MPEG-7-like) Powered Localized dEscriptors (SIMPLE) has shown that the answer is \"No\". Convinced by this work but without ignoring its limitations, in this article, we aim to show that, by carefully fine-tuning a SIMPLE based system applying on deep analyses, the performance can still be improved significantly. At the end, using Gaussian distribution to detect interest regions and using Jensen-Shannon Divergence as the distance measure seems to be the key points to bootstrap the system performance. The significant relative Mean Average Precision improvements on UCID (5--7%) and UKBench (24--36%), the two well-known datasets for CBIR, support this conclusion. However, clustering the final results according to another similarity measure based on manual tags (textual features), we show that the semantic gap remains large. The integration of visual and textual features for CBIR is therefore necessary.