Dmitry Ustalov — TagBag: Annotating a Foreign Language Lexical Resource with Pictures
2 likes930 views
The document discusses an approach for annotating a foreign language lexical resource with images, highlighting the importance of mapping images to word senses for various applications. It evaluates the efficacy of this annotation method using a dataset, presenting results that indicate moderately high accuracy and identifying common sources of error. The proposed 'tagbag' technique is an unsupervised method that improves mapping quality through visual saliency and considers ontology biases.
Dmitry Ustalov — TagBag: Annotating a Foreign Language Lexical Resource with Pictures
- 1. Annotating a Foreign Language Lexical Resource with Pictures Dmitry Ustalov IMM UB RAS / UrFU Yekaterinburg, Russia
- 3. Introduction •The problem of mapping images to the word senses is quite important: • multimedia search, • text illustration, • quality assessment. •It is also interesting to assess the Yet Another RussNet lexical resource. (Braslavski et al, 2014). 3
- 4. Related Work • PicNet, a proprietary resource (Mihalcea & Leong, 2008). • ImageNet annotates WordNet with pictures & bounding boxes (Deng et al., 2009). • Intersection with WordNet.ru is negligible. • ImageCLEF creates software and datasets for image indexing (Mü̈ller et al., 2010). 4
- 5. Related Work: Flickr •Single-query image retrieval (Reiter et al., 2007). •Semantic Web-based approach (Trojahn et al., 2008). •Wikipedia-based approach (Stampouli et al., 2010). •Flickr tags with visual saliency of images (Jiang et al., 2014). 5
- 6. Problem Given an annotated image I, a bilingual dictionary B, and a lexical resource S, produce a mapping Is. “cat”,“tomcat”,“kitten” → «кот, кошка, котёночек» 6
- 7. TagBag: Assumptions •The most image tags are nouns. •Tags may be polysemous and the redundant tags may be present. • “crane” is «журавль» or «кран»? •The image has a “main” object. 7
- 8. TagBag •Tag. Initialize an empty vector. • Iterate over image tags and retrieve all the translations for each tag. • Add each occurrence to a dimension. •Bag. Prune that vector. • Remove the low frequency dimensions with the cut-off value. • Return the resulting vector. 8
- 10. Evaluation •The present approach is pretty simple. Let’s evaluate it empirically. •Take the top 1500 English nouns and search for Flickr photos. http://www.talkenglish.com/Vocabulary/T op-1500-Nouns.aspx •Get the V.K. Mueller’s dictionary. http://ustalov.imm.uran.ru/pub/mueller.tar.gz 10
- 11. Experimental Setup •Yet Another RussNet (CC BY-SA). http://russianword.net/ •Similarity measures: • cosine similarity, • Jaccard index. •Ask three annotators to submit judgements. 11
- 16. пища, провизия, питание, корм 16https://www.flickr.com/photos/80972943@N00/16396295195
- 17. Results •The accuracy is moderately high and the agreement level is good. •Both measures demonstrate the same performance. 17 http://ustalov.imm.uran.ru/pub/tagbag-aist.tar.gz
- 18. Discussion •Some mappings are the same w.r.t. the similarity measures and 13 of 43 of these mapping are wrong. •Three sources of errors: • sloppy image tags (7 of 13), • actual mapping errors (3 of 13), • batch uploads (3 of 13). 18
- 19. Conclusion •TagBag is an unsupervised approach for mapping images to synsets. •The performance depends both on image tags and ontology bias. •Visual saliency and spam filtering may increase the quality. 19
- 20. Further Work 20
- 21. Thank you! Dmitry Ustalov a post-graduate student @ IMM UB RAS, Yekaterinburg, Russia. https://ustalov.name/ dau@imm.uran.ru The present work is supported by the Russian Foundation for the Humanities, project no. 13-04-12020. 21