Read Aloud the Text Content

This audio was created by Woord's Text to Speech service by content creators from all around the world.

Text Content or SSML code:

Although these properties bring quality and completeness, we found only a few methods [10]–[13] which exploit only one of the properties. Ordonez et al. [10] use the parent-child relation in ImageNet to find entry-level (natural) names of an object. Mathews et al. [11] proposed methods for predicting basic-level (natural) concepts by a series of classification and ranking tasks. Guillaumin et al.[9] and Chen et al.[12] use weighted neighbors and a tf-idf-like score, respectively to boostrare tags. To the best of our knowledge, newness and generality have not yet been incorporated in any of the existing approaches. In this paper, we consider all four properties. Our idea is to explore knowledge bases to find tags withRNGN properties. A knowledge base (KB) is referred to as a repository of concepts with commonsense features which are in the form of relations. Some of the existing KBs are WordNet [13], ConceptNet [14], and NELL [15]. According to our intuition, appropriate tags can be found in KBs using their relational information. These tags help in identifying different semantic levels of images. In this paper, we try to substantiate our intuition. Most of the commonly used knowledge bases such as WordNet, ConceptNet, WebChild [16], Probase [17] and NELL are built using only textual information. WordNet has organized nouns and adjectives into lexical classes, with distinction between words and word senses. However, nouns and adjectives are not connected by any semantic relation. ConceptNet is a huge collection of commonsense assertions, but the vast majority are instances of generic relations like parent-child, PartOf, ConceptuallyRelatedTo, or DerivedFrom. WordNet and ConceptNet are constructed manually and curated by experts and by crowdsourcing, respectively. Probase extracts parent-child relations automatically from a large corpus of web pages and has a higher number of such relations compared to WordNet. NELL and WebChild auto extract many relations from the web that are similar to the relations in ConceptNet. Webchild identifies relations from the static dump of Google N-gram which has 1.2 billion 5-grams. NELL tends to suffer from confusion and inconsistency as it gathers data from a variety of web pages. Most of the text-based KBs are constructed using a variant quality of text documents or experts and thus have multiple types of relations and representations which cannot be used in describing images. Therefore, representations and relations obtained by these text-based KBs are not consistent with the semantics of images.