In recent years, the so called "Bag of Words" (BoW) representation has drastically grown in importance in the Pattern Recognition research area: in such representation, the object is characterized by the repetition of basic, "constituting" elements called words. By assuming that all possible words are stored in a dictionary, the bag of words vector for one particular object is obtained by counting the number of times each element of the dictionary occurs in the object. This paradigm has been introduced in the text processing community, where all its ingredients have a clear semantic meaning (words, documents, dictionary). After that, it has been exported to an impressively large set of scenarios becoming a standard reference representation.
In some cases, however, its usage has been extended too far, having been applied "as it is" in situations where not all the concepts/tools of this paradigm have a meaningful role.
For example, there are many scenarios where it is impossible to define natural words/dictionaries, or where such definitions fail in properly capturing all the facets of the data. In other cases, the straightforward application of the bag of words paradigm is too simple to capture the complexity of the problem, or can spoil all the information contained in the data.
Against this background, in this project we aim at critically review and analyse the bag of words representation paradigm, describing possible structural failures and proposing alternative solutions.
We will pay particular attention to the fundamental assumptions and definitions of such paradigm (like definition of words, the counting procedure, the definition of dictionary); moreover, significant efforts will be devoted to the formal management of the uncertainty, which models the fact that some words could be more expressive, or general than others, deserving novel BoW modelling methodologies, absent so far in the literature.