Dear all,
I am looking for an unsupervised approaches to identify the category of each Tag. In other terms, there is some approaches that use taxonomies or thesaurus to categorize Tags.
The goal is to classify a list of generated Tags into a set of different categories. Most of generated content are extracted from audio and video contents.
Topic Modelling may be of interest to you as it does not require any linguistic modelling and is unsupervised although it is mainly used to generate themes (that often require interpretation) rather than 'categories' per se. There is a tool I have used as a researcher which you may find useful:
http://code.google.com/p/topic-modeling-tool/
References are (Blei et al 2003), (Balagopalan 2011) and (McCallum 2002).
There are many other OpenSource clustering tools and algorithms out there that may be better.
If you have an existing set of categories, then you can use semi-supervised learning by providing labelled training data. So for each 'category' practioner heuristics seem to indicate you need around 50 'documents' as clean examples for each category. Then tools can build statistical models (clues) which you can apply to the rest of your dataset, sample to see how accurate the model is and iterate. This seems to work well when you have a small number of broad categories and good examples for each category. If you have a detailed deep domain taxonomy (including synonyms) then you could apply very simple look-up inference (if A then B) as you parse the text to for togging, just watch out for polysemy where you may need to disambiguate meanings if a term in your taxonomy is mentioned in your text collection in more than one meaning.
Thank you so much for your quick and clear answer. Up to now, we don't have an already taxonomies.
To this end, we thought to try already taxonomies used in literature (Such as ODP) and see what we got using theses taxonomies. Thus, have you an idea about existing approaches using such taxonomies on categorization ?
Our main goal is to calssify a set of generated tags on different categories and sub-categories. We don't have up to now a clear idea which strategy we will adopt exactly. We prefer to use taxonomies from literature. It is also to note that generated tags are mostly generated from audio and video contents (Film, Tv programs...). We are planning to try better categorization methods in literature in our system. We prefer unsupervised or semi-unsupervised approaches than supervised approaches. However, If the results are not good with such approaches we will move to supervised approaches..
I would give a look at http://digitalhumanities.org/answers/topic/what-is-the-easiest-tool-for-topic-modelling and in particular to WEKA tool.
- Badges
- Science topic
- Similar topics
- Physical Sciences
- Materials Science
- Materials
- Inorganic Materials
- Metals
- Alloys
- Steel