Around the Taaable research project
|Around the Taaable research project|
|Participants :|| Amedeo Napoli|
Florence Le Ber
The Taaable project has been originally created as a challenger of the Computer Cooking Contest (ICCBR Conference). A candidate to this contest is a system whose goal is to solve cooking problems on the basis of a recipe book (common to all candidates), where each recipe is a shallow XML document with an important plain text part. The size of the recipe book (about 800 in 2008 and about 1500 in 2009 and in 2010) prevents from a manual indexing of recipes: this indexing is performed using semi-automatic techniques.
The first version of the Taaable system (2008) was the European vice-champion of the contest. The second version (2009) was theWorld vice-champion of the contest. The third version (2010) was theWorld champion: it has won the main challenge and the adaptation challenge . A fourth version for the 2011’s contest is under conception.
The partners of the 2010’s Taaable project are members of Orpailleur and of Score (INRIA projects in Nancy). Beyond its participation to the CCCs, the Taaable project aims at federating various research themes: case-based reasoning, information retrieval, knowledge acquisition and extraction, knowledge representation, minimal change theory, ontology engineering, semantic wikis, text-mining, etc.
A general description of the 2010’s Taaable system can be found in . The most important original features of this version are:
- A module for adapting quantities. In the previous versions of Taaable, only a substitution of ingredient types by other ingredient types was proposed by the system. Now, there is the possibility to adapt the ingredient quantities. In this way, there is a maximum preservation of some features of the global recipe, such as the quantity of sugar, of calories, etc. This implementation is based on a theoretical research published in 2009 .
- A module for adapting recipe preparation texts. Another adaptation that was not studied before this year is the adaptation of the texts that describe the preparations . Such an adaptation module has been implemented, using natural language processing techniques in order to transform recipes in a tree structure whose root is the final dish, whose leaves are the ingredients, and whose internal nodes represent the actions. The adaptation is performed on the tree structure and, thanks to links between the text and the tree, this adaptation has repercussions on the text.
Several theoretical studies have been carried out that should be applied to some future versions of Taaable:
- The representation of preparations in a temporal qualitative algebra .
- An algorithm for adapting cases defined in an expressive description logic  ).
- The study of the relations between rule-based adaptation and adaptation based on belief revision, that enables to incorporate rules in a revision-based adaptation .
- The study of the extension of the domain ontology to make the retrieval step of a case-based reasoning system more accurate  .
The fourth aspect involves text mining within CBR. In the Taaable system, similar cases are searched according to an ontology which is used to progressively refine or generalize a given target problem. The extension the domain ontology is based on the application of FCA on specific resources collected for this purpose. For example, for refining the ingredient hierarchy, a set of actions applied to ingredients are extracted from the text of recipes. The linguistic anaphoras in recipes require the use of a syntactic and dynamic semantic analysis for building a formal representation of a recipe  from which relations between ingredients and actions are extracted. Based on this textual analysis, the formal representation of the recipe can be considered as a tree structure whose root is the desired meal, whose leaves are ingredients, and whose internal nodes correspond to actions. In this way, a textual adaptation process can be defined where adaptation consists in a subtree substitution, i.e. replacing an initial subtree with a final and more accurate subtree. The search of the final subtree is based on FCA which is used to organize recipes w.r.t. their content.