Cover Page

B.3.2.b Work Plan

We consider GALATEAS as a middle sized software integration project with many innovative aspects. Therefore, since preliminary phases, it is important to spell out all dependencies among tasks, in order to achieve a harmonic development process and a timely deliverable of all planned artefacts. For these reasons we decided to plan time-based dependency on tasks rather than work-packages: the idea is that WPs are mostly thematic management units, whereas tasks are goal driven activities. Work packages are naturally built on set of tasks.

The following GANTT presents an overview of the project (Task identifiers are explained in the relative work package):

The task on system design (T6.1) provides input for nearly all tasks. The first technical task to be initiated will concern language resources which must provide input to algorithm tuning. The only task which is considered as an exception is Named Entity Extraction (T2.2) which spans over a longer period: as it is considered a major innovation point, it is likely that it will need waterfall integration with all other project modules.

Log Analysis tasks (T3.1 and T3.2) can proceed in parallel, as they do not depend on linguistic analysis, and they are just considered as predecessor of the first milestone, i.e. the availability of the LangLog service (id 23, in the diagram). Also two of the algorithms optimized in WP4 are predecessor of this milestone, namely Clustering and topic computation. Once the first milestone is achieved (after one year since project start) evaluation activities can start, especially the one concerning clustering and topic computation.

If the project achieves the first milestone with no major delay, than all the activities connected with the delivery of the query based machine translation system (QueryTrans) can start as well. These presuppose the availability of logs, and the fine tuning of the TLike algorithm (T4.1). The reason why the WP on Machine Translation actually starts before the milestone is that the consortium has already enough query data to start tuning the MOSES system in order to understand which are the parameters which will need to be set in order to reach optimal performance.

The goal of the second year of activities is therefore to deliver the first version of the QueryTrans service at the second year of the project (second milestone). From that time on, GALATEAS activities will be focussed on:

• 
  Improving the service by repeated cycles of evaluation and tuning on real data;
  
• 
  Customer acquisition;
  
• 
  Machine translation tuning on newly acquired customers;
  
• 
  Optimisation of the business model;
  
• 
  Dissemination activities, specially concerned with non academic dissemination.
  

Logical dependencies among principal tasks are described by the following PERT diagram:

. Performance Monitoring Table

Indicator No.	Objective/expected result	Indicator name	Expected Progress
			Year 1	Year 2	Year 3
1	To gather language resources on which different algorithms and MT can rely	Linguistic coverage	7 languages, 4 language families
2	To gather as much query log files as possible. Abundance of query log is a quality factor for query machine translation	Log availability	10M queries	20M queries	40M queries
3	Customize several algorithms to deal with user queries. They are crucial for both the LangLog and QueryTrans services	Algorithm tuning	70% (F)Topic Classification 80% (F) TLike	80%(F)Topic Classification 90% (F) TLike
4	Train the Machine translation system in order to handle the biggest number of query pairs	Machine Translation	English French Italian pairs	German and Spanish	Remaining languages with transitive translation for cases with less than 50.000 query pairs
5	To perform integration of available components and run and maintain LangLog and QueryTrans	Optimisation	LangLog	QueryTrans	Maintenance
6	To attract customers to subscribe in free or paying modality GALATEAS services	Sustainability	Availability of the LangLog Service	10 customers for LangLog	20 Customers for LangLog and 10 Customer for QueryTrans