Humble Opinions Again in february 2013

We have seen the state of the art, there two possible directions
That research in coreference resolution Should follow are the use of more expres-
sive than mention-pairs models to manage the problem, Such as entity-mention,
and the incorporation of new information, Such as world knowledge and discourse coherence. In some cases, This Information can not be Expressed in terms of pairs of mentions. That is, it is information That Involves several mentions Either partial or entities at once. THEREFORE, an experimental approach In This Requires the expressiveness of the combined entity-mention model With The MOST typical features of the mention-pair model.
we dened an approach based on constraint satisfaction That
Represented the problem in a hypergraph and solved it by relaxation labeling,
Reducing coreference resolution to a hypergraph partitioning problem under a set of constraints. Our approach managed mention-pair and entity-mention models at the same time, and was Able to enter new information by adding as many
Necessary as constraints. Furthermore, our approach overcame the Weaknesses of previous Approaches in state-of-the-art systems, Such as linking Contradictions, classications without context, and A Lack of information in Evaluating pairs.
The system Developed, RelaxCor,'ve Achieved state-of-the-art results using only the mention-pair model without new knowledge. Moreover, Experiments with the entity-mention model Showed how the system Is Able to enter knowledge in a constructive way.
In Addition, as Explained in Section, We Have Proposed a method based on the clustering of all positive examples in Which examples are included, while the negative examples similar to the positive MOST ones are kept and the rest are discarded. This method you reduce the number of negative examples positive without losing any information.
Regarding the feature selection function, many works just manually select the MOST informative feature functions and discard the noisy ones. Few Researchers have incorporated an automatic feature selection function process.
We have made a small contribution in This Area by Selecting feature functions through to Hill Climbing process.
The other Contributions include techniques for performance
Such as balance optimization, pruning, and reordering. The balance parameter used was the optimal point tond Between Precision and recall, the pruning process while the computational cost and Reduced Avoids the system performance being dependent on the size of the documents. Were Both techniques included in the development process Facilitated That the optimization of the system for a target measure. The reordering process performance improved by Reducing the number of possible labels Assigned To The Most informative mentions, Which Caused The Most Reliable coreferential relations to be resolved rst.
Experiments to add world knowledge in order Were performed to Improve the coreference resolution performance. Although These experiments did not last Achieve a signicant improvement, the reason Seems to be more related to the type and source of information and its extraction than the approach used
to Incorporate it.