Applying Specific Clusterization and Fingerprint Density Distribution with Genetic Algorithm Overall Tuning in External Plagiarism Detection

One of the biggest challenges encountered at PAN'11 External Plagiarism Detection was the need for different clusterization methods for different types of plagiarism within the corpus. The existence of sparse sections of highly obfuscated, low obfuscated and translated plagiarism sections alongside with verbatim plagiarism parts, made single pass clusterization inefficient as it produced negative effects in one of the above cases. At PAN'11 we used a single pass fixed length clusterization algorithm with a fixed value defining the maximum distance for cluster formation. The main issue with the fixed cauterization value is that large numbers (1600-1800) perform best for high obfuscation, medium (900) for translated and low (40) for verbatim sections. We decided to develop the system that will be able to either dynamically adjust the clusterization distance depending on the type of detected sections or try out multi-pass clusterization with different distance value with the exclusion of already detected clusters and heuristic post processing. For each detected cluster in several clusterization runs we measured Diagonal Density Distribution (DDD) and Mean Average Diagonal Fingerprint Distance (MADFD). These two values reflect the relative distribution of detected equal fingerprints within the cluster diagonal and allows to effectively tell which type of plagiarism is actually there. One more important role that these values play is the negation of cluster merging if the resulting DDD is less than any of the two clusters merged. This was particularly effective preventing accidental fingerprints merging the resulting clusters. Additionally we discovered that the total number of parameters that affect the system performance is already large and decided to apply the genetic algorithm in order to tackle the best possible meta values instead of picking them by hand. In PAN'12 prototype application we employed a dot plot visualization with both detected clusters and master clusters overlay that allowed us to efficiently control the training process and to measure the overall progress for each separate document pair.