230 likes | 382 Views
Auto-grouping Emails for Faster eDiscovery. Sachindra Joshi , Danish Contractor, Kenney Ng*, Prasad M Deshpande, and Thomas Hampp* IBM Research – India *IBM Software Group. Outline of the Talk. eDiscovery Process A new way of eDiscovery Review: Group Level Review Creating Syntactic Groups
E N D
Auto-grouping Emails for Faster eDiscovery Sachindra Joshi, Danish Contractor, Kenney Ng*, Prasad M Deshpande, and Thomas Hampp* IBM Research – India *IBM Software Group
Outline of the Talk • eDiscovery Process • A new way of eDiscovery Review: Group Level Review • Creating Syntactic Groups • Creating Semantic Groups • Experiments and Conclusion |
eDiscovery Process • Discovery: Process in pre-trial phase • Produce relevant information • eDiscovery: FRCP 2006 amendment • Produce relevant Electronically Stored Information (ESI) • Emails, chats, word docs, presentations etc. • Huge volumes of ESI - Process is expensive • 60% of cases warrant some form of eDisovery • 4.8 billion dollars industry in 2011 |
eDiscovery Process • High cost due to review stage • Lawsuit between Clinton administration and tobacco companies (U.S. Vs. Philip Morris) Apply Text Mining Techniques to reduce high costs involved in eDiscovery Process |
Named entity annotator Language Annotator Signature Annotator Architecture of eDiscovery Review Systems |
Group Level Review • Review groups of documents that are “related” instead of individual documents • Mark whole group as responsive/unresponsive or privileged • Efficient and consistent • Syntactically Similar Documents • Automated messages, Near and exact duplicates • Semantically Similar Documents • Threads, semantic categories |
Detecting Near Duplicates • S1: I am away from 17/2/2011 to 19/2/2011. Please mail xyz@in.ibm.com in case of any need • S2: I am away from 26/7/2011 to 31/7/2011. Please mail abc@us.ibm.com in case of any need • Notion of Similarity: Resemblance • Use fingerprinting (Rabin) instead of actual chunks. |
Efficient Detection of Near Duplicates • For a document of length n words there would be • n-K+1 chunks with a window size of K • It suffices to keep for each document a relatively small fixed size signature • Let Sn be the set of permutations of [n] • And let P be chosen uniformly at random over Sn |
Signature Annotator • In practice choosing the permutations randomly is hard • Use a set of n one-to-one functions fi and keep only the smallest value for each fi • Keep only j lowest significant bits for each value |
Discovering Automated Messages • Generating groups of near duplicate – Index Based Clustering • For each document d in index I do • If d is not covered • Let S = {S1, S2, …, Sn} be the signature of document d • D = Query(I, atleast(S,k)) • For each document d’ in D • d’ is covered • Discovering Groups of Automated Messages • Automated Messages, Group of bulk emails, Group of forward emails • Use MD5 to detect bulk emails. Emails with one segment are automated messages |
Detecting Semantic Groups: Email Threads • A tree like structure • A link denotes that the child node was written as a reply to the parent node. • Capture the context in which an email was written |
Detecting Email Threads • Meta data based methods • Headers are not consistently used • Content of old mail remains in the new mail • A segment contains text of only one communication • An email ei contains ej iff ei approximately contains all the segment of ej |
Method for Thread Detection • Email Segment Generator (ESG) • Creates segments of it where each segment contains content of only one email. • Segment Signature Generator (SSG): • Generates a signature for a segment • Use near duplicate signatures • For practical implementation, we limit on the number of segment signatures (N) that can be associated with an email, e.g. 20 segments.
Word index Meta index ESG SSG Signature index w1 w2 wn Method: Processing at Indexing Time
Word index Meta index Signature index w1 w2 wn Method: Processing at Query Time q Use Signature Of First Segment Generating Candidate Thread Set
Detecting Email Threads • Given a Candidate Thread Set • Identify the email with only root segment • An email ec is child of an email ep if ecminimally contains ep |
Creating Semantic Categories • Focus Categories • Documents that are likely to be responsive • Legal Content, Financial Communication, Intellectual Property • High recall • Filter Categories • Documents that are likely to be unresponsive • Bulk emails, Private communication, Jokes • High precision |
Creating Semantic Categories • Email Segmentation • Pattern based annotation: Use System T based method • Consolidation • Each concept is independent • Apply additional constraints over concepts |
Experiments – Near Duplicate Detection • Enron Corpus • 517K emails from 150 users • Measuring precision • Manually evaluated near duplicate set for 500 queries • With more bits precision is 100% even with 40% similarity threshold • Only 33.3 % emails are unique |
Experiments – Email Thread Detection • No ground truth for threads • Subject approximation Method: Based on “Re:”, “Fw:” etc in subject • Manually verified the results of thread for our method and subject approximation method • The union of correct emails in thread for both approaches is treated as ground truth. |
Experiments – Semantic Group • Ground truth: Sampled 2200 emails using generic keywords and then manually labeled |
Conclusions • We developed a framework that allow group level review of documents • We developed methods for finding syntactic groups such as automated messages for creating groups • We developed methods for finding email threads and semantic groups • We showed significant reduction in the review time by using the group level review and integrated the proposed techniques with IBM Infosphere eDiscovery Analyzer product |