ACM corpus annotation analysis Andrew Rosenberg 2/26/2004
2 Overview Motivation Corpus Description Kappa Shortcomings Kappa Augmentation Classification of messages Corpus annotation analysis Next step: Sharpening method Summary
3 Motivation The ACM corpus annotation raises two problems. –By allowing annotators to assign a message one or two labels, there is no clear way to calculate an annotation statistic. An augmentation to the kappa statistic is proposed –Interannotator reliability is low (K <.3) Annotator reeducation and/or annotation material redesign are most likely necessary. Available annotated data can be used, hypothetically, to improve category assignment.
4 Corpus Description 312 messages exchanged between the Columbia chapter of the ACM. Annotated by 2 annotators with one or two of the following 10 labels –question, answer, broadcast, attachment transmission, planning, planning scheduling, planning-meeting scheduling, action item, technical discussion, social chat
5 Kappa Shortcomings Before running ML procedures, we need confidence in assigning labels to the messages. In order to compute kappa (below) we need to count up the number of agreements. How do you determine agreement with an optional secondary label? –Ignore the secondary label?
6 Kappa Shortcomings (ctd.) Ignoring the secondary label isn’t acceptable for two reasons. –It is inconsistent with the annotation guidelines. –It ignores partial agreements. {a,ba}- singleton matches secondary {ab,ca}- primary matches secondary {ab,cb}- secondary matches secondary {ab,ba}- secondary matches primary, and vice versa Note: The purpose is not to inflate the kappa value, but to accurately assess the data.
7 Kappa Augmentation When a labeler employs a secondary label, consider it as a single annotation divided between two categories Select a value of p, where 0.5≤p≤1.0, based on how heavily to weight the secondary label –Singleton annotations assigned a score of 1.0 –Primary p –Secondary 1-p
Kappa Augmentation example AB 1a,bb,d 2b,aa,b 3bb 4ca,d 5b,cc Annotator labels Judge Aabcd Total Judge Babcd Total Annotation Matrices with p=0.6
9 Kappa Augmentation example (ctd.) abcd Total Agreement matrix Judge Aabcd Total Judge Babcd Total Annotation Matrices
10 Kappa Augmentation example (ctd.) To calculate p(E), use the relative frequencies of each annotators label usage. P(Topic)Judge AJudge BP(A)*P(B) a b c d p(E)=0.312 Kappa is then computed as originally:
11 Classification of messages This augmentation allows us to classify messages based their individual kappa’ values at different values of p. –Class 1: high kappa’ at all values of p. –Class 2: low kappa’ at all values of p. –Class 3: high kappa’ at p = 1.0 –Class 4: high kappa’ at p = 0.5 Note: mathematically kappa’ needn’t be monotonic w.r.t. p, but with 2 annotators it is.
12 Corpus Annotation Analysis Agreement is low at all values of p –K’(p=1.0) = –K’(p=0.5) = Other views of the data will provide some insight into how to revise the annotation scheme. –Category distribution –Category co-occurrence –Category confusion –Class distribution –Category by class distribution
13 Corpus Annotation Analysis: Category Distribution totalgrdb Question Answer Broadcast Attachment Transmission312 Planning Meeting Scheduling Planning Scheduling27225 Planning Action Item19109 Technical Discussion31229 Social Chat36297
14 Corpus Annotation Analysis: Category Co-occurrence QABA.T.P.M.SP.S.P.A.IT.DS.C Questionx Answerxx Broadcastxxx Attachment Transmissionxxxx Planning Meeting Schedulingxxxxx21000 Planning Schedulingxxxxxx0000 Planningxxxxxxx320 Action Itemxxxxxxxx10 Technical Discussionxxxxxxxxx1 Social Chatxxxxxxxxxx
15 Corpus Annotation Analysis: Category Confusion QABA.T.P.M.S.P.SPA.IT.D.S.C. Question Answerx Broadcastxx Attachment Transmissionxxx Planning Meeting Schedulingxxxx Planning Schedulingxxxxx24110 Planningxxxxxx7550 Action Itemxxxxxxx121 Technical Discussionxxxxxxxx21 Social Chatxxxxxxxxx4
16 Corpus Annotation Analysis: Class Distribution Constant High (Class 1): Constant Low (Class 2): Low to High (Class 3): High to Low (Class 4): Total Messages312
17 Corpus Annotation Analysis: Category by Class Distribution-1/2 Num messages Class : Total Question Answer Broadcast Attachment Transmission00 Planning Meeting Scheduling Planning Scheduling Planning Action Item00 Technical Discussion Social Chat Num messages Class : Total Question Answer Broadcast Attachment Transmission31 Planning Meeting Scheduling Planning Scheduling Planning Action Item Technical Discussion Social Chat Class 1:const. highClass 2:const. low
Corpus Annotation Analysis: Category by Class Distribution-2/2 Num messages Class : Total Question Answer Broadcast Attachment Transmission00 Planning Meeting Scheduling Planning Scheduling Planning Action Item Technical Discussion Social Chat Num messages Class : Total Question Answer Broadcast Attachment Transmission00 Planning Meeting Scheduling Planning Scheduling Planning Action Item Technical Discussion Social Chat Class 3:low to highClass 4:high to low
19 Next step: Sharpening method In determining interannotator agreement with kappa, etc., two available pieces of information are overlooked: –Some annotators are “better” than others –Some messages are “easier to label” than others By limiting the contribution of known poor annotators and difficult messages, we gain confidence in the final category assignment of each message. How do we rank annotators? Messages?
20 Sharpening Method (ctd.) Ranking Annotators –Calculate kappa between each annotator and the rest of the group. –“Better” annotators have a higher agreement with the group Ranking messages –Variance (or -p*log(p)) of label vector summed over annotators. –Messages with high variance are more consistently annotated
21 Sharpening Method (ctd.) How do we use these ranks? –Weight the annotators based on their rank. –Recompute the message matrix with weighted annotator contributions. –Weight the messages based on their rank. –Recompute the kappa values with weighted message contributions. –Repeat these steps until the weights change beneath a threshold.
22 Summary The ACM corpus annotation raises two problems. –By allowing annotators to assign a message one or two labels, there is no clear way to calculate an annotation statistic. An augmentation to the kappa statistic is proposed –Interannotator reliability is low (K <.3) Annotator reeducation and/or annotation material redesign are most likely necessary. Available annotated data can be used, hypothetically, to improve category assignment.