1 Modeling Intention in Vitor R. Carvalho Ph.D. Thesis DefenseThesis Committee: Language Technologies Institute William W. Cohen (chair) School of Computer Science Tom M. Mitchel Carnegie Mellon University Robert E. Kraut July 22 th 2008 Lise Getoor (Univ. Maryland)

2 Outline 1. Motivation 2. Acts 3. Leaks 4. Recommending Recipients 5. Learning Robust Rank Models 6. User Study

3 Why  The most successful e-communication application.  Great tool to collaborate, especially in different time zones.  Very cheap, fast, and convenient.  Multiple uses: task manager, contact manager, doc archive, to-do list, etc.  Increasingly popular  Clinton adm. left 32 million s to the National Archives  Bush adm….more than 100 million in 2009 (expected)  Visible impact  Office workers in the U.S. spend at least 25% of the day on – not counting handheld use [Shipley & Schwalbe, 2007]

4 Hard to manage  People get overwhelmed  Costly interruptions  Serious impacts on work productivity  Increasingly difficult to manage requests, negotiate shared tasks and keep track of different commitments  People make horrible mistakes  Send messages to the wrong persons  Forget to address intended recipients  “Oops, Did I just hit reply-to-all?” [Dabbish & Kraut, CSCW-2006]. [Belloti et al. HCI-2005]

5 Thesis ►We present evidence that management can be potentially improved by the effective use of machine learning techniques to model different aspects of user intention.

6 Outline 1. Motivation 2. Acts ◄ 3. Preventing Information Leaks 4. Recommending Recipients 5. Learning Robust Rank Models 6. User Study

7 Classifying into Acts [Cohen, Carvalho & Mitchell, EMNLP-04] Nouns Verbs  An Act is described as a verb- noun pair (e.g., propose meeting, request information) - Not all pairs make sense  One single message may contain multiple acts  Try to describe commonly observed behaviors, rather than all possible speech acts in English  Also include non-linguistic usage of (e.g. delivery of files)

8 Data & Features  Data: Carnegie Mellon MBA students competition  Semester-long project for CMU MBA students. Total of 277 students, divided in 50 teams (4 to 6 students/team). Rich in task negotiation.  messages (from 5 teams) were manually labeled. One of the teams was double labeled, and the inter-annotator agreement ranges from 0.72 to 0.83 (Kappa) for the most frequent acts.  Features  N-grams: 1-gram, 2-gram, 3-gram,4-gram and 5-gram  Pre-Processing  Remove Signature files, quoted lines (in-reply-to) [Jangada package]  Entity normalization and substitution patterns:  “Sunday”…”Monday” →[day], [number]:[number] → [hour],  “me, her, him,us or them” → [me], “after, before, or during” → [time], etc

9 Classification Performance 5-fold cross-validation over s, SVM with linear kernel [Carvalho & Cohen, HLT-ACTS-06] [Cohen, Carvalho & Mitchell, EMNLP-04]

10 Predicting Acts from Surrounding Acts Deliver Request Commit Propose Request Commit Deliver Commit Deliver Strong correlation between previous and next message’s acts Example of Thread Sequence Both Context and Content have predictive value for act classification [Carvalho & Cohen, SIGIR-05] Collective classification problem → Dependency Network

11 Collective Classification with Dependency Networks (DN) In DNs, the full joint probability distribution is approximated with a set of conditional distributions that can be learned independently. The conditional probabilities are calculated for each node given its Markov blanket. Parent Message Child Message Current Message Request Commit Deliver … …… [Heckerman et al., JMLR-00] [Neville & Jensen, JMLR-07] Inference: Temperature-driven Gibbs sampling [Carvalho & Cohen, SIGIR-05]

12 Act by Act Comparative Results Modest improvements over the baseline Only on acts related to negotiation: Request, Commit, Propose, Meet, Commissive, etc. Kappa values with and without collective classification, averaged over four team test sets in the leave-one-team out experiment.

13 Key Ideas Summary Introduced a new taxonomy of acts tailored to communication Good levels of inter-annotator agreement Showed that it can be automated Proposed a collective classification algorithm for threaded messages Related Work:  Speech Act Theory [Austin, 1962;Searle,1969], Coordinator system [Winograd,1987], Dialog Acts for Speech Recognition, Machine Translation, and other dialog-based systems. [Stolcke et al., 2000] [Levin et al., 03], etc.  Related applications: Focus message in threads/discussions [Feng et al, 2006], Action-items discovery [Bennett & Carbonell, 2005], Task-focused summary [Corsten- Oliver et al, 2004], Predicting Social Roles [Leusky, 2004], etc.

14 Applications of Acts Iterative Learning of Tasks and Acts Predicting Social Roles and Group Leadership Detecting Focus on Threaded Discussions Semantically Enhanced Act Taxonomy Refinements [Kushmerick & Khousainov, IJCAI-05] [Leusky,SIGIR-04][Carvalho et al, CEAS-07] [Feng et al., HLT/NAACL-06] [Scerri et al, DEXA-07] [Lampert et al, AAAI ]

15 Outline 1. Motivation 2. Acts 3. Preventing Information Leaks ◄ 4. Recommending Recipients ◄ 5. Learning Robust Rank Models 6. User Study

16

17

18

19 Preventing Info Leaks 1.Similar first or last names, aliases, etc 2.Aggressive auto- completion of addresses 3.Typos 4.Keyboard settings Disastrous consequences: expensive law suits, brand reputation damage, negotiation setbacks, etc. Leak: accidentally sent to wrong person [Carvalho & Cohen, SDM-07]

20 Preventing Info Leaks 1.Similar first or last names, aliases, etc 2.Aggressive auto- completion of addresses 3.Typos 4.Keyboard settings [Carvalho & Cohen, SDM-07] Method 1. Create simulated/artificial recipients 2. Build model for (msg.recipients): train classifier on real data to detect synthetically created outliers (added to the true recipient list). Features: textual(subject, body), network features (frequencies, co- occurrences, etc). 3. Detect potential outliers - Detect outlier and warn user based on confidence.

21 Simulating Leaks Else: Randomly select an address book entry Generate a random address NOT in Address Book  Several options:  Frequent typos, same/similar last names, identical/similar first names, aggressive auto-completion of addresses, etc. We adopted the 3g-address criteria:  On each trial, one of the msg recipients is randomly chosen and an outlier is generated according to:

22 Data and Baselines Enron dataset, with a realistic setting  For each user, ~10% most recent sent messages were used as test  Some basic preprocessing Baseline methods:  Textual similarity  Common baselines in IR Rocchio/TFIDF Centroid [1971] Create a “TfIdf centroid” for each user in Address Book. For testing, rank according to cosine similarity between test msg and each centroid. Knn-30 [Yang & Chute, 1994] Given a test msg, get 30 most similar msgs in training set. Rank according to “sum of similarities” of a given user on the 30-msg set.

23 Using Non-Textual Features 1. Frequency features  Number of received, sent and sent+received messages (from this user) 2. Co-Occurrence Features  Number of times a user co-occurred with all other recipients. 3. Auto features  For each recipient R, find Rm (=address with max score from 3g- address list of R), then use score(R)- score(Rm) as feature. Combine with text-only scores using perceptron-based reranking, trained on simulated leaks Non-textual Features Text-based Feature (KNN30 score or TFIDF score)

24 Leak Results [Carvalho & Cohen, SDM-07]

25 Finding Real Leaks in Enron How can we find it?  Grep for “mistake”, “sorry” or “accident”  Note: must be from one of the Enron users Found 2 valid cases: 1. Message germany-c/sent/930, message has 20 recipients, leak is 2. kitchen-l/sent items/497, it has 44 recipients, leak is Prediction results:  The proposed algorithm was able to find these two leaks “Sorry. Sent this to you by mistake.”, “I accidentally sent you this reminder”

26 Another Addressing Problem Sometimes people just forget an intended recipient

27 Forgetting an intended recipient  Particularly in large organizations,  it is not uncommon to forget to CC an important collaborator: a manager, a colleague, a contractor, an intern, etc.  More frequent than expected (from Enron Collection)  at least 9.27% of the users have forgotten to add a desired recipient.  At least 20.52% of the users were not included as recipients (even though they were intended recipients) in at least one received message.  Cost of errors in task management can be high:  Communication delays, deadlines can be missed  Opportunities wasted, costly misunderstandings, task delays [Carvalho & Cohen, ECIR-2008]

28 Data and Features  Easy to obtain labeled data  Two Ranking problems  Predicting TO+CC+BCC  Predicting CC+BCC  Features & Methods  Textual: Rocchio (TFIDF) and KNN  Non-Textual: Frequency, Recency and Co-Occurrence  Number of messages received and/or sent (from/to this user)  How often was a particular user addressed in the last 100 msgs  Number of times a user co-occurred with all other recipients. Co-occurr means “two recipients were addressed in the same message in the training set”

29 Recipient Recommendation 36 Enron users queries Avg: ~1267 q/user [Carvalho & Cohen, ECIR-08] MRR ~ 0.5

30 Rank Aggregation Many “Data Fusion” methods  2 types: Normalized scores: CombSUM, CombMNZ, etc. Unnormalized scores: BordaCount, Reciprocal Rank Sum, etc. Reciprocal Rank:  The sum of the inverse of the rank of document in each ranking. [Aslam & Montague, 2001]; [Ogilvie & Callan, 2003]; [Macdonald & Ounis, 2006]

31 Rank Aggregation Results

32 Intelligent Auto-completion TOCCBCC CCBCC [Carvalho & Cohen, ECIR-08]

33 Related Work Leak  [Boufaden et al., 2005]: proposed a privacy enforcement system to monitor specific privacy breaches (student names, student grades, IDs).  [Lieberman and Miller, 2007]: Prevent leaks based on faces Recipient Recommendation  [Pal & McCallum, 2006], [Dredze et al, 2008]: CC Prediction problem, Recipient prediction based on summary keywords  Expert Search in [Dom et al.,2003], [Campbell et al,2003], [Balog & de Rijke, 2006], [Balog et al, 2006],[Soboroff, Craswell, de Vries (TREC-Enterprise …)]

34 Outline 1. Motivation 2. Acts 3. Preventing Information Leaks 4. Recommending Recipients 5. Learning Robust Ranking Models ◄ 6. User Study

35 Can we learn a better ranking function? Learning to Rank: machine learning to improve ranking  Many recently proposed methods: RankSVM RankBoost Committee of Perceptrons Meta-Learning Method  Learn Robust Ranking Models in the pairwise-based framework [Elsas, Carvalho & Carbonell, WSDM-08] [Freund et al, 2003] [Joachims, KDD-02]

36 Pairwise-based Ranking Rank q d 1 d 2 d 3 d 4 d 5 d 6... d T We assume a linear function f Goal: induce a ranking function f(d) s.t. Constraints: Paired instance O(n) mislabels produce O( n^2 ) mislabeled pairs

37 Effect of Pairwise Outiers RankSVM SEAL-1

38 Effect of Pairwise Outiers Loss Function RankSVM Pairwise Score P l

39 Effect of Pairwise Outiers RankSVM Loss Function Robust to outliers, but not convex Pairwise Score P l

40 Ranking Models: 2 Stages Base Ranker Sigmoid Rank Non-convex: Final model Base ranking model e.g., RankSVM, Perceptron, ListNet, etc. Minimize (a very close approximation for) the empirical error: number of misranks Robust to outliers (label noise)

41 Learning SigmoidRank Loss Learning with Gradient Descent

42 Recipient Results 36 Enron users 12.7% 1.69%-0.09% 13.2%0.96%2.07% queries Avg: ~1267 q/user p=0.74p<0.01 p=0.06p<0.01 [Carvalho, Elsas, Cohen and Carbonell, SIGIR 2008 LR4IR]

43 Recipient Results [Carvalho, Elsas, Cohen and Carbonell, SIGIR 2008 LR4IR]

44 Recipient Results [Carvalho, Elsas, Cohen and Carbonell, SIGIR 2008 LR4IR]

45 Set Expansion (SEAL) Results [Wang & Cohen, ICDM-2007] [18 features, ~120/60 train/test splits, ~half relevant] [Carvalho et al, SIGIR 2008 LR4IR]

46 Letor Results [ #queries/#features: (106/25) (50/44) (75/44)] [Carvalho et al, SIGIR 2008 LR4IR]

47 Related Work Classification with non-convex loss functions: tradeoff for outlier robustness, accuracy, scalability, etc. [Perez-Cruz et al, 2003], [Xu et al., 2006], [Zhan & Shen, 2005], [Collobert et al, 2006], [Liu et al, 2005], [Yang and Hu, 2008] Ranking with other non-convex loss functions FRank [Tsai et al, 2007]: a fidelity-based loss function optimized in the boosting framework, query normalization may be interfering in performance gains, not a general stage (meta) learner

48 Outline 1. Motivation 2. Acts 3. Preventing Information Leaks 4. Recommending Recipients 5. Learning Robust Ranking Models 6. User Study ◄

49 User Study Choosing an System:  Gmail, Yahoo!Mail, etc. Widely adopted, interface/compatibility issues  Develop a new client Perfect control, longer development, low adoption.  Mozilla Thunderbird Open source community, easy mechanism to install extensions, millions of users.

50 User Study: Cut Once Cut Once, a Mozilla Thunderbird extension for Leak Detection and Recipient Recommendation A few issues …  Poor documentation and limited customization of interface  JavaScript is slow: Imposed computational restrictions Disregard rare words and rare users. Implement two ‘lightweight’ ranking methods:  1) TFIDF  2) MRR (Frequency, Recency, TFIDF) [Balasubramanyan, Carvalho and Cohen, AAAI ]

51 Cut Once Screenshots Main Window after installation

52 Cut Once Screenshots

53 Cut Once Screenshots Logged: Cut Once Usage - Time, Confidence and Position in rank of clicked recommendations - Baseline Ranking Method

54 User Study Description 4 week long study most subjects from Pittsburgh After 1 week, qualified users were invited to continue. 20% of compensation was paid after 1 week After 4 weeks, users were fully compensated (+ final questionnaire) 26 subjects finished study 4 female and 22 male. Median age Total of 2315 sent messages. Averages: 113 address book entries Mostly students. A few sys admin, 1 professor, 2 staff member. Randomly assigned to two different ranking methods: TFIDF and MRR

55 Recipient Suggestions 17 subjects used the functionality (in 5.28% of their sent msgs). Average of 1 accepted suggestion per sent messages.

56 Comparison of Ranking Methods Difference is not statistically significant MRR better than TFIDF Clicked Rank Average Rank: 3.14 versus 3.69 Rank Quality: 3.51 versus 3.43 Rough estimate: factor of *4 = 22 weeks of user study or 5.5*26 = 143 subjects for 4 weeks

57 Results: Leak Detection 18 subj used the leak deletion (in 2.75% of their sent msgs). Most frequent reported use was to “clean up” the addressee list:  Removing unwanted people (inserted by Reply-all)  Remove themselves (automatically added) 5 real leaks were reported, from 4 different users These users did not use Cut Once to remove the leaks  Clicked on the “cancel” button, and removed manually Uncomfortable or unfamiliar with interface Under pressure because of 10-sec timer

58 Results: Leak Detection 5 leaks → 4 users.  Network Admin: two users with similar userIDs  System Admin: wrong auto-completion in 2 or 3 situations  Undergrad: two acquaintances with similar names  Grad student: reply-all case Correlations  2 users used TFIDF, 2 MRR  No significant correlation with size of Address Book or #sent msgs  Correlation with “non-student” occupations (95% confidence) Estimate: 1 leak every 463 sent messages  Assuming a binomial dist with p=5/2315, then 1066 messages are required send at least one leak (with 90% confidence).

59 Final Questionnaire (Higher = Better) Not

60 Frequent complaints Training and Optimization Interface

61 Final Questionnaire Compose-then-address instead of address-then-compose behavior

62 Conclusions acts A taxonomy of intentions in communication Categorization can be automated Addressed Two Types of Addressing Mistakes Leaks (accidentally adding non-intended recipients) Recipient recommendation (forgetting intended recipients)  Framed as a supervised learning problems  Introduced new methods for Leak detection  Proposed several models for recipient recommendation, including combinations of base methods. Proposed a new general purpose ranking algorithm Robust to outliers - outperformed state-of-the-art rankers on the recipient recommendation task (and other ranking tasks) User Study using a Mozilla Thunderbird extension  Caught 5 real leaks, and showed reasonably good prediction quality  Showed clear potential to be adopted by a large number of users

63 Proof: non-existence of a better advisor Given the finite set of good advisors A n

64 Proof: non-existence of a better advisor Q.E.D. Assuming Given the finite set of advisors A n

65 Thank you.