Modeling Intention in Speech Acts, Information Leaks and User Ranking Methods Vitor R. Carvalho Carnegie Mellon University

2 Other Research Interests More details at Info Extraction & Sequential Learning IJCAI-05 (Stacked Sequential Learning) CEAS-04 (Extract Signatures and Reply Lines in ) Richard Stallman, founder of the Free Software Foundation, declared today… Scalable Online & Stacked Algorithms KDD-06 (Single-pass online learning) NIPS-07 WEML (Online Stacked Graphical Learning) Keywords for Advertising & Implicit Queries WWW-06 (Finding Advertising Keywords in WebPages) CEAS-05 (Implicit Queries for )

3 Outline 1. Motivation 2. Speech Acts  Modeling textual intention in messages 3. Intelligent Addressing  Preventing information leaks  Ranking potential recipients 4. Fine-tuning Ranking Models  Ranking in two optimization steps

4 Why  The most successful e-communication application.  Great tool to collaborate, especially in different time zones.  Very cheap, fast, convenient and robust. It just works.  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]

5 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.  “I accidentally sent that message to the wrong person”  “Oops, I forgot to CC you his final offer”  “Oops, Did I just hit reply-to-all?” [Dabbish & Kraut, CSCW-2006]. [Belloti et al. HCI-2005]

6 Outline 1. Motivation√ 2. Speech Acts  Modeling textual intention in messages 3. Intelligent Addressing  Preventing information leaks  Ranking potential recipients 4. Fine-tuning Ranking Models  Ranking in two optimization steps

7 Example From: Benjamin Han To: Vitor Carvalho Subject: LTI Student Research Symposium Hey Vitor When exactly is the LTI SRS submission deadline? Also, don’t forget to ask Eric about the SRS webpage. Thanks. Ben Request - Information Reminder - Action/Task Prioritize by “intention” Help keep track of your tasks: pending requests, commitments, reminders, answers, etc. Better integration with to-do lists

8 Request Time/date Add Task: follow up on: “request for screen shots” by ___ days before -? 2 “next Wed” (12/5/07) “end of the week” (11/30/07) “Sunday” (12/2/07) - other -

9 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)

10 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

11 Error Rate for Various Acts 5-fold cross-validation over s, SVM with linear kernel [Carvalho & Cohen, HLT-ACTS-06] [Cohen, Carvalho & Mitchell, EMNLP-04]

12 Best features Ciranda : Java package for Speech Act Classification

13 Idea: Predicting Acts from Surrounding Acts Deliver Request Commit Propose Request Commit Deliver Commit Deliver Act has little or no correlation with other acts of same message 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] Context: Collective classification problem

14 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]

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

16 Applications of Acts Iterative Learning of Tasks and Acts Predicting Social Roles and Group Leadership Detecting Focus on Threaded Discussions Automatically Classifying s into Activities [Kushmerick & Khousainov, IJCAI-05] [Leusky,SIGIR-04][Carvalho,Wu & Cohen, CEAS-07] [Feng et al., HLT/NAACL-06] [Dredze, Lau & Kushmerick, IUI-06]

17 Outline 1. Motivation √ 2. Speech Acts √  Modeling textual intention in messages 3. Intelligent Addressing  Preventing information leaks  Ranking potential recipients 4. Fine-tuning User Ranking Models  Ranking in two optimization steps

18

19

20

21

22 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]

23 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.Rank potential outliers - Detect outlier and warn user based on confidence.

24 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] Rec 6 Rec 2 … Rec K Rec 5 Most likely outlier Least likely outlier P(rec t ) P(rec t ) =Probability recipient t is an outlier given “message text and other recipients in the message”.

25 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:

26 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.

27 Enron Data Preprocessing ISI version of Enron –Remove repeated messages and inconsistencies Disambiguate Main Enron addresses –List provided by Corrada-Emmanuel from UMass Bag-of-words –Messages were represented as the union of BOW of body and BOW of subject Some stop words removed Self-addressed messages were removed

28 Leak Results 1 Average Accuracy in 10 trials: On each trial, a different set of outliers is generated Rocc

29 Using Network 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.

30 Using Network 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 Network Features Text-based Feature

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

32 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 good 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”

33 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 good 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: “Sorry. Sent this to you by mistake.”, “I accidentally sent you this reminder” Kitchen-l has 4 unseen addresses out of the 44 recipients, Germany has only one.

34 Precision at rank 1 α = 0

35 Not the only problem when addressing s…

36 Sometimes we just… forget 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]

37 Data and Features  Two Ranking problems:  Predicting TO+CC+BCC  Predicting CC+BCC  Easy to obtain labeled data  Features  Textual: Rocchio (TfIdf) and KNN  Network (from Headers)  Frequency  # messages received and/or sent (from/to this user)  Recency  How often was a particular user addressed in the last 100 msgs  Co-Occurrence  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”

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

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

40 [Carvalho & Cohen, ECIR-08]

41 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]

42 Aggregated Ranking Results [Carvalho & Cohen, ECIR-08]

43 Closer look on CC+BCC Results

44 Leak warnings: hit x to remove recipient Timer: msg is sent after 10sec by default Suggestions: hit + to add Mozilla Thunderbird plug-in (Cut Once)

45 Outline 1. Motivation √ 2. Speech Acts √  Modeling textual intention in messages 3. Intelligent Addressing √  Preventing information leaks  Ranking potential recipients 4. Fine-tuning User Ranking Models  Ranking in two optimization steps

46 Recipient Recommendation 36 Enron users

47 Learning to Rank Can we do better ranking? –Learning to Rank: machine learning to improve ranking –Many recently proposed methods: RankSVM ListNet RankBoost Perceptron Variations –Online, scalable. –Learning to rank using 2 optimization steps Pairwise-based ranking framework [Elsas, Carvalho & Carbonell, WSDM-08] [Joachims, KDD-02] [Cao et al., ICML-07] [Freund et al, 2003]

48 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:

49 Ranking with Perceptrons Nice convergence and mistake bounds –bound on the number of misranks Online, fast and scalable Many variants –Voting, averaging, committee, pocket, etc. –General update rule: –Here: Averaged version of perceptron [Elsas, Carvalho & Carbonell, 2008] [Collins, 2002; Gao et al, 2005]

50 Rank SVM Minimizing number of misranks (hinge loss approx.) Equivalent to maximizing AUC Lowerbound on MAP, MRR, etc. [Joachims, KDD-02], [Herbrich et al, 2000] Equivalent to: [Elsas, Carvalho & Carbonell, WSDM-08]

51 Loss Function

52 Loss Function

53 Loss Function Robust to outliers

54 Fine-tuning Ranking Models 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)

55 Learning SigmoidRank Loss Learning with Gradient Descent

56 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

57 Recipient Results 36 Enron users 26.7% 1.22% 0.67% 15.9% 0.77% 4.51% p<0.01 p=0.55 p<0.01

58 Recipient Results 36 Enron users 8.71% 1.78% -0.2% 12.9% 2.68% 1.53%

59 Recipient Results 36 Enron users MAP values CCBCC task

60 Set Expansion (SEAL) Results [Wang & Cohen, ICDM-2007] [18 features, ~120/60 train/test splits, ~half relevant] 1.76% 0.46% 3.22% 2.68% 0.11% 3.20% 1.94% 0.44% 1.81%

61 Letor Results [Liu et al, SIGIR-LR4IR 2007] [ #queries/#features: (106/25) (50/44) (75/44)] 41.8% 0.38% -0.2% 261% 7.86% 19.5% 21.5% 2.51% 2.02%

62 TOCCBCC Enron: user lokay-m Learning Curve A few steps to convergence - good starting point

63 Sigma Parameter

64 Regularization Parameter TREC3TREC4Ohsumed

65 Conclusions acts Managing/tracking commits, requests... (semi) automatically Preventing User’s Mistakes Leaks (accidentally adding non-intended recipients) Recipient prediction (forgetting intended recipients) Mozilla Thunderbird extension Ranking in two-optimization steps Closer approximation minimizing number of misranks (empirical risk minimization framework) Robust to outliers (when compared to convex losses) Fine-tune any base learner in few steps - good starting point

66 Related Work 1 acts: –Speech Act Theory [Austin, 1962;Searle,1969] – classification: spam, folder, etc. –Dialog Acts for Speech Recognition, Machine Translation, and other dialog-based systems. [Stolcke et al., 2000] [Levin et al., 03] Typically, 1 act per utterance (or sentence) and more fine-grained taxonomies, with larger number of acts. is new domain –Winograd’s Coordinator (1987) users manually annotated with intent. –Related applications: Focus message in threads/discussions [Feng et al, 2006], Action-items discovery [Bennett & Carbonell, 2005], Activity classification [ Dredze et al., 2006], Task-focused summary [Corsten-Oliver et al, 2004], Predicting Social Roles [Leusky, 2004], etc.

67 Related Work 2 Leak –[Boufaden et al., 2005] proposed a privacy enforcement system to monitor specific privacy breaches (student names, student grades, IDs). 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 …)]

68 Related Work 3 Ranking in two-optimization steps –[Perez-Cruz et al, 2003] similar idea for the SVM-classification context (Empirical Risk Minimization) –[Xu, Crammer & Schuurman, 2006][Krause & Singer, 2004][Zhan & Shen, 2005], etc. SVM robust to outliers and label noise –[Collobert et al, 2006], [Liu et al, 2005] convexity tradeoff

69 Thank you.

70 Thank you.

71 Thank you.

72 Inter-Annotator Agreement 443 messages were double labeled Inter-Annotator Agreement Act Kappa Deliver 0.75 Commit 0.72 Request 0.81 Amend 0.83 Meeting 0.82 Propose 0.72

73 Collective Classification Procedure

74 Results in Letor

75 A Taxonomy of “ Acts” Noun ActivityInformation Meeting Logistics Data Opinion Ongoing Activity Data Single Event Meeting Other Short Term Task Other Data Committee Future work: integration with task-oriented clustering Only will consider predicting top-level tasks, not recursive structure

76 Closer look on CC+BCC Results

77 PreProcessing: Substitution SymbolPattern [number]any sequence of numbers [hour][number]:[number] [wwhh]“why, where, who, what, or when” [day]“Monday, Tuesday…or Sunday” [day]“Mon, Tue, … or Sun” [pm]AM, PM, A.M., P.M. [me]“me, her, him,us or them” [person]“I, we, you, he, she or they” [after]“after, before, during” [filetype]“.doc,.pdf, …,.txt,.xls”

78 Precision Recall Curves

79 Content versus Context: both Have Predictive Value Content: Bag of Words features only (using only 1g features) Context: Parent and Child Features only ( table below) 8 MaxEnt classifiers, trained on 3F2 and tested on 1F3 team dataset Only 1 st child message was considered (vast majority – more than 95%) Kappa Values using Relational (Context) features and Textual (Content) features. Parent Boolean Features Child Boolean Features Parent_Request, Parent_Deliver, Parent_Commit, Parent_Propose, Parent_Directive, Parent_Commissive Parent_Meeting, Parent_dData Child_Request, Child_Deliver, Child_Commit, Child_Propose, Child_Directive, Child_Commissive, Child_Meeting, Child_dData Set of Context Features (Relational) Delivery Request Commit Proposal Request ??? Parent messageChild message

80 Evidence of Sequential Correlation of Acts Transition diagram for most common verbs from CSPACE corpus It is NOT a Probabilistic DFA Act sequence patterns: (Request, Deliver+), (Propose, Commit+, Deliver+), (Propose, Deliver+), most common act was Deliver Less regularity than the expected ( considering previous deterministic negotiation state diagrams)

81 Finding Real Leaks in Enron How can we find it? –Look for “mistake”, “sorry” or “accident ”. We were looking for sentences like “Sorry. Sent this to you by mistake. Please disregard.”, “I accidentally sent you this reminder”, etc. How many can we find? –Dozens of cases. Unfortunately, most of these cases were originated by non-Enron addresses or by an Enron address that is not one of the 151 Enron users whose messages were collected. Our method requires a collection of sent (+received) messages from a user. Found 2 real “valid” cases! (“valid” = testable) –Message germanyc/sent/930, message has 20 recipients, leak is –kitchen-l/sent items/497, it has 44 recipients, leak is

82 Finding Real Leaks in Enron –Disappointing Results!! –Reason: and were never observed in the training set! [Accuracy, Average Rank], 100 trials

83 Combining Textual and Network Features 10-fold crossvalidation: textual scores from Knn-30 Training –Turn each train example into |R| binary examples, where |R| is the number of recipients of the message. |R|-1 negative (the real recipients) 1 positive (leak-recipient) –Augment “textual score” with network features –Apply a classification-based reranking scheme (Voted Perceptron algorithm with 5 passes over training set)

84 Some Results: Kitchen-l has 4 unseen addresses out of the 44 recipients, Germany-c has only one, out of 20.

85 Mixture parameter  :

86 Mixture parameter  :

87 Back to the simulated leaks:

88 Office Automation  storm motivates new research in Office Automation: –Adaptive spam filtering [Cormack & Lynam, TOIS-06] – foldering [Klimt & Yang, ECML-04; Brutlag & Meek, ICML-00] –Automatic learning of user’s activities [Surendran et al., CEAS-05; Huang et al., CEAS-04] –Integration with search engines [Goodman & Carvalho, CEAS-05] –Integration with to-do lists [Bennett & Carbonell, SIGIR-05] – contextual search [Minkov & Cohen, SIGIR-06] –Expertise search [Balog & de Rijke, WWW-06]

89 Related Work  Privacy Enforcement System  [Boufaden et al., CEAS-05] - used information extraction techniques and domain knowledge to detect privacy breaches via in a university environment. Breaches: student names, student grades and student IDs. Keyword patterns.  CC Prediction  [Pal & McCallum, CEAS-06] : Prediction of most likely intended recipients of s. Used Naïve Bayes and Factor-graphs models.  Tested on one single user, limited evaluation, private data. Different formulations, but 44%-60% comparison.  Expert finding in  [Balog & de Rijke, WWW-06], [Campbell et al., CIKM-03]  No user focus (single user vs organization), specific queries, etc. Leak Prediction Recipient Prediction

90 Goal  Goal: Automate aspects of exchange that can help users manage work-related s more effectively, potentially improving task management and productivity.  We focus on two aspects related to intention in 1) Textual Intention: Can we make systems aware of the sender’s intentions? How? How well? 2) Intended Recipients: Can we assist the user on selecting the intended recipients for a particular message under composition? How well? Why?