Multilingual Vandalism Detection using Language-Independent & Ex Post Facto Evidence Notebook for PAN at CLEF 2011 Andrew G. West and Insup Lee Dept. of Computer and Information Science University of Pennsylvania - Philadelphia, PA {westand, lee}@cis.upenn.edu Abstract There is much literature on Wikipedia vandalism detection. However, this writing addresses two facets given little treatment to date. First, prior efforts emphasize zero-delay detection, classifying edits the moment they are made. If classification can be delayed (e.g., compiling offline distributions), it is possible to leverage ex post facto evidence. This work describes/evaluates several features of this type, which we find to be overwhelmingly strong vandalism indicators. Second, English Wikipedia has been the primary test-bed for research. Yet, Wikipedia has 200+ language editions and use of localized features impairs porta- bility. This work implements an extensive set of language-independent indicators and evaluates them using three corpora (German, English, Spanish). The work then extends to include language-specific signals. Quantifying their performance benefit, we find that such features can moderately increase classifier accuracy, but significant effort and language fluency are required to capture this utility. Aside from these novel aspects, this effort also broadly addresses the task, implementing 65 total features. Evaluation produces 0.840 PR-AUC on the zero- delay task and 0.906 PR-AUC with ex post facto evidence (averaging languages). Performance matches the state-of-the-art (English), sets novel baselines (German, Spanish), and is validated by a first-place finish over the 2011 PAN-CLEF test set. 1 Introduction Unconstructive or ill-intentioned edits (i.e., vandalism) on Wikipedia erode the encyclo- pedia’s reputation and waste the utility of those who must locate/remove the damage. Moreover, while Wikipedia is the focus of this work, these are issues that affect all wiki environments and collaborative software [9]. Classifiers capable of detecting vandalism can mitigate these issues by autonomously undoing poor edits or prioritizing human ef- forts in locating them. Numerous proposals have addressed this need, as well surveyed in [2,6,9]. These techniques span multiple domains, including natural language pro- cessing (NLP), reputation algorithms, and metadata analysis. Recently, our own prior work [2] combined the leading approaches from these domains to establish a new per- formance baseline; our technique herein borrows heavily from that effort. The 2011 edition of the PAN-CLEF vandalism detection competition, however, has slightly redefined the task relative to the 2010 competition [6] and the bulk of existing anti-vandalism research. In particular, two differences have motivated novel analysis and feature development. First, the prior edition permitted only zero-delay features: an edit simultaneously committed and evaluated at time tn can only leverage information from time t ≤ tn . However, if evaluation can be delayed until time tn+m , it is possi- ble to use ex post facto evidence from the tn < t ≤ tn+m interval to aid predictive efforts. While such features are not relevant for “gate-keeping,” they still have applica- tions. For example, the presence of vandalism would severely undermine static content distributions like the Wikipedia 1.0 project1 , which targets educational settings. This work describes/evaluates several ex post facto features and finds them to be very strong vandalism predictors. The second redefinition is that this year’s corpus contains edits from three lan- guages: German, English, and Spanish. Prior research, however, has been conducted almost exclusively in English, and the 2010 PAN-CLEF winning approach heavily uti- lized English-specific dictionaries [6,8]. Such techniques do not lend themselves to portability across Wikipedia’s 200+ language editions, motivating the use of language- independent features. While these are capable of covering much of the problem space, we find the addition of language-specific features still moderately improves classifier performance. Orthogonal to the issue of portability, we also use the multiple corpora to examine the consistency of feature performance across language versions. While discussion concentrates on these novel aspects, we also implement a breadth of features (65 in total). Performance measures, as detailed in Sec. 3.2, vary based on language and task. The complete feature set produces cross-validation results consistent with the state-of-the-art for English [2] and establishes novel performance benchmarks for Spanish and German (PR-AUC=0.91, weighing languages equally). Though perfor- mance varied considerably over the label-withheld PAN-CLEF 2011 test set, our ap- proach took first-place in the associated competition, reinforcing its status as the most accurate known approach to vandalism classification. 2 Feature Set This section describes the features implemented. Discussion begins with a core feature- set that is both zero-delay and language independent (Sec. 2.1). Then, two extensions to that set are handled: ex post facto (Sec. 2.2) and language-specific (Sec. 2.3). Any fea- ture which cannot be calculated directly from the provided corpus utilizes the Wikipedia API2 . Readers should consult cited works to learn about the algorithms and parameters of complex features (i.e., reputations and lower-order classifiers). 2.1 Zero-Delay, Language-Independent Features Tab. 1 presents features that are: (1) zero-delay and (2) language-independent. Note that features utilizing standardized language localization are included in this category (e.g., “User Talk” in English, is “Benutzer Diskussion” in German). Nearly all of these features have been described in prior work [2,6], so their discus- sion is abbreviated here. Even so, these signals are fundamental to our overall approach, given that a single implementation is portable across all language versions. This is pre- cisely why an extensive quantity of these features have been encoded. 1 http://en.wikipedia.org/wiki/Wikipedia:1.0 2 http://en.wikipedia.org/w/api.php FEATURE DESCRIPTION USR_IS_IP Whether the editor is anonymous/IP, or a registered editor USR_IS_BOT Whether the editor has the “bot” flag (i.e., non-human user) USR_AGE Time, in seconds, since the editor’s first ever edit USR_BLK_BEFORE Whether the editor has been blocked at any point in the past USR_PG_SIZE Size, in bytes, of the editor’s “user talk” page USR_PG_WARNS Quantity of vandalism warnings on editor’s “user talk” (EN only) USR_EDITS_* Editor’s revisions in last, t ∈ {hour, day, week, month, ever} USR_EDITS_DENSE Normalizing USR_EDITS_EVER by USR_AGE USR_REP Editor reputation capturing vandalism tendencies [10] (EN only) USR_COUNTRY_REP Reputation for editor’s geo-located country of origin [10] (EN only) USR_HAS_RB Whether the editor has ever been caught vandalizing [10] (EN only) USR_LAST_RB Time, in seconds, since editor last vandalized [10] (EN only) ART_AGE Time, in seconds, since the edited article was created ART_EDITS_* Article revisions in last, t ∈ {hour, day, week, month, ever} ART_EDITS_DENSE Normalizing ART_EDITS_EVER by ART_AGE ART_SIZE Size, in bytes, of article after the edit under inspection was made ART_SIZE_DELT Difference in article size, in bytes, as a result of the edit ART_CHURN_CHARS Quantity of characters added or removed by edit ART_CHURN_BLKS Quantity of non-adjacent text blocks modified by edit ART_REP Article reputation, capturing vandalism tendencies [10] (EN only) TIME_TOD Time-of-day at which edit was committed (UTC locale) TIME_DOW Day-of-week on which edit was committed (UTC locale) COMM_LEN Length, in characters, of the “revision comment” left with the edit COMM_HAS_SEC Whether the comment indicates the edit was “section-specific” COMM_LEN_NO_SEC Length, in chars., of the comment w/o auto-added section header COMM_IND_VAND Whether the comment is one typical of vandalism removal WT_NO_DELAY WikiTrust [1] score w/o ex post facto evidence (DE, EN only) PREV_TIME_AGO Time, in seconds, since the article was last revised PREV_USR_IP Whether the previous editor of the article was IP/anonymous PREV_USR_SAME Whether the previous article editor is same as current editor LANG_CHAR_REP Size, in chars., of longest single-character repetition added by edit LANG_UCASE Percent of text added which is in upper-case font LANG_ALPHA Percent of text added which is alphabetic (vs. numeric/symbolic) LANG_LONG_TOK Size, in chars., of longest added token (per word boundaries) LANG_MARKUP Measure of the addition/removal of wiki syntax/markup Table 1. Zero-delay, language-independent features. Some features are not calculated for all lan- guages. These are not fundamental limitations, rather, the source APIs are yet to extend support (but trivially could). See Sec. 2.3 for discussion regarding features of the “LANG_*” form. 2.2 Leveraging Ex Post Facto Evidence More novel is the utilization of ex post facto data in the classification task. To the best of our knowledge, only the WikiTrust system of Adler et al. [1,2] has previously described features of this type. Tab. 2 lists the ex post facto signals implemented in our approach, which includes our own novel contributions (the first 4 features), as well as those proposed and calculated by Adler et al. (the remainder). EX POST FEAT. DESCRIPTION USR_BLK_EVER Whether the editor has ever been blocked on the wiki USR_PG_SZ_DELT Size change of “user talk” page between edit time and +1 hour ART_DIVERSITY Percentage of recent revisions (±10 edits) made by editor HASH_REVERT Whether article content hash-codes indicate edit was reverted WIKITRUST WikiTrust [1] score with ex-post-facto evidence (DE, EN only) WT_DELAY_DELT Difference in WIKITRUST and WT_NO_DELAY (DE, EN only) NEXT_TIME_AHEAD Time, in seconds, until article was next revised NEXT_USR_IP Whether the next editor of the article is an IP/anonymous editor NEXT_USR_SAME Whether the next article editor is same as current editor NEXT_COMM_VAND Whether the next “comment” indicates vandalism removal Table 2. Ex-post-facto features: Leveraging evidence after edit save, but before evaluation. No doubt, the strongest of these features is the WikiTrust score (WIKITRUST). This captures the notion of reputation-weighted content-persistence: text that survives is trustworthy, especially when the subsequent editors have good reputations. The Wik- iTrust values we obtain are from a lower-order classifier, encompassing ≈70 data points. However, it may be possible to improve upon or supplement the WikiTrust score. First, WikiTrust is computationally intense, having to track word-level histories. Sec- ond, content is sometimes removed or re-authored for reasons other than malicious intent. Third, WikiTrust is not presently enabled for all languages. This motivated our creation of feature HASH_REVERT, a more efficient and coarse-grained measure. The hash-code is computed for the article version prior-to, and immediately-after, the edit under inspection (scope is expanded if the editor makes multiple consecutive edits). If the hashes match it indicates an identity revert, the wholesale removal of the editor’s contributions, which is highly indicative of vandalism. Another novel feature, USR_PG_SZ_DELT, captures that poor contributors are of- ten notified/warned of their transgressions on their “talk page”. Informal analysis sug- gested that German and Spanish versions lack the standardized warning system that English employs [3]. Thus, a generic “size change” feature was implemented to detect such talk page contributions. 2.3 On Language-Driven Features When talking about language features, realize that is possible to produce language- driven features that are not language-specific (i.e., generic properties). Examples in- clude our features of the form LANG_*, as found at the bottom of Tab. 1. These mea- sures are certainly applicable to the languages used herein (German, English, Spanish) and analogues likely exist in many languages. However, these properties are unlikely to be universal in nature. In particular, different character sets (e.g., Hindi, Chinese, Japanese) might prove problematic, but this is ultimately outside the authors’ range of expertise. It should be noted that languages similar to those under evaluation (i.e., use of Latin characters, letter casing, space-delimited words, and Arabic numerals) represent a significant portion of Wikipedia’s article space3 . 3 http://meta.wikimedia.org/wiki/List_of_Wikipedias_by_language_group LANG-SPEC. FEAT. DESCRIPTION {DE,EN,ES}_OFFEND Quantity of offensive terms added/removed by edit *_OFFEND_IMPACT Normalizing *_OFFEND by ART_SIZE_DELT {DE,EN,ES}_PRONOUN Quantity of 1st-person pronouns added/removed *_PRONOUN_IMPACT Normalizing *_PRONOUN by ART_SIZE_DELT Table 3. Features requiring natural-language customization. Each feature is implemented inde- pendently, per-language. Spanish and German edits are also processed by the English versions. While generic language features are portable, they lack the intuition of language- specific ones. After all, profanity and slang have little place in encyclopedic content. Not only are such measures intuitive, they are effective, as the 2010 PAN-CLEF win- ning approach of Velasco [8] used multiple dictionaries (profanity, sexual terms, bi- ased words, etc.). This is disheartening as such features: (1) lack portability, (2) can be evaded with obfuscation, (3) require time-consuming implementation by fluent speak- ers, and (4) tend to be computationally expensive. Velasco, however, did not include many of the language-independent features we present in Tab. 1. Thus, as [2] sug- gested, language-independent features might overlap and render language-specific ones less critical. We extend that analysis here and do so across multiple natural languages. Unfortunately, Velasco’s dictionaries are not open source and the German and Span- ish equivalents must be implemented. Not NLP experts ourselves, we intend only to create proof-of-concept and non-exhaustive language-specific features, as per Tab. 3. This also allows us to perform cost-benefit analysis (i.e., the coverage of dictionaries vs. the performance improvement) and motivates our decision to encode three different approaches to compiling the offensive word lists (“offensive” here is just the combina- tion of all undesirable text categories): – S PANISH (ES): We re-purposed a scoring list designed for Spanish Wikipedia use4 . The list contains 800+ manually constructed regexps of extensive complexity (cap- turing intra-word permutations of diacritics, case, repeated letters, etc.). Manual inspection removed regexps not specific to offensive terminology. – E NGLISH (EN): A generic list of 1300+ offensive words (not regexps) is utilized5 . The list is not Wikipedia-specific, but does enumerate conjugated verb forms. – G ERMAN (DE): Unable to locate a dictionary of sufficient breadth, we decided to examine the feasibility of a programmatic approach. We took the union of infor- mal profanity lists and ran a stemming algorithm to produce roots which could be searched for as embedded (i.e., non word-boundary delimited) regexp matches. The text added and removed by an edit is scanned for word/regexp matches. The number of matches are quantified (+1 for additions, -1 for removals) and these form the {DE,EN,ES}_OFFEND features. The first-person “pronoun” features are straightfor- ward and intend to capture bias in authorship and possible non-neutral points-of-view. 4 http://es.wikipedia.org/wiki/Usuario:AVBOT/Lista_del_bien_y_del_mal 5 http://www.cs.cmu.edu/~biglou/resources/ ENGLISH FEATURE # . . . FEATURE . . . # . . . FEATURE . . . # WIKITRUST (F) 1 ART_SIZE_DELT 21 USR_LAST_RB 41 WT_DELAY_DELT (F) 2 USR_PG_SIZE 22 COMM_HAS_SEC 42 WT_NO_DELAY 3 ART_REP 23 ART_CHURN_CHARS 43 HASH_REVERT (F) 4 USR_PG_WARNS 24 COMM_IND_VAND 44 NEXT_COMM_VAND (F) 5 LANG_MARKUP 25 ART_CHURN_BLKS 45 USR_EDITS_MONTH 6 LANG_LONG_TOK 26 ART_EDITS_WEEK 46 USR_EDITS_WEEK 7 LANG_UCASE 27 ART_SIZE 47 USR_EDITS_EVER 8 EN_PRONOUN_IMPCT 28 ART_EDITS_DAY 48 USR_COUNTRY_REP 9 ART_EDITS_TOTAL 29 TIME_DOW 49 USR_EDITS_DENSE 10 USR_REP 30 ART_EDITS_HOUR 50 USR_IS_IP 11 ART_AGE 31 NEXT_USR_SAME (F) 51 USR_EDITS_DAY 12 LANG_ALPHA 32 USR_HAS_RB 52 USR_PG_SZ_DELT (F) 13 LANG_MARKUP 33 PREV_USR_IP 53 NEXT_TIME_AHEAD (F) 14 EN_PRONOUN 34 USR_BLK_EVER (F) 54 USR_AGE 15 ART_EDITS_DENSE 35 USR_BLK_BEFORE 55 COMM_LEN_NO_SEC 16 ART_DIVERSITY (F) 36 USR_IS_BOT 56 EN_OFFEND_IMPACT 17 LANG_CHAR_REP 37 NEXT_USR_IP (F) 57 USR_EDITS_HOUR 18 PREV_USR_SAME 38 TIME_TOD 58 EN_OFFEND 19 PREV_TIME_AGO 39 COMM_LEN 20 ART_EDITS_MONTH 40 Table 4. Kullback-Leibler divergence (i.e., information-gain) ranking for English features. Ex post facto signals are indicated by “(F)” (but ranking is independent, so a zero-delay list would have the same relative ordering). Foreign language features are not included for brevity. 3 Evaluation This section describes and evaluates the machine-learning model built atop our feature set. We begin by describing our choice of classification algorithm (Sec. 3.1). Then, this model is used to evaluate feature effectiveness over the labeled training set, paying particular attention to novel subsets (Sec. 3.2). Finally, we summarize performance over the PAN-CLEF 2011 competition test set (Sec. 3.3). 3.1 Classification Model The Weka [4] implementation of the alternating decision tree algorithm (ADTree) is used for scoring/classification. This method was chosen because it: (1) produces human- readable models, (2) handles missing features (API failures, missing data, etc.), and (3) supports enumerated features (our strategy has many booleans). ADTrees have one parameter of interest: the quantity of “boosting iterations” (i.e., tree-depth). German and Spanish classifiers utilize 18 iterations and English uses 30, quantities arrived at via cross-validation (the English training corpus [5] is 32× the size of the other two). 3.2 Training Set Evaluation All results are produced via 10-fold cross-validation over the training corpus [5]. The labels of the test corpus were withheld for the competition, as discussed in Sec. 3.3. # GERMAN ENGLISH SPANISH 1 WT_NO_DELAY WT_NO_DELAY USR_EDITS_MONTH 2 USR_EDITS_EVER USR_EDITS_MONTH USR_EDITS_WEEK (a) 3 USR_IS_IP USR_EDITS_WEEK USR_EDITS_EVER 4 USR_EDITS_MONTH USR_EDITS_EVER USR_IS_IP 5 USR_EDITS_WEEK USR_COUNTRY_REP ES_OFFEND_IMPACT 1 NEXT_COMM_VAND (F) WIKITRUST (F) NEXT_COMM_VAND (F) 2 WIKITRUST (F) WT_DELAY_DELT (F) NEXT_TIME_AHEAD (F) (b) 3 WT_NO_DELAY WT_NO_DELAY HASH_REVERT (F) 4 HASH_REVERT (F) HASH_REVERT (F) USR_PG_SZ_DELT (F) 5 NEXT_USR_IP (F) NEXT_COMM_VAND (F) USR_EDITS_MONTH Table 5. Extending Tab. 4 for all language corpora. Portion (a) permits only zero-delay features, while portion (b) also includes ex post facto signals, as indicated by “(F)”. Core Features and Cross-Language Consistency: We begin with the “core” set of features (Tab. 1). Though these have been described in the past, their cross-language evaluation is novel. Although space considerations prevent showing the full feature- ranking for all languages (Tab. 5a), they are remarkably similar to those presented for English (Tab. 4, ignoring “(F)” entries), especially when binned by the info-gain metric. That is, a feature tends to be equally effective no matter the language of evaluation. It is unsurprising that the zero-delay WikiTrust feature (WT_NO_DELAY) is the top- performing feature where available (English, German) – it is a lower-order classifier that wraps many data points. Beyond that, user participation statistics and registration status are also dominant. Generic language features tend to perform moderately (not all edits add content), with article-driven signals tending towards the bottom of the rankings. While the feature ranking is not unexpected, the cross-language consistency has stronger implications. It is a sociologically interesting observation that misbehavior is characterized similarly across language and cultural boundaries. More technically, it suggests the creation of language-independent classifiers might be feasible, eliminating the need for new corpora to be amassed for each new Wikipedia edition. Ex Post Facto Inclusion: As Tab. 5b demonstrates, the inclusion of ex post facto features dramatically modifies the list of “best features,” with 4 of the top 5 being of this type for all languages. Such signals also positively affect overall performance, varying between 3.6% (English) and 13.6% (Spanish) PR-AUC increase (see Tab. 6). While these improvements are not overwhelming, it should be emphasized that the high- accuracy of zero-delay approaches decreases the possible margin for improvement. These ex post facto features are redundant, however, all trying to capture the same notion: “was the edit reverted?” (particularly WIKITRUST, NEXT_COMM_VAND, and HASH_REVERT). While all are features of exemplary performance, they vary in effi- ciency and robustness. For example, WikiTrust employs a complex but secure algorithm that mines reputation from implicit Wikipedia actions. In contrast, NEXT_COMM_VAND parses explicit summaries for keywords, which while simple, could easily be gamed. The degree to which secure features are required is not immediately apparent. Vandals are typically poorly incentivized [7] and therefore may not evade crude protections. GERMAN ENGLISH SPANISH METRIC RND ZD ALL RND ZD ALL RND ZD ALL PR-AUC 0.302 0.878 0.930 0.074 0.773 0.801 0.310 0.868 0.986 ROC-AUC 0.500 0.958 0.981 0.500 0.963 0.968 0.500 0.946 0.993 Table 6. Area-under-curve (AUC) measurements for feature sets over training data. This is done for precision-recall (PR) and receiver-operating characteristic (ROC) curves. Feature sets include a control classifier (random, RND), zero-delay (ZD), and including ex post facto data (ALL). LANG ZD-WO ZD-W DIFF% ALL-WO ALL-W DIFF% (PR-AUC) DE 0.881 0.878 -0.34% 0.930 0.930 ±0.00% (PR-AUC) EN 0.737 0.773 +4.89% 0.776 0.801 +3.22% (PR-AUC) ES 0.805 0.868 +7.83% 0.988 0.986 -0.20% Table 7. Measuring the impact of language-specific features (Tab. 3). Feature sets are evaluated with (W) and without (WO) the inclusion of language-specific signals. Otherwise, acronyms are as defined as in Tab. 6. PR-AUC is the singular metric used in this comparison. Cost vs. Benefits of Language-Specific Signals: As Tab. 7 shows, the performance benefit of language-specific features varies dramatically. They prove most helpful when targeting zero-delay detection, and the extensiveness and expertise involved in creating the “offensive word list” correlates with performance gains. Recall from Sec. 2.3 that our German approach was quite crude (a stemming algorithm over informal profanity lists). Such attempts did not translate positively, adding only noise to the classifier. At the other extreme, a third-party, Wikipedia-customized, and complex set of regular- expressions was able to increase zero-delay PR-AUC by nearly 8% in the Spanish case. Where infrastructure already exists for these purposes, it can and should be re- utilized (as we did for English and Spanish). Where it does not, it would seem casual attempts should be avoided. More broadly, it seems wise to investigate autonomous (and language-independent) means to produce robust dictionaries (e.g., n-grams). Cumulative Performance: A broader viewer of classifier performance is presented numerically in Tab. 6 and visualized in Fig. 1. One interesting observation is the vary- ing performance between languages. English, despite having the most enabled features, and 32× more training examples, is classified much poorer than Spanish and German. At current, we have two hypotheses why this is the case. First, English has a tool called the “Edit Filter” which prevents trivial vandalism from being saved6 (and becoming a corpus member). We are unaware of any German/Spanish equivalent, meaning obvious vandalism (i.e., “low-hanging fruit”) would be corpus members in those cases. Sec- ond, vandalism tagging is a subjective process. The labeling of the English corpus was done via Amazon Mechanical Turk [5] (utilizing random persons), whereas the smaller German/Spanish versions involved Wikipedia researchers. The latter group is likely to be more consistent in upholding the standards of the Wikipedia community, and such agreement is particularly important for features like NEXT_COMM_VAND. 6 http://en.wikipedia.org/wiki/Wikipedia:Edit_Filter Random Zero-Delay w/Ex-Post-Facto 1 1 0.8 0.8 precision precision 0.6 0.6 0.4 0.4 0.2 0.2 0 0 0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1 (a) German (de) (b) English (en) (c) Spanish (es) Figure 1. Precision-recall curves over training data. # GERMAN ENGLISH SPANISH 1 WT_NO_DELAY EN_OFFEND_IMPACT ES_OFFEND_IMPACT 2 USR_EDITS_MONTH USR_PG_WARNS USR_IS_IP (a) 3 ART_CHURN_CHARS WT_NO_DELAY TIME_TOD 4 USR_PG_SIZE USR_EDITS_MONTH LANG_UCASE 5 ART_SIZE_DELT LANG_UCASE PREV_USR_IP 1 NEXT_COMM_VAND (F) WIKITRUST (F) NEXT_COMM_VAND (F) 2 USR_IS_IP NEXT_COMM_VAND (F) USR_EDITS_WEEK (b) 3 LANG_UCASE LANG_MARKUP NEXT_TIME_AHEAD (F) 4 LANG_ALPHA USR_COUNTRY_REP PREV_TIME_AGO 5 ART_CHURN_CHARS LANG_LONG_TOK LANG_LONG_TOK Table 8. Top feature subsets of size n = 5, calculated using greedy step-wise analysis. Portion (a) permits only zero-delay features; (b) includes ex post facto ones. Regardless, English-language performance (the only known baseline) is comparable to the state-of-the-art. That benchmark was set in our prior work [2], which this writing re-implements with slight modifications. It should be emphasized that it was not our intention to best that prior work, rather, we sought to use the expanded PAN-CLEF 2011 rules/corpora to analyze novel portions of the problem space. Finally, it is interesting to produce the most effective feature subsets for each lan- guage (Tab. 8). Unlike Tab. 5, this list considers feature correlation and overlap; dis- playing the features weighted most heavily in the actual ADTree models. These or- derings are quite unique compared to Tabs. 4 & 5, and greater analysis is needed to determine what correlations give rise to these rule chains. For instance, English feature LANG_MARKUP ranked 25th in info-gain, yet was the 3rd highest ranking in subset form. Results like these imply a large degree of overlap between features, suggesting that small (and therefore, efficient) feature sets/trees can produce accurate results. 3.3 Test Set Performance When applied to the label-withheld test set, our model won the 2011 PAN-CLEF com- petition. The PR-AUCs (EN= 0.706, EN= 0.822, ES= 0.489) show a slight perfor- mance increase for English, but a dramatic drop for German/Spanish relative to cross- validation over training data (Tab. 6). When the test corpus labels are revealed, they should be inspected to see if some type of systematic bias gave rise to this discrepancy. 4 Conclusions Our novel research directions in this paper were motivated by changes in the 2011 PAN-CLEF competition with respect to both the 2010 edition and the bulk of exist- ing Wikipedia vandalism research. First, the competition permitted features to leverage evidence after the edits were made. We identified multiple metrics of this type, which were extremely effective, and whose implementation made clear the trade-off between feature efficiency and robustness. Second, the competition spanned three natural languages. For language-independent features (i.e., metadata) this was the first non-English evaluation of such signals, though relative order was found to be surprisingly consistent across languages. Multiple lan- guages, however, imply costly localization for language-specific features (e.g., profanity lists), forcing examination of their effectiveness. Including these atop an extensive set of language-independent features, we find that minor-to-moderate contributions are still possible, and the degree of improvement correlates with the localization’s complexity. We hope that this work continues to promote and improve the autonomous detection of vandalism. Such progress frees editors of monitoring roles and allows them to better contribute to a growing body of collaborative knowledge. Acknowledgements: This research was supported in part by ONR MURI N00014-07-1-0907. The authors recognize those colleagues whose techniques were a component in the described approach: B. Thomas Adler, Luca de Alfaro, Santiago Mola-Velasco, Sampath Kannan, Ian Pye, and Paolo Rosso (see [2]). Andreas Haeberlen is thanked for his German language assistance. Martin Potthast is acknowledged for his continued dedication to the vandalism detection task. References 1. Adler, B.T., de Alfaro, L.: A content-driven reputation system for the Wikipedia. In: WWW’07: Proc. of the 16th International World Wide Web Conference (May 2007) 2. Adler, B., de Alfaro, L., Mola-Velasco, S.M., Rosso, P., West, A.G.: Wikipedia vandalism detection: Combining natural language, metadata, and reputation features. In: CICLing’11 (Comp. Linguistics and Intelligent Text Processing) and LNCS 6609 (February 2011) 3. Geiger, R.S., Ribes, D.: The work of sustaining order in Wikipedia: The banning of a vandal. In: CSCW’10: Proc. of the Conf. on Computer Supported Cooperative Work (2010) 4. Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., Witen, I.H.: The WEKA data mining software: An update. SIGKDD Explorations 11(1) (2009) 5. Potthast, M.: Crowdsourcing a Wikipedia vandalism corpus. In: SIGIR‘10: Proc. of the 33rd International ACM SIG Information Retrieval Conference. pp. 189–790 (2010) 6. Potthast, M., Stein, B., Holfeld, T.: Overview of the 1st International competition on Wikipedia vandalism detection. In: PAN-CLEF 2010 Labs and Workshops (2010) 7. Priedhorsky, R., Chen, J., Lam, S.K., Panciera, K., Terveen, L., Riedl, J.: Creating, destroying, and restoring value in Wikipedia. In: ACM GROUP’07 (2007) 8. Velasco, S.M.M.: Wikipedia vandalism detection through machine learning: Feature review and new proposals. Tech. rep., Lab Report for PAN at CLEF 2010 (2010) 9. West, A.G., Chang, J., Venkatasubramanian, K., Lee, I.: Trust in collaborative web applications. Future Generation Comp. Sys. section on Trusting Software Behavior (2011) 10. West, A.G., Kannan, S., Lee, I.: Detecting Wikipedia vandalism via spatio-temporal analysis of revision metadata. In: EUROSEC’10: European Wkshp. on Sys. Security (2010)