《电子商务 E-business》阅读文献：Human-competitive tagging using automatic keyphrase extraction

Human-competitive tagging using automatic keyphrase extraction Olena Medelyan, Eibe frank, lan h. witten Computer Science Department University of Waikato olena, eibe, ihw)@cs. waikato. ac nz tags. How well do human taggers perform? How abstract consistent are they with each other? One potential solution to inconsistency This paper connects two research areas: auto- folksonomies is to use suggestion tools matic tagging on the web and statistical key automatically compute tags for new documents phrase extraction. First, we analyze the quality (e.g. Mishne, 2006 Sood et al, 2007; Heymann using traditional evaluation techniques. Next, et al, 2008). Interestingly, the blooming research we demonstrate how documents can be tagged on automatic tagging has so far not been con automatically with a state-of-the-art keyphrase nected to work on keyphrase extraction(e.g traction algorithm, and further improve per Frank et al, 1999, Turney, 2003, Hulth, 2004) ormance in this new domain using a new al which can be used as a tool for the same task gorithm,"Maui, that utilizes semantic infor (note: we use tag and keyph mation extracted from Wikipedia. Maui out- Instead of simple heuristics based on term fre performs existing approaches and extracts tags quencies and co-occurrence of tags, keyphrase hat are competitive with those assigned by the extraction methods apply machine learning to best performing human tagger determine typical distributions of properties common to manually assigned phrases, and can 1 Introduction include analysis of semantic relations betv Tagging is the process of labeling web resources candidate tags (Turney, 2003) state-of-the-art keyphrase extraction systems per- based on their content. Each label, or tag, corre- form compared to simple tagging techniques? sponds to a topic in a given document. Unlike How consistent are they with human taggers? metadata assigned by authors, or by professional indexers in libraries, tags are assigned by end These are questions we address in this paper users for organizing and sharing information that Until now, keyphrase extraction methods have is of interest to them. The organic system of tags primarily been evaluated using a single set of assigned by all users of a given web platform is keyphrases for each document, thereby largely called a folksonomy ignoring the subjective nature of the tas Collaboratively tagged documents, on the other In contrast to traditional taxonomies painstak- hand, offer multiple tag assignments by inde- ingly constructed by experts, a user can add any tags to a folksonomy. This leads to the greatest pendent users, a unique basis for evaluation that downside of tagging, inconsistency, which orig- we capitalize upon in this paper nates in the synonymy and polysemy of human The experiments reported in this paper language, as well as in the varying degrees of these gaps in the research on automatic tagging specificity used by taggers(Golder and Huber- and keyphrase extraction. First, we analyze tag man, 2006). In traditional libraries, consistency is gng consistency on the CiteULike. org platform the primary evaluation criterion of indexing or organizing academic citations Methods tradi (Rolling, 1981). Much work has been done on tionally used for the evaluation of professional describing the statistical properties of folksono- indexing will provide insight into the quality of mies such as tag distribution and co-occurrences this folksonomy. Next, we extract a high quality (Halpin et aL, 2007; Sigurbjornsson et al., 2008. corpus from CiteULike, containing documents Sood et al., 2007), but to our knowledge there that have been tagged consistently by the best has been none on assessing the actual quality of human taggers Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1318-1327 Singapore, 6-7 August 2009. 2009 ACL and AFNLP

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1318–1327, Singapore, 6-7 August 2009. c 2009 ACL and AFNLP Human-competitive tagging using automatic keyphrase extraction Olena Medelyan, Eibe Frank, Ian H. Witten Computer Science Department University of Waikato {olena,eibe,ihw}@cs.waikato.ac.nz Abstract This paper connects two research areas: auto￾matic tagging on the web and statistical key￾phrase extraction. First, we analyze the quality of tags in a collaboratively created folksonomy using traditional evaluation techniques. Next, we demonstrate how documents can be tagged automatically with a state-of-the-art keyphrase extraction algorithm, and further improve per￾formance in this new domain using a new al￾gorithm, “Maui”, that utilizes semantic infor￾mation extracted from Wikipedia. Maui out￾performs existing approaches and extracts tags that are competitive with those assigned by the best performing human taggers. 1 Introduction Tagging is the process of labeling web resources based on their content. Each label, or tag, corre￾sponds to a topic in a given document. Unlike metadata assigned by authors, or by professional indexers in libraries, tags are assigned by end￾users for organizing and sharing information that is of interest to them. The organic system of tags assigned by all users of a given web platform is called a folksonomy. In contrast to traditional taxonomies painstak￾ingly constructed by experts, a user can add any tags to a folksonomy. This leads to the greatest downside of tagging, inconsistency, which origi￾nates in the synonymy and polysemy of human language, as well as in the varying degrees of specificity used by taggers (Golder and Huber￾man, 2006). In traditional libraries, consistency is the primary evaluation criterion of indexing (Rolling, 1981). Much work has been done on describing the statistical properties of folksono￾mies, such as tag distribution and co-occurrences (Halpin et al., 2007; Sigurbjörnsson et al., 2008; Sood et al., 2007), but to our knowledge there has been none on assessing the actual quality of tags. How well do human taggers perform? How consistent are they with each other? One potential solution to inconsistency in folksonomies is to use suggestion tools that automatically compute tags for new documents (e.g. Mishne, 2006; Sood et al., 2007; Heymann et al., 2008). Interestingly, the blooming research on automatic tagging has so far not been con￾nected to work on keyphrase extraction (e.g. Frank et al., 1999; Turney, 2003; Hulth, 2004), which can be used as a tool for the same task (note: we use tag and keyphrase as synonyms). Instead of simple heuristics based on term fre￾quencies and co-occurrence of tags, keyphrase extraction methods apply machine learning to determine typical distributions of properties common to manually assigned phrases, and can include analysis of semantic relations between candidate tags (Turney, 2003). How well do state-of-the-art keyphrase extraction systems per￾form compared to simple tagging techniques? How consistent are they with human taggers? These are questions we address in this paper. Until now, keyphrase extraction methods have primarily been evaluated using a single set of keyphrases for each document, thereby largely ignoring the subjective nature of the task. Collaboratively tagged documents, on the other hand, offer multiple tag assignments by inde￾pendent users, a unique basis for evaluation that we capitalize upon in this paper. The experiments reported in this paper fill these gaps in the research on automatic tagging and keyphrase extraction. First, we analyze tag￾ging consistency on the CiteULike.org platform for organizing academic citations. Methods tradi￾tionally used for the evaluation of professional indexing will provide insight into the quality of this folksonomy. Next, we extract a high quality corpus from CiteULike, containing documents that have been tagged consistently by the best human taggers. 1318

2.1 Extracting a high quality tagged corpus 713, 600 all C teULike articles The CiteULike data set is freely available and 513, 100 at least 1 tac contains information about which documents 367, 600 at least 2 tags were tagged with what tags by which users(al though identities are not provided) CiteULike's 22 300 users have tagged 713, 600 documents with 2. 4M"tag assignments'-sin- 16,600 at least 3 taggers gle applications of a tag by a user to a document 2, 500 at least 3 agreed tag The two most popular tags, bibtex-import and no-tag, indicate an information source and a missing tag respectively. Most of the remainder describe particular concepts relevant to the documents. We exclude non-content tags fron Figure 1. Quality control of CiteULike data our experiments, e.g. personal tags like to-read or todo. Note that spam entries have been elimi- Following that, our goal is to build a system nated from the data set Because CiteULike tagger at matches the performance of these taggers. sional indexers, high quality of the assigned top- We first apply an existing approach proposed by Brooks and Montanez(2006) and compare it to s cannot be guaranteed. In fact, manual as- sessment of users' tags by human evaluators the keyphrase extraction algorithm Kea(Frank et shows precision of 59%(Mishne, 2006)and 49% al, 1999). Next we create a new algorithm alled Maui, that enhances Kea's successful ma. (Sood et al. 2006). However, why is the opinion chine learning framework with semantic know of human evaluators valued more than the edge retrieved from Wikipedia, new features, and lon of taggers? We propose an alternative way of a new classification model. We evaluate mau determining ground truth using an automatic ap- proach to determine reliable tags: We concen using tag sets assigned to the same documents by trate on a subset of CiteULike containing docu- with Citeulike users as they are with each other. ments that have been indexed with at least three Most of the computation required for auto matic tagging with this method can be performed sistency between the users, and then compare it offline. In practice, it can be used as a tag sug- to the algorithm's consistency, we need taggers gestion tool that provides users with tags describ who have tagged documents that some others ing the main topics of newly added documents, had tagged. We say that two users are "co- which can then be corrected or enhanced by per- taggers"if they have both tagged at least one sonal tags if required. This will improve consis common document. As well as restricting the tency in the folksonomy without compromising document set, we only include taggers who have its flexibility at least two co-taggers 2 Collaboratively-tagged data Figure 1 shows the proportions of CiteULike documents that are discarded in order to produce ike. org is a bookmarking service that re- our high quality data set. The final set contains sembles the popular deL icio us, but concentrates only 2, 100 documents(0.3% of the original) on scholarly papers. Rather than replicating the Unfortunately, many of these are unavailable for fulltextoftaggedpapersitsimplypointstothemdownload-forexamplebooksatAmazon.com on the web(e.g. PubMed, Cite Seer, Science Di- and ArXiv. org references cannot be crawled. We rect, Amazon)or in journals(e. g. High Wire, Na- further restrict attention to two sources: High ture). This avoids violating copyright but means Wire and Nature, both of which provide easily that the full text of articles is not necessarily accessible PDFs of the full text available. When entering new resources, users The result is a set of 180 documents indexed re encouraged to assign tags describing their by 332 taggers. A total of 4,638 tags were content or reflecting their own grouping of the signed by all taggers to documents in this set; information. However, the system does not sug- however, the number of tags on which gest tags. Moreover, users do not see other users' two users agreed is significantly smaller, n tags and are thus not biased in their tag choices. 946. Still, this results in accurate tag sets that contain an average of five tags per document

Following that, our goal is to build a system that matches the performance of these taggers. We first apply an existing approach proposed by Brooks and Montanez (2006) and compare it to the keyphrase extraction algorithm Kea (Frank et al., 1999). Next we create a new algorithm, called Maui, that enhances Kea’s successful ma￾chine learning framework with semantic knowl￾edge retrieved from Wikipedia, new features, and a new classification model. We evaluate Maui using tag sets assigned to the same documents by several users and show that it is as consistent with CiteULike users as they are with each other. Most of the computation required for auto￾matic tagging with this method can be performed offline. In practice, it can be used as a tag sug￾gestion tool that provides users with tags describ￾ing the main topics of newly added documents, which can then be corrected or enhanced by per￾sonal tags if required. This will improve consis￾tency in the folksonomy without compromising its flexibility. 2 Collaboratively-tagged Data CiteULike.org is a bookmarking service that re￾sembles the popular del.icio.us, but concentrates on scholarly papers. Rather than replicating the full text of tagged papers it simply points to them on the web (e.g. PubMed, CiteSeer, ScienceDi￾rect, Amazon) or in journals (e.g. HighWire, Na￾ture). This avoids violating copyright but means that the full text of articles is not necessarily available. When entering new resources, users are encouraged to assign tags describing their content or reflecting their own grouping of the information. However, the system does not sug￾gest tags. Moreover, users do not see other users’ tags and are thus not biased in their tag choices. 2.1 Extracting a high quality tagged corpus The CiteULike data set is freely available and contains information about which documents were tagged with what tags by which users (al￾though identities are not provided). CiteULike’s 22,300 users have tagged 713,600 documents with 2.4M “tag assignments”— sin￾gle applications of a tag by a user to a document. The two most popular tags, bibtex-import and no-tag, indicate an information source and a missing tag respectively. Most of the remainder describe particular concepts relevant to the documents. We exclude non-content tags from our experiments, e.g. personal tags like to-read or todo. Note that spam entries have been elimi￾nated from the data set. Because CiteULike taggers are not profes￾sional indexers, high quality of the assigned top￾ics cannot be guaranteed. In fact, manual as￾sessment of users’ tags by human evaluators shows precision of 59% (Mishne, 2006) and 49% (Sood et al., 2006). However, why is the opinion of human evaluators valued more than the opin￾ion of taggers? We propose an alternative way of determining ground truth using an automatic ap￾proach to determine reliable tags: We concen￾trate on a subset of CiteULike containing docu￾ments that have been indexed with at least three tags on which at least two users have agreed. In order to be able to measure the tagging con￾sistency between the users, and then compare it to the algorithm’s consistency, we need taggers who have tagged documents that some others had tagged. We say that two users are “co￾taggers” if they have both tagged at least one common document. As well as restricting the document set, we only include taggers who have at least two co-taggers. Figure 1 shows the proportions of CiteULike documents that are discarded in order to produce our high quality data set. The final set contains only 2,100 documents (0.3% of the original). Unfortunately, many of these are unavailable for download—for example, books at Amazon.com and ArXiv.org references cannot be crawled. We further restrict attention to two sources: High￾Wire and Nature, both of which provide easily￾accessible PDFs of the full text. The result is a set of 180 documents indexed by 332 taggers. A total of 4,638 tags were as￾signed by all taggers to documents in this set; however, the number of tags on which at least two users agreed is significantly smaller, namely 946. Still, this results in accurate tag sets that contain an average of five tags per document. Figure 1. Quality control of CiteULike data 1319

O-taggers documents consistency personalized tag choices as Chirita et al.(2007) 1234567 suggest, but rather stem from the lack of guide 714 1661 555625686 lines and uniform tag suggestions that a book marking service could provide 2.2 Measuring tagging consistency 248 48.3 Traditional indexers aim for consistency, on the 47.1 45.4 basis that this will enhance document retrieval 44.4 (Leonard, 1975). Consistency is measured using 8665 43.5 experiments in which several people index the 7787942490 same documents--usually a small set of a few 123456789022 40.9 dozen documents. It is computed for pairs of in 39.7 38.8 deers, by formulae such as rollings(1981) 35948 37. Consistency(/,)=2C. where C is the number of tags(index terms)in dexers I, and 1, have in common and a and b is 7671 16506 the size of their tag sets respectively In our experiments, before computing the number of terms in common. we stem each tas with the Porter(1980)stemmer. For example, the 678 overlap C between the tag sets (complex systems, network, small world, and theoretical, small 10 12 29 28.8 world, networks, dynamics) consist of the two 30 10 279 tags (network, small world,, and the consistency 31 10 26.7 is2×2/(3+4)=0.57. To compute the overall consistency of a par 34 21.0 ticular indexer, this figure is averaged over all documents and co-indexers. There were no cases where the same user reassigned tags to the same average 7.5 37.7 articles, so computing intra-tagger consistency Table 1. Consistency of the most prolific and although interesting, was not impossible most consistent taggers To our knowledge. traditional indexing consis- yet been applied to col Note that traditionally much smaller data sets are laboratively tagged data. However, experiments used to assess consistency of human indexers, on determining tagging quality do follow the because such sets need to be created specifically same idea. For example Xu et al.(2006) define for the experiment. Collaborative tagging plat- an authority metric that assigns high scores to forms like CiteULike can be mined for large col- those users who match other users'choices on lections of this kind in natural settings the same documents. in order to eliminate the area of bioinformatics. To give an example, a 2.3 Consistency of CiteULike taggers document entitled Initial sequencing and com parative analysis of the mouse genome was In the collec 80 documents tagged by 332 tagged by eight users with a total of 22 tags. Four users described in Section 3.1, each tagger has 18 of them agreed on the tag mouse, but one used co-taggers on average, ranging from 2 to 129 the broader term rodents. Three agreed on the tag and has indexed I to 25 documents. For each genome, but one added genome paper, and an- user we compute the consistency with all other other used the more specific comparative genom- users who tagged the e document. Consis- ics. There are also cases when tags are written tency is then averaged across documents. We together, e.g. genomepaper, or with a prefix key found that the distribution of per-user consis- genome, or in a different grammatical form: se- tency resembles a power law with a few users quence vs sequencing. This example shows achieving high consistency values and a long tail many inconsistencies in tags are not caused of inconsistent taggers The maximum consis

Note that traditionally much smaller data sets are used to assess consistency of human indexers, because such sets need to be created specifically for the experiment. Collaborative tagging plat￾forms like CiteULike can be mined for large col￾lections of this kind in natural settings. Most documents in the extracted set relate to the area of bioinformatics. To give an example, a document entitled Initial sequencing and com￾parative analysis of the mouse genome was tagged by eight users with a total of 22 tags. Four of them agreed on the tag mouse, but one used the broader term rodents. Three agreed on the tag genome, but one added genome paper, and an￾other used the more specific comparative genom￾ics. There are also cases when tags are written together, e.g. genomepaper, or with a prefix key genome, or in a different grammatical form: se￾quence vs. sequencing. This example shows that many inconsistencies in tags are not caused by personalized tag choices as Chirita et al. (2007) suggest, but rather stem from the lack of guide￾lines and uniform tag suggestions that a book￾marking service could provide. 2.2 Measuring tagging consistency Traditional indexers aim for consistency, on the basis that this will enhance document retrieval (Leonard, 1975). Consistency is measured using experiments in which several people index the same documents—usually a small set of a few dozen documents. It is computed for pairs of in￾dexers, by formulae such as Rolling’s (1981): , where C is the number of tags (index terms) in￾dexers I1 and I2 have in common and A and B is the size of their tag sets respectively. In our experiments, before computing the number of terms in common, we stem each tag with the Porter (1980) stemmer. For example, the overlap C between the tag sets {complex systems, network, small world} and {theoretical, small world, networks, dynamics} consist of the two tags {network, small world}, and the consistency is 2×2/(3+4) = 0.57. To compute the overall consistency of a par￾ticular indexer, this figure is averaged over all documents and co-indexers. There were no cases where the same user reassigned tags to the same articles, so computing intra-tagger consistency, although interesting, was not impossible. To our knowledge, traditional indexing consis￾tency metrics have not yet been applied to col￾laboratively tagged data. However, experiments on determining tagging quality do follow the same idea. For example, Xu et al. (2006) define an authority metric that assigns high scores to those users who match other users’ choices on the same documents, in order to eliminate spammers. 2.3 Consistency of CiteULike taggers In the collection of 180 documents tagged by 332 users described in Section 3.1, each tagger has 18 co-taggers on average, ranging from 2 to 129, and has indexed 1 to 25 documents. For each user we compute the consistency with all other users who tagged the same document. Consis￾tency is then averaged across documents. We found that the distribution of per-user consis￾tency resembles a power law with a few users achieving high consistency values and a long tail of inconsistent taggers. The maximum consis￾tagger co-taggers documents consistency 1 1 5 71.4 2 1 5 71.4 3 6 5 57.9 4 6 6 51.0 5 11 12 50.4 6 2 5 50.1 7 4 6 48.3 8 8 8 47.1 9 13 16 45.4 10 12 8 44.4 11 7 6 43.5 12 7 6 41.7 13 8 5 40.9 14 7 6 39.7 15 9 13 38.8 16 4 5 38.4 17 12 9 37.3 18 4 14 36.1 19 9 8 35.9 20 10 11 33.7 21 7 6 33.1 22 6 5 33.0 23 7 10 32.1 24 11 16 31.7 25 8 13 30.6 26 6 8 30.6 27 9 6 29.8 28 10 12 29.0 29 8 6 28.8 30 9 10 27.9 31 10 8 26.7 32 8 7 26.3 33 10 5 25.6 34 8 7 21.0 35 9 9 18.3 36 3 6 7.9 average 7.5 8.1 37.7 Table 1. Consistency of the most prolific and most consistent taggers 1320

tency in this group is 92.3% and the average is 3.1 Features indicating significance 18.5%. The average consistency of the most pro- We now describe the features used in the classi- lific 70 indexers-those who have indexed fication model to determine whether a phrase is least five documents-is in the same range, likely to be a tag. We begin with three baseline namely 18.4%. The consistency of traditional approaches to free indexing is reported to be be- extend the set with three features that have been tween 4% and 67%, with an average of 27% de found useful in previous work. We also add three pending on what aids are used (Leininger, 2000). new features that have not been evaluated before It is instructive to consider the group of best taggers. We define these as the ones who (a)ex. Wikipedia linkage. All Wikipedia-based features It greater than average consistency with all are computed using the WikipediaMiner toolkit others, and(b) are su tagged at least five documents. There are 36 such 1. TFXIDF combines the frequency of a taggers Table I lists their consistency within this occurrence frequency in general use(Salton and group. The average consistency they achieve as a McGill, 1983 ). This score is high for rare phrases group is 37.7%, which is the similar to the aver- that appear frequently in a document and there- age consistency of professionals (Leininger, fore are more likely to be significant 2. Position of the first occurrence is com The above consistency analysis provides in- puted as the relative distance of the first occur sight into the tagging quality of the best rence of the candidate tag from the beginning of CiteULike users, based on High Wire and Nature the document. Candidates with very high or very articles. For the purposes of this paper, it shows low values are likely to be tags, because they how the tagging community can be restricted to a appear either in the opening document parts such best-performing group of taggers by measuring as title, abstract, table of contents, and introduc- their consistency. This is helpful for testing the tion, or in the documents final sections such as performance of automatic tagging(Section 4. 4). conclusion and reference lists 3 Automatic tagging with maui 3. Keyphraseness quantifies how often a can didate phrase appears as a tag in the training cor- Maui is a general algorithm for automatic topical pus. Automatic tagging approaches utilize the indexing based on the Kea system(Frank et al, same information: Mishne(2006)and Sood et al. 1999). It works in two stages: candidate selec-(2006)automatically suggest tags previously as tion and machine learning based filtering. In this signed to similar documents. However, in Maui paper, we apply it to automatic tagging In the (as in Kea)this feature is just one component of candidate selection stage, Maui first determines the overall model. Thus if a candidate never ap- textual sequences defined by orthographic pears as a keyphrase in the training corpus, it can boundaries and splits these sequences into to- still be extracted if its other feature values are kens. Then all n-grams up to a maximum length significant enough 4. Phrase length is measured in words. gen- of 3 words that do not begin or end with a stop- erally speaking, the longer the phrase, the more word are extracted as candidate tags. To reduce the number of candidates, all those that appear specific it is. Training captures and quantifies the only once are discarded. This speeds up the train- specificity preference in a given training corpus ing and the extraction process without impacting 5. Node degree quantifies the semantic relat the results In the filtering stage several features edness of a candidate tag to other candidates are computed for each candidate, which are then Turney(2003)computes semantic relatedness input to a machine learning model to obtain the using search engine statistics. Instead, following bability that the candidate is indeed Medelyan et al.(2008), we utilize wikipedia Maui 's architecture resembles that hyperlinks for this task. We first map each can other supervised keyphrase extraction systems didate phrase to its most common Wikipedia (Turney, 2000; Hulth 2004; Medelyan et al., page. For example, the word Jaguar appears as a 2008). However, this architecture has not previ- link anchor in wikipedia 927 times. In 466 cases ously been applied to the task of automatic tag monness of this mapping is 0.5. In 203 cases it gIng links to the animal description, a commonness of Maui is open-source and available for download hp:∥ ttp: //wikipedia sourceforge. net/ 1321

tency in this group is 92.3% and the average is 18.5%. The average consistency of the most pro￾lific 70 indexers—those who have indexed at least five documents—is in the same range, namely 18.4%. The consistency of traditional approaches to free indexing is reported to be be￾tween 4% and 67%, with an average of 27% de￾pending on what aids are used (Leininger, 2000). It is instructive to consider the group of best taggers. We define these as the ones who (a) ex￾hibit greater than average consistency with all others, and (b) are sufficiently prolific, i.e. have tagged at least five documents. There are 36 such taggers; Table 1 lists their consistency within this group. The average consistency they achieve as a group is 37.7%, which is the similar to the aver￾age consistency of professionals (Leininger, 2000). The above consistency analysis provides in￾sight into the tagging quality of the best CiteULike users, based on HighWire and Nature articles. For the purposes of this paper, it shows how the tagging community can be restricted to a best-performing group of taggers by measuring their consistency. This is helpful for testing the performance of automatic tagging (Section 4.4). 3 Automatic tagging with Maui Maui is a general algorithm for automatic topical indexing based on the Kea system (Frank et al., 1999).1 It works in two stages: candidate selec￾tion and machine learning based filtering. In this paper, we apply it to automatic tagging. In the candidate selection stage, Maui first determines textual sequences defined by orthographic boundaries and splits these sequences into to￾kens. Then all n-grams up to a maximum length of 3 words that do not begin or end with a stop￾word are extracted as candidate tags. To reduce the number of candidates, all those that appear only once are discarded. This speeds up the train￾ing and the extraction process without impacting the results. In the filtering stage several features are computed for each candidate, which are then input to a machine learning model to obtain the probability that the candidate is indeed a tag. Maui’s architecture resembles that of many other supervised keyphrase extraction systems (Turney, 2000; Hulth 2004; Medelyan et al., 2008). However, this architecture has not previ￾ously been applied to the task of automatic tag￾ging. 1 Maui is open-source and available for download at http://maui-indexer.googlecode.com 3.1 Features indicating significance We now describe the features used in the classi￾fication model to determine whether a phrase is likely to be a tag. We begin with three baseline features used in Kea (Frank et al., 1999), and extend the set with three features that have been found useful in previous work. We also add three new features that have not been evaluated before: spread, semantic relatedness and inverse Wikipedia linkage. All Wikipedia-based features are computed using the WikipediaMiner toolkit. 2 1. TF×IDF combines the frequency of a phrase in a particular document with its inverse occurrence frequency in general use (Salton and McGill, 1983). This score is high for rare phrases that appear frequently in a document and there￾fore are more likely to be significant. 2. Position of the first occurrence is com￾puted as the relative distance of the first occur￾rence of the candidate tag from the beginning of the document. Candidates with very high or very low values are likely to be tags, because they appear either in the opening document parts such as title, abstract, table of contents, and introduc￾tion, or in the document’s final sections such as conclusion and reference lists. 3. Keyphraseness quantifies how often a can￾didate phrase appears as a tag in the training cor￾pus. Automatic tagging approaches utilize the same information: Mishne (2006) and Sood et al. (2006) automatically suggest tags previously as￾signed to similar documents. However, in Maui (as in Kea) this feature is just one component of the overall model. Thus if a candidate never ap￾pears as a keyphrase in the training corpus, it can still be extracted if its other feature values are significant enough. 4. Phrase length is measured in words. Gen￾erally speaking, the longer the phrase, the more specific it is. Training captures and quantifies the specificity preference in a given training corpus. 5. Node degree quantifies the semantic relat￾edness of a candidate tag to other candidates. Turney (2003) computes semantic relatedness using search engine statistics. Instead, following Medelyan et al. (2008), we utilize Wikipedia hyperlinks for this task. We first map each can￾didate phrase to its most common Wikipedia page. For example, the word Jaguar appears as a link anchor in Wikipedia 927 times. In 466 cases it links to the article Jaguar cars, thus the com￾monness of this mapping is 0.5. In 203 cases it links to the animal description, a commonness of 2 http://wikipedia-miner.sourceforge.net/ 1321

0. 22. We compute the node degree of the corre- 3.2 Machine learning in Maui sponding Wikipedia article as the number of In order to build the model, we use the subset of hyperlinks that connect it to other Wikipedia the CiteULike collection described in Section pages that have been identified for other candi- date tags from the same document. a document 31. For each document we know a that describes a particular topic will cover many that at least two users have agreed on. This is related concepts, so high node degree--which used as ground truth for building the model. For indicates strong connectivity to other phrases in ch training document, candidate phrases (i.e the same document-means that a candidate is n-grams)are identified and their feature values more likely to be significant are calculated as described above 6. wikipedia-based keyphraseness is the Each candidate is then marked as a positive or likelihood of a phrase bein link in the negative example, depending on whether users Wikipedia pages in which the phrase appears in document. The machine-learning model is con- the anchor text of a link by the total number of structed automatically from these labeled train- We multiply this ing examples using the WEKA machine learning number by the phrase s document frequency The new features proposed in this paper are Naive Bayes classifier, which implicitly assumes the following that the features are independent of each other 7. Spread of a phrase is the distance between given the classification. However, Kea uses only its first and last occurrences in a document. Both two or three features, whereas Maui combines values are computed relative to the length of the nine features amongst which there are many ob- document(see feature 2). High values help to vious relationships, e.g. first occurrence and determine phrases that are mentioned both in the spread, or node degree and semantic relatedness beginning and at the end of a document. Consequently, we also consider bagged decision A8. Semantic relatedness of a phrase has al- trees, which can model attribute interactions and eady been captured as the node degree(see fea- do not require parameter tuning to yield good ture 5). However, recent research allows us to results. Bagging learns an ensemble of classifiers compute semantic relatedness with better tech- and uses them in combination, thereby often niques than mere hyperlink counts. Milne and achieving significantly better results than the in- Witten(2008) propose an efficient Wikipedia dividual classifiers(Breiman, 1996). Different based approach that is nearly as accurate as hu- trees are generated by sampling from the original man subjects at quantifying the relationship be- dataset with replacement. Like Naive Bayes, tween two given concepts. Given a set of candi- bagged trees yield probability estimates that can date phrase determine the most likely Wikipedia articles that describe them (as ex- To select tags from a new document, Maui de plained in feature 5), and then determine the total termines candidate phrases and their feature val relatedness of a given phrase to all other candi- ues, and then applies the classifier built during dates. The higher the value, the more likely is the training. This classifier determines the probabil phrase to be a tag ity that a candidate is a tag based on relat 9. Inverse Wikipedia linkage is another fea- quencies observed from the training data. ture that utilizes wikipedia as a source of lan- guage usage statistics. Here, again given the 4 Evaluation most likely Wikipedia article for a given phrase, Here we describe the data used in the experi that link to it and normalize this value as inin- ments and the results obtained, addressing the verse document frequenc following questions 1. How does a state-of-the-art keyphrase ex links To(Ap) traction method perform on collaboratively tagged data, compared to a baseline auto- where links To(Ap) is the number of incoming matic tagging method? links to the article A representing the candidate 2. What is the performance of Maui with old rase P. and n is the total number of links in and new features? our Wikipedia snapshot (52M). This feature 3. How consistent are Maui's tags compared to highlights those phrases that refer to concepts those assigned by human taggers? commonly used to describe other concepts

0.22. We compute the node degree of the corre￾sponding Wikipedia article as the number of hyperlinks that connect it to other Wikipedia pages that have been identified for other candi￾date tags from the same document. A document that describes a particular topic will cover many related concepts, so high node degree—which indicates strong connectivity to other phrases in the same document—means that a candidate is more likely to be significant. 6. Wikipedia-based keyphraseness is the likelihood of a phrase being a link in the Wikipedia corpus. It divides the number of Wikipedia pages in which the phrase appears in the anchor text of a link by the total number of Wikipedia pages containing it. We multiply this number by the phrase’s document frequency. The new features proposed in this paper are the following: 7. Spread of a phrase is the distance between its first and last occurrences in a document. Both values are computed relative to the length of the document (see feature 2). High values help to determine phrases that are mentioned both in the beginning and at the end of a document. 8. Semantic relatedness of a phrase has al￾ready been captured as the node degree (see fea￾ture 5). However, recent research allows us to compute semantic relatedness with better tech￾niques than mere hyperlink counts. Milne and Witten (2008) propose an efficient Wikipedia based approach that is nearly as accurate as hu￾man subjects at quantifying the relationship be￾tween two given concepts. Given a set of candi￾date phrases we determine the most likely Wikipedia articles that describe them (as ex￾plained in feature 5), and then determine the total relatedness of a given phrase to all other candi￾dates. The higher the value, the more likely is the phrase to be a tag. 9. Inverse Wikipedia linkage is another fea￾ture that utilizes Wikipedia as a source of lan￾guage usage statistics. Here, again given the most likely Wikipedia article for a given phrase, we count the number of other Wikipedia articles that link to it and normalize this value as in in￾verse document frequency: where linksTo(AP) is the number of incoming links to the article A representing the candidate phrase P, and N is the total number of links in our Wikipedia snapshot (52M). This feature highlights those phrases that refer to concepts commonly used to describe other concepts. 3.2 Machine learning in Maui In order to build the model, we use the subset of the CiteULike collection described in Section 3.1. For each document we know a set of tags that at least two users have agreed on. This is used as ground truth for building the model. For each training document, candidate phrases (i.e. n-grams) are identified and their feature values are calculated as described above. Each candidate is then marked as a positive or negative example, depending on whether users have assigned it as a tag to the corresponding document. The machine-learning model is con￾structed automatically from these labeled train￾ing examples using the WEKA machine learning workbench. Kea (Frank et al., 1999) uses the Naïve Bayes classifier, which implicitly assumes that the features are independent of each other given the classification. However, Kea uses only two or three features, whereas Maui combines nine features amongst which there are many ob￾vious relationships, e.g. first occurrence and spread, or node degree and semantic relatedness. Consequently, we also consider bagged decision trees, which can model attribute interactions and do not require parameter tuning to yield good results. Bagging learns an ensemble of classifiers and uses them in combination, thereby often achieving significantly better results than the in￾dividual classifiers (Breiman, 1996). Different trees are generated by sampling from the original dataset with replacement. Like Naïve Bayes, bagged trees yield probability estimates that can be used to rank candidates. To select tags from a new document, Maui de￾termines candidate phrases and their feature val￾ues, and then applies the classifier built during training. This classifier determines the probabil￾ity that a candidate is a tag based on relative fre￾quencies observed from the training data. 4 Evaluation Here we describe the data used in the experi￾ments and the results obtained, addressing the following questions: 1. How does a state-of-the-art keyphrase ex￾traction method perform on collaboratively tagged data, compared to a baseline auto￾matic tagging method? 2. What is the performance of Maui with old and new features? 3. How consistent are Maui’s tags compared to those assigned by human taggers? 1322

R I Top words based on TFxIDF 16.8 17.3 17.0 1 TFXIDF 2 Top phrases based on TFxIDF 14.4 16.0 15.2 2 Ist occurrence 444 52 3Kea(TF×IDF,1ocur) 0.4 321.3 3 Keyphraseness 25.226.325.5 4 Kea(+keyphrasene 41.143.1 4 2.12.121 Table 2. Baseline auto-tagging approach vs 6 Wi 16918.3176 4.1 Evaluation method 8 Semantic relatedness The evaluation was performed using a set of 180 9 Inverse Wikipedia linkage 7.3 6.8 7.0 documents. described in Section 3. 1. each tagged Table 3. Evaluation of individual features with at least three tags on which two users have agreed. In the following, unless explicitly stated that of manual tags. Note that they only use one- otherwise, these are the only tags we use. We word tags. We evaluate this approach using our consider them to be ground truth. There are on 180 test documents and cross-validation, and average five such tags per document, and our compare the top five extracted tags with those goal is to extract tag sets that contain them all assigned manually. Comparing the first two rows We regard a predicted tag as"correct " if it of Table 2 shows that using multi-word phrase: matches one of the ground truth tags after using as candidate tags( Section 4) is less accurate than the porter stemmer. We measure performance by using single words, which gives an overall F- computing Precision(the percentage of correct Measure of 17%. Multi-words have higher extracted tags out of all extracted), Recall (the TFXIDF values, but single words are the majority percentage of correct extracted tags out of all among the users'tags. The length feature applied correct)and F-Measure(the harmonic mean of in the next section helps to capture this character the two ). Given the set east(4), network (3), istic, without compromising Maui's ability to tags, where the numbers in parenthesis show how Adding a second feature, the position of the many users have assigned each one, and the set first occurrence, and Keas Naive bayes network, metabolic, regulatory ChIP-chip, model to learn their conditional distribution, im transcription) of predicted tags, three out of five proves the results by 5 percentage points(row 3) predicted terms are correct, yielding a precision Adding the keyphraseness feature(row 4)nearly of 60%, and three out of four ground-truth terms doubles the F-Measure, from 21.3 to 42. 1%. This are extracted. a recall of 75%. The F-measure shows that CiteULike users tend to re-assign ex- combining the two values is 67% Isting tags The reported precision and recall values are averaged over all test documents. We use 10-fold 4.3 Maui with additional features cross-validation for evaluation, which allows us To evaluate Maui let us first consider the indi to use all 180 documents as test documents with- vidual performance of old and new features, as out introducing optimistic bias in the perform- shown in Table 3 Rows 1 to 3 evaluate the stan- ance measures obtained dard features used by Frank et al.(1999); Rows 4 e The results obtained in Sections 4.2 and 4.3 to 6 evaluate features that were previously used ing this evaluation provide answers to the first in Kea for controlled indexing(Medelyan et al two questions above. To answer the third we 2006) and which we have adapted in Maui for compare the indexing consistency of Maui to that free indexing. Rows 7 to 9 evaluate the three new of CiteULike users in Section 4. 4. Here, we con- features of Maui. The values can be compared to sider the assigned tag sets individually and com- keyphrase extraction by chance(F-Measure pute the consistency of Maui with each tagger as 1%)and to the multi-word TFxIDF baseline in described in Section 3. 2. We compare Maui both Table 2, row 2(F-Measure = 15.2%).The to all 332 users who tagged these documents, and strength of these features varies from 2.1to to the 36 best taggers identified in Section 3.3 25.5%(F-Measure). The strongest ones are key 4.2 Keyphrase extraction vs auto-tagging phraseness, Wikipedia keyphraseness, TFXIDI and spread As noted earlier, Brooks and Montanez(2006) Table 4 demonstrates Maui's performance automatically determine tags by extracting terms when the features are combined and shows how with the highest TFxIDF values for each post the two different classifiers, Naive Bayes(left) and argue that their quality is perhaps better than and bagged decision trees(right), compare to

4.1 Evaluation method The evaluation was performed using a set of 180 documents, described in Section 3.1, each tagged with at least three tags on which two users have agreed. In the following, unless explicitly stated otherwise, these are the only tags we use. We consider them to be ground truth. There are on average five such tags per document, and our goal is to extract tag sets that contain them all. We regard a predicted tag as “correct” if it matches one of the ground truth tags after using the Porter stemmer. We measure performance by computing Precision (the percentage of correct extracted tags out of all extracted), Recall (the percentage of correct extracted tags out of all correct) and F-Measure (the harmonic mean of the two). Given the set {yeast (4), network (3), regulation (2), metabolic (2)} of ground truth tags, where the numbers in parenthesis show how many users have assigned each one, and the set {network, metabolic, regulatory, ChIP-chip, transcription} of predicted tags, three out of five predicted terms are correct, yielding a precision of 60%, and three out of four ground-truth terms are extracted, a recall of 75%. The F-measure combining the two values is 67%. The reported precision and recall values are averaged over all test documents. We use 10-fold cross-validation for evaluation, which allows us to use all 180 documents as test documents with￾out introducing optimistic bias in the perform￾ance measures obtained. The results obtained in Sections 4.2 and 4.3 using this evaluation provide answers to the first two questions above. To answer the third we compare the indexing consistency of Maui to that of CiteULike users in Section 4.4. Here, we con￾sider the assigned tag sets individually and com￾pute the consistency of Maui with each tagger as described in Section 3.2. We compare Maui both to all 332 users who tagged these documents, and to the 36 best taggers identified in Section 3.3. 4.2 Keyphrase extraction vs. auto-tagging As noted earlier, Brooks and Montanez (2006) automatically determine tags by extracting terms with the highest TF×IDF values for each post and argue that their quality is perhaps better than that of manual tags. Note that they only use one￾word tags. We evaluate this approach using our 180 test documents and cross-validation, and compare the top five extracted tags with those assigned manually. Comparing the first two rows of Table 2 shows that using multi-word phrases as candidate tags (Section 4) is less accurate than using single words, which gives an overall F￾Measure of 17%. Multi-words have higher TF×IDF values, but single words are the majority among the users’ tags. The length feature applied in the next section helps to capture this character￾istic, without compromising Maui’s ability to assign correct multi-words tags. Adding a second feature, the position of the first occurrence, and using Kea’s Naïve Bayes model to learn their conditional distribution, im￾proves the results by 5 percentage points (row 3). Adding the keyphraseness feature (row 4) nearly doubles the F-Measure, from 21.3 to 42.1%. This shows that CiteULike users tend to re-assign ex￾isting tags. 4.3 Maui with additional features To evaluate Maui let us first consider the indi￾vidual performance of old and new features, as shown in Table 3. Rows 1 to 3 evaluate the stan￾dard features used by Frank et al. (1999); Rows 4 to 6 evaluate features that were previously used in Kea for controlled indexing (Medelyan et al., 2006) and which we have adapted in Maui for free indexing. Rows 7 to 9 evaluate the three new features of Maui. The values can be compared to keyphrase extraction by chance (F-Measure = 1%) and to the multi-word TF×IDF baseline in Table 2, row 2 (F-Measure = 15.2%). The strength of these features varies from 2.1 to 25.5% (F-Measure). The strongest ones are key￾phraseness, Wikipedia keyphraseness, TF×IDF and spread. Table 4 demonstrates Maui’s performance when the features are combined and shows how the two different classifiers, Naïve Bayes (left) and bagged decision trees (right), compare to P R F 1 Top words based on TF×IDF 16.8 17.3 17.0 2 Top phrases based on TF×IDF 14.4 16.0 15.2 3 Kea (TF×IDF, 1st occur) 20.4 22.3 21.3 4 Kea (+keyphraseness) 41.1 43.1 42.1 Table 2. Baseline auto-tagging approach vs. Kea P R F 1 TFxIDF 14.4 16 15.2 2 1st occurrence 5.4 5.4 5.4 3 Keyphraseness 25.2 26.3 25.5 4 Length 2.1 2.1 2.1 5 Node degree 8.3 9.0 8.6 6 Wikipedia keyphraseness 16.9 18.3 17.6 7 Spread 12.1 13.0 12.5 8 Semantic relatedness 7.1 7.3 7.2 9 Inverse Wikipedia linkage 7.3 6.8 7.0 Table 3. Evaluation of individual features 1323

Naive Bayes 2 Features 1-6 38.9 42.6 3 Features 1-3.7-9 39.3 41.1 46.2 44.9 376 Table 4. Combining all features in Maui each other. The baseline in row 1(left)shows the contribution of the strongest feature- Keas performance, using TFXIDF, first occur- keyphraseness-is, as expected, the highest, add rence, keyphraseness and Naive Bayes to com- ing 16.9 points. The second most important fea bine them(same as row 4 in Table 2). Using de- ture is Wikipedia keyphraseness, contributing 4 cision trees with these three features does not percentage points to the overall result improve the performance(row 1, right). The fol Since some of the features in the best perform lowing row combines the three original features ing combination rely on Wikipedia as a knowl with length, node degree and Wikipedia-based edge source, it is interesting to determine keyphraseness. In contrast to previous research Wikipedia's exact contribution. The last row of (Medelyan et al., 2008), in this setting we do not Table 5 combines the following features observe an improvement with either Naive Bayes TFXIDF, first occurrence, keyphraseness, length or bagged decision trees. In row 3 we combine and spread. The F-Measure is 5.4 points lower the three original features with the three new than that of maui with all 9 features combined ones introduced in this work. While Naive Therefore, the contribution of wikipedia-based Bayes values are lower than the baseline, with features is significant bagged decision trees Maui's F-Measure im- proves from 41.2 to 44.9%. The best results are 4.4 Maui's consistency with human taggers obtained by combining all nine features, again In Section 2.3 we discussed the indexing consis- using bagged decision trees, giving in row 4 tency of CiteULike users on our data. There are a (right)a notably improved F-Measure of 47. 1%. total of 332 taggers and their consistency with The recall of 48.6% shows that we match nearly each other is 18.5%. Now. we use results ob- half of all tags on which at least two human tag- tained with Maui during the cross-validation, when all 9 features and bagged decision trees are Given this best combination of features used(Table 4, row 4, right; see examples in Ta eliminate each feature one by one starting from ble 5), and compute how consistent Maui is with he individually weakest feature, in order to de- each human user. based on whatever document termine the contribution of each feature to this this user has tagged. Then we average the results overall result. Table 5 compares the values and to obtain the overall consistency with all 332 only bagged decision trees are used this time. users The 'Difference' column quantifies the differ Maui's consistency with the 332 human tag ence between the best F-Measure achieved with gers ranges from 0 to 80%, with an average of all features and excluding the one that is exam- 23.8%. The only cases where very low consis- ned in that row. Interestingly, one of the strong- tency was achieved are those where the human est features, TEXIDE, is the one that contributes has only assigned a few tags per document(one the least when all features are combined, while to three), or has some idiosyncratic tagging be havior (for example, one tagger adds the word Features F-Measure Difference cey in front of most tags). Still, with an average of 23. 8%, Maui's performance is over 5 points I st occu 45. 2.1 higher than that of an average CiteULike tagger -Inverse (18.5%and note this group only includes tag Semantic relatedness gers who have at least two co-taggers Node degree In Section 2.3 we were also able to determine 46.4 TFXIDF smaller group of users who perform best and wiki keyphraseness 43.1 are most prolific. This group consists of 36 tag gers whose consistency exceeds the average of the original 332 users. These 36 taggers hay Table 5 Evaluation using feature elimination tagged a total of 143 documents with an average consistency of 37.6%. Maui's consistency with

each other. The baseline in row 1 (left) shows Kea’s performance, using TF×IDF, first occur￾rence, keyphraseness and Naïve Bayes to com￾bine them (same as row 4 in Table 2). Using de￾cision trees with these three features does not improve the performance (row 1, right). The fol￾lowing row combines the three original features with length, node degree and Wikipedia-based keyphraseness. In contrast to previous research (Medelyan et al., 2008), in this setting we do not observe an improvement with either Naïve Bayes or bagged decision trees. In row 3 we combine the three original features with the three new ones introduced in this work. While Naïve Bayes’ values are lower than the baseline, with bagged decision trees Maui’s F-Measure im￾proves from 41.2 to 44.9%. The best results are obtained by combining all nine features, again using bagged decision trees, giving in row 4 (right) a notably improved F-Measure of 47.1%. The recall of 48.6% shows that we match nearly half of all tags on which at least two human tag￾gers have agreed. Given this best combination of features, we eliminate each feature one by one starting from the individually weakest feature, in order to de￾termine the contribution of each feature to this overall result. Table 5 compares the values and only bagged decision trees are used this time. The ‘Difference’ column quantifies the differ￾ence between the best F-Measure achieved with all 9 features and excluding the one that is exam￾ined in that row. Interestingly, one of the strong￾est features, TF×IDF, is the one that contributes the least when all features are combined, while the contribution of the strongest feature— keyphraseness—is, as expected, the highest, add￾ing 16.9 points. The second most important fea￾ture is Wikipedia keyphraseness, contributing 4 percentage points to the overall result. Since some of the features in the best perform￾ing combination rely on Wikipedia as a knowl￾edge source, it is interesting to determine Wikipedia’s exact contribution. The last row of Table 5 combines the following features: TF×IDF, first occurrence, keyphraseness, length and spread. The F-Measure is 5.4 points lower than that of Maui with all 9 features combined. Therefore, the contribution of Wikipedia-based features is significant. 4.4 Maui’s consistency with human taggers In Section 2.3 we discussed the indexing consis￾tency of CiteULike users on our data. There are a total of 332 taggers and their consistency with each other is 18.5%. Now, we use results ob￾tained with Maui during the cross-validation, when all 9 features and bagged decision trees are used (Table 4, row 4, right; see examples in Ta￾ble 5), and compute how consistent Maui is with each human user, based on whatever document this user has tagged. Then we average the results to obtain the overall consistency with all 332 users. Maui’s consistency with the 332 human tag￾gers ranges from 0 to 80%, with an average of 23.8%. The only cases where very low consis￾tency was achieved are those where the human has only assigned a few tags per document (one to three), or has some idiosyncratic tagging be￾havior (for example, one tagger adds the word key in front of most tags). Still, with an average of 23.8%, Maui’s performance is over 5 points higher than that of an average CiteULike tagger (18.5%)—and note this group only includes tag￾gers who have at least two co-taggers. In Section 2.3 we were also able to determine a smaller group of users who perform best and are most prolific. This group consists of 36 tag￾gers whose consistency exceeds the average of the original 332 users. These 36 taggers have tagged a total of 143 documents with an average consistency of 37.6%. Maui’s consistency with Naïve Bayes Bagged decision trees P R F P R F 1 Features 1 – 3 41.1 43.1 42.1 40.3 42.2 41.2 2 Features 1 – 6 38.9 41.1 40.0 40.3 42.6 41.4 3 Features 1 – 3, 7 – 9 39.3 41.1 40.2 43.7 46.2 44.9 4 Features 1 – 9 37.6 39.6 38.6 45.7 48.6 47.1 Table 4. Combining all features in Maui Features F-Measure Difference All 9 Features 47.1 – Length 45 2.1 – 1st occurrence 45.6 1.5 – Inverse Wikip linkage 45.1 2 – Semantic relatedness 45.4 1.7 – Node degree 46 1.1 – Spread 46.4 0.7 – TFxIDF 46.8 0.3 – Wikip keyphraseness 43.1 4 – Keyphraseness 30.2 16.9 Non-Wikip features 41.7 5.4 Table 5. Evaluation using feature elimination 1324

document 6865 Neural correlates 4. Exploring complex 353537. Computational 101.Netwo of decision variables in networks. Strogatz Nature roles for dopamine in imple building blocks arietal cortex Platt and 410, 8 (2001 behavioural control. Mon- of complex networks Glimcher Nature 400.15 tague et al. Nature 431 Milo et al. Science 298 14(2004) 324(2002) decision making compl assigned decisionmaking complexity euroscience combinatorics complex networks reinforcement learning complexity CiteULike monkey graph review motifs networks Idiosyncratic: network Idiosyncratic: brain small world action selection, attention, sub graph pattern social networks behavior. behavioral con- decision, trol, cognitive control electrophysiology, eye- learning, network, rein- datamining, data min- ovements, limitations, Idiosyncratic: 2001, adap- forcementlearning, re- onkeys, neuroecono tive systems, bistability ward. td model is,modularity, net lICs, neurons, neuro- pled oscilla ce,other, ppc, qua aphs, exploring, science,sysbio, web reinforcementlearning network biological, neu- characterization. web rons, strogatz grap Tags cortex complex networks complex networks assigned decision networks network euroscience motifs synchronization review gene visual graph reward Table 6. Tags assigned by CiteULike taggers and Maui to four sample documents these taggers ranges from 11.5% to 56%, with an Brooks and Montanez(2006) extract terms average of 35%. This places it only 2.6 percent- with the highest TFXIDF values as tags for posts age points behind the average performance of the on technorati. com. They do not report precision best CiteULike taggers. In fact, it outperforms 17 and recall values for their system, but our re- of them(cf. Table 1) implementation resulted in precision of 16.8% and recall of 17. 3% for the top five assigned 4.5 Examples tags, compared to those agreed to by at least two Table 6 compares Maui with some of CiteULike users on 180 documents. Adding CiteULike's best human taggers on four ran- eight additional features and combining them domly chosen test documents. Boldface in the using machine learning gives a clear improve taggers row indicates a tag that has been chosen ment-Maui achieves 45.7% and 48. 7% preci by at least two other human taggers; the remain- sion and recall respectively ng tags have been chosen by just one human Mishne(2006)uses TFXIDF-weighted terms Boldface in Maui's row shows tags that match as full-text queries to retrieve posts similar to the human tags. For each document Maui extracts one being analyzed. Tags assigned to these posts several tags assigned by at least two humans. are analyzed to retrieve the best ones using clus- The other tags it chooses are generally chosen by tering and heuristic ranking, tags assigned by the at least one human tagger, and even if not, they given user receive extra weight. Mishne per- are still related to the main theme of the docu- forms manual evaluation on 30 short articles and reports precision and recall for the top ten tags of 38% and 47% respectively. We matched Maui's 5 Discussion and related work op ten terms to all tags assigned to 180 docu ments automatically and obtained precision and It is possible to indirectly compare the results of recall of 44% and 29% respectively(We believe several previously published automatic tagging that manual rather than automatic evaluation approaches with Maui's. For each paper, we would be likely to give a far more favorable as compute Maui's results in settings closest to the sessment of our system. reported ones Chirita et al.(2007) gs. Given a web they first retrieve

these taggers ranges from 11.5% to 56%, with an average of 35%. This places it only 2.6 percent￾age points behind the average performance of the best CiteULike taggers. In fact, it outperforms 17 of them (cf. Table 1). 4.5 Examples Table 6 compares Maui with some of CiteULike’s best human taggers on four ran￾domly chosen test documents. Boldface in the taggers’ row indicates a tag that has been chosen by at least two other human taggers; the remain￾ing tags have been chosen by just one human. Boldface in Maui’s row shows tags that match human tags. For each document Maui extracts several tags assigned by at least two humans. The other tags it chooses are generally chosen by at least one human tagger, and even if not, they are still related to the main theme of the docu￾ment. 5 Discussion and related work It is possible to indirectly compare the results of several previously published automatic tagging approaches with Maui’s. For each paper, we compute Maui’s results in settings closest to the reported ones. Brooks and Montanez (2006) extract terms with the highest TF×IDF values as tags for posts on technorati.com. They do not report precision and recall values for their system, but our re￾implementation resulted in precision of 16.8% and recall of 17.3% for the top five assigned tags, compared to those agreed to by at least two CiteULike users on 180 documents. Adding eight additional features and combining them using machine learning gives a clear improve￾ment—Maui achieves 45.7% and 48.7% preci￾sion and recall respectively. Mishne (2006) uses TF×IDF-weighted terms as full-text queries to retrieve posts similar to the one being analyzed. Tags assigned to these posts are analyzed to retrieve the best ones using clus￾tering and heuristic ranking; tags assigned by the given user receive extra weight. Mishne per￾forms manual evaluation on 30 short articles and reports precision and recall for the top ten tags of 38% and 47% respectively. We matched Maui’s top ten terms to all tags assigned to 180 docu￾ments automatically and obtained precision and recall of 44% and 29% respectively. (We believe that manual rather than automatic evaluation would be likely to give a far more favorable as￾sessment of our system.) Chirita et al. (2007) aim to extract personal￾ized tags. Given a web page, they first retrieve Document 86865. Neural correlates of decision variables in parietal cortex. Platt and Glimcher. Nature 400,15 (1999) 44. Exploring complex networks. Strogatz. Nature 410, 8 (2001) 353537. Computational roles for dopamine in behavioural control. Mon￾tague et al. Nature 431, 14 (2004) 101. Network motifs: simple building blocks of complex networks. Milo et al. Science 298, 824 (2002) Tags assigned by CiteULike taggers decision making decisionmaking lip monkey neurophysiology reward Idiosyncratic: brain, choice, cortex, decision, electrophysiology, eye￾movements, limitations, monkeys, neuroecono￾mics, neurons, neuro￾science, other, ppc, quals, reinforcementlearning complex complexity complex networks graph networks review small world social networks survey Idiosyncratic: 2001, adap￾tive systems, bistability, coupled oscillator, graph mining, graphs, explorig, network biological, neu￾rons, strogatz dopamine neuroscience reinforcement learning review Idiosyncratic: action selection, attention, behavior, behavioral con￾trol, cognitive control, learning, network, rein￾forcementlearning, re￾ward, td model applied math combinatorics complexity motifs network original sub graph pattern Idiosyncratic: 2002, datamining, data min￾ing, graphs, link analy￾sis, modularity, net paper, patterns, protein, science, sysbio, web characterization, web graph Tags assigned by Maui cortex decision lip monkey visual complex networks networks review synchronization graph dopamine learning neuroscience review reward complex networks network motifs gene complex Table 6. Tags assigned by CiteULike taggers and Maui to four sample documents 1325

similar documents stored on the users desktop par with that achieved by the best human taggers and then determine keywords for these docu- Our results also show a substantial improvement ments. They evaluate different term scoring on an existing automatic tagging approach based techniques, such as term and document fre- on TFxIDF, and the results compare well to quency, lexical dispersion, sentence scoring, and other systems term co-occurrence. Like the Kea algorithm, the The success of automatic keyphrase extraction best formula combines term frequency with the depends primarily on the quality of the features position of the first occurrence of the term, nor- that are provided to the machine learning algo- malized by page length. It yields a precision of rithm involved In this paper we have evaluated 80% for the top four tags assigned to 30 large nine different features, including two novel websites (32Kbytes), again evaluated manually. Wikipedia-based semantic features, and found Our documents are considerably longer that their combination used in conjunction with (47Kbytes)and thus more difficult to work with, bagged decision trees produces the best perform nevertheless Maui achieves only slightly lower ance values, from 66% to 80%, when evaluating automatically against user-assigned tags. The References above caveat regarding automatic and manual assessment applies here too. Breiman, L. 1996. Bagging predictors. Machine e Budura et al.(2008)develop a scoring for- ming24(2):123-140 mula that combines three features(tag frequency, Brooks, C. H. and N. Montanez. 2006. Improved an- tag co-occurrence and document similarity )and notation of the blogosphere via autotagging and manually evaluate it on ten CiteULike docu- hierarchical clustering. In Proc. Int. Conf.on ments. Their precision for the top three to five World wide Web, Edinburgh, UK. pp. 625-632 tags ranges from 66% to 77%, slightly New York, NY, USA. ACM Press than in our paper(66% to 80%) Budura, A.,S. Michel, P. Cudre-Mauroux, and K. The only reported automatic evaluation of tags Aberer. 2008. To tag or not to tag -harvesting ad- was found in Sood et al. (2006), here Tag Assist acent metadata in large-scale tagging systems. In Proc. Int. ACM SIGIR Conf on Research and De was tested on 1000 blog posts. This algorithm is velopment in Information Retrieval, Singapore. pp similar to Mishne's(2006), but uses centroid 733-734. New York. NY. USA: ACM Press based clustering. Exact matching of Tag Assist's tags against existing ones yielded precision and Chirita, P. A.,S. Costache, W. Nejdl, and S.Hand- recall of 13. 1% and 22.8% respectively. This is Schuh. 2007. P-tag: large scale automatic genera substantially lower than Maui's 45.75% and tion of personalized annotation tags for the web 48.7% obtained with best settings(Section 4.3) Proc. Int Conf. on World wide Web, Banff, Can- ada. pp. 845-854. New York, NY, USA: ACM Note that this indirect comparison does not re Press veal the true ranking of approaches, because their task definitions and test sets are slightly differ- Frank,E, G.w. Paynter, L.H. Witten, C. Gutwin, and ent. It would be interesting to compare other sys- keyphrase extraction. In Proc. of the 16th Interna- tems on the multiple tagger set described in this paper, as we believe this would more objectively tional Joint Conference on Artificial Intelligence, Stockholm, Sweden. pp. 668-673. San Francisco, reflect the performance of humans and algo- CA: Morgan Kaufmann Publishers rithms Golder, s.A. and B. A. Huberman. 2006. Usage pat 6 Conclusions terns of collaborative tagging systems. Journal of Information Science, 32(2): 198-208 This paper has introduced a systematic way of Halpin, H, V. Robu, and H. Shepherd. 2007.The complex dynamics of collaborative tagging. In the need for manual inspection. We have shown Proc. Int Conf. on World Wide Web, pp. 211-220 how documents with multiple tag sets can New York. NY USA. ACM Press used in conjunction with a standard consistency measure to identify a robust test corpus for these Heymann, P, D. Ramage, and H. Garcia-Molina 2008. Social tag prediction. In Proc. Int. ACM techniques. Based on the evaluation methodol SIGIR Conf on Research and Development in In ogy developed, we have shown that machine- formation Retrieval, Singapore. pp. 531-538Nev learning-based automatic keyphrase extraction York, NY, USA: ACM Press produces tag sets that exhibit consistency on a

similar documents stored on the user’s desktop and then determine keywords for these docu￾ments. They evaluate different term scoring techniques, such as term and document fre￾quency, lexical dispersion, sentence scoring, and term co-occurrence. Like the Kea algorithm, the best formula combines term frequency with the position of the first occurrence of the term, nor￾malized by page length. It yields a precision of 80% for the top four tags assigned to 30 large websites (32Kbytes), again evaluated manually. Our documents are considerably longer (47Kbytes) and thus more difficult to work with, nevertheless Maui achieves only slightly lower values, from 66% to 80%, when evaluating automatically against user-assigned tags. (The above caveat regarding automatic and manual assessment applies here too.) Budura et al. (2008) develop a scoring for￾mula that combines three features (tag frequency, tag co-occurrence and document similarity) and manually evaluate it on ten CiteULike docu￾ments. Their precision for the top three to five tags ranges from 66% to 77%, slightly worse than in our paper (66% to 80%). The only reported automatic evaluation of tags was found in Sood et al. (2006), where TagAssist was tested on 1000 blog posts. This algorithm is similar to Mishne’s (2006), but uses centroid￾based clustering. Exact matching of TagAssist’s tags against existing ones yielded precision and recall of 13.1% and 22.8% respectively. This is substantially lower than Maui’s 45.75% and 48.7% obtained with best settings (Section 4.3). Note that this indirect comparison does not re￾veal the true ranking of approaches, because their task definitions and test sets are slightly differ￾ent. It would be interesting to compare other sys￾tems on the multiple tagger set described in this paper, as we believe this would more objectively reflect the performance of humans and algo￾rithms. 6 Conclusions This paper has introduced a systematic way of evaluating automatic tagging techniques without the need for manual inspection. We have shown how documents with multiple tag sets can be used in conjunction with a standard consistency measure to identify a robust test corpus for these techniques. Based on the evaluation methodol￾ogy developed, we have shown that machine￾learning-based automatic keyphrase extraction produces tag sets that exhibit consistency on a par with that achieved by the best human taggers. Our results also show a substantial improvement on an existing automatic tagging approach based on TF×IDF, and the results compare well to other systems. The success of automatic keyphrase extraction depends primarily on the quality of the features that are provided to the machine learning algo￾rithm involved. In this paper we have evaluated nine different features, including two novel Wikipedia-based semantic features, and found that their combination used in conjunction with bagged decision trees produces the best perform￾ance. References Breiman, L. 1996. Bagging predictors. Machine Learning 24(2): 123–140. Brooks, C. H. and N. Montanez. 2006. Improved an￾notation of the blogosphere via autotagging and hierarchical clustering. In Proc. Int. Conf. on World Wide Web, Edinburgh, UK. pp. 625–632. New York, NY, USA. ACM Press. Budura, A., S. Michel, P. Cudre-Mauroux, and K. Aberer. 2008. To tag or not to tag - harvesting ad￾jacent metadata in large-scale tagging systems. In Proc. Int. ACM SIGIR Conf. on Research and De￾velopment in Information Retrieval, Singapore. pp. 733–734. New York, NY, USA: ACM Press. Chirita, P. A., S. Costache, W. Nejdl, and S. Hand￾schuh. 2007. P-tag: large scale automatic genera￾tion of personalized annotation tags for the web. In Proc. Int. Conf. on World Wide Web, Banff, Can￾ada. pp. 845–854. New York, NY, USA: ACM Press. Frank, E., G. W. Paynter, I. H. Witten, C. Gutwin, and C. G. Nevill-Manning. 1999. Domain-specific keyphrase extraction. In Proc. of the 16th Interna￾tional Joint Conference on Artificial Intelligence, Stockholm, Sweden. pp. 668–673. San Francisco, CA: Morgan Kaufmann Publishers. Golder, S. A. and B. A. Huberman. 2006. Usage pat￾terns of collaborative tagging systems. Journal of Information Science, 32(2): 198–208. Halpin, H., V. Robu, and H. Shepherd. 2007. The complex dynamics of collaborative tagging. In Proc. Int. Conf. on World Wide Web, pp. 211–220. New York, NY, USA: ACM Press. Heymann, P., D. Ramage, and H. Garcia-Molina. 2008. Social tag prediction. In Proc. Int. ACM SIGIR Conf. on Research and Development in In￾formation Retrieval, Singapore. pp. 531–538. New York, NY, USA: ACM Press. 1326

Hulth, A. 2004. Combining machine learning and Rolling, L. 1981. Indexing Consistency, Quality and natural language processing for automatic key- Efficiency Information Processing Management word extraction. Ph. D. thesis, Dep. of Computer 17(2): 69-76 and Systems Sciences, Stockholm University Salton. G. and M.J. McGill. 1983. Introduction to Leonard, L.E. 1975. Inter-indexer consistency an Modern Information Retrieval. McGraw-Hill New retrieval effectiveness: measurement of relation- York ence.Univ of Illinois, Urbana-Champaign, ILCK ships. Ph. D. thesis, Grad. School of Library Scr Sigurbjornsson, B. and R. van Zwol. 2008. Flickr tag recommendation based on collective knowledge. In Leininger, K. 2000. Interindexer consistency in Psyc- Proc. Int. Conf on World Wide Web, Beijing, Info. Journal of Librarianship and Information China. pp 327-336. New York, NY, USA: ACM Science 32(1): 4-8 edelyan, O, I H Witten and D. Milne. 2008. Topic Sood, S, K. Hammond, S. Owsley, and L. Birnbau In Proc. of AAAl08 2007. Tag Assist: Automatic tag suggestion Workshop on Wikipedia and Artificial Intelligence. blog posts. of Int. Conf on Weblogs and Social an Evolving Synergy, Chicago, USA. pp. 19-24 Media, Boulder, Colorado. Menlo Park, CA Mishne, G. 2006. Autotag: a collaborative approach Turney, P. D. 2003. Coherent keyphrase extraction to automated tag assignment for weblog posts In via web mining. In Proc. of the 18th Int. Joint Proc. Int. Conf. on World Wide Web, Edinburgh, Conf on Artificial Intelligence, Acapulco, Mexico UK. pp. 953-954. New York, NY, USA. ACM pp. 434439. San Francisco, CA: Morgan Kauf- Press mann Publishers Porter, M. F. 1980. An algorithm for suffix stripping, Xu, Z, Fu,Y, Mao, J, and D. Su. 2006. Towards the Progran,14(④3):130-137 Semantic Web: Collaborative tag suggestions. In Proc, Collaborative Web Tagging Workshop at the Int. Joint Conf. on Artificial Intelligence, Stock holm. swede

Hulth, A. 2004. Combining machine learning and natural language processing for automatic key￾word extraction. Ph.D. thesis, Dep. of Computer and Systems Sciences, Stockholm University. Leonard, L. E. 1975. Inter-indexer consistency and retrieval effectiveness: measurement of relation￾ships. Ph.D. thesis, Grad. School of Library Sci￾ence, Univ. of Illinois, Urbana-Champaign, IL. Leininger, K. 2000. Interindexer consistency in Psyc￾Info. Journal of Librarianship and Information Science 32(1): 4–8. Medelyan, O., I. H. Witten and D. Milne. 2008. Topic indexing with Wikipedia. In Proc. of AAAI’08 Workshop on Wikipedia and Artificial Intelligence: an Evolving Synergy, Chicago, USA. pp. 19–24. Mishne, G. 2006. Autotag: a collaborative approach to automated tag assignment for weblog posts. In Proc. Int. Conf. on World Wide Web, Edinburgh, UK. pp. 953–954. New York, NY, USA. ACM Press Porter, M. F. 1980. An algorithm for suffix stripping, Program, 14(3): 130−137. Rolling, L. 1981. Indexing Consistency, Quality and Efficiency. Information Processing & Management 17(2): 69–76. Salton, G. and M. J. McGill. 1983. Introduction to Modern Information Retrieval. McGraw-Hill New York. Sigurbjörnsson, B. and R. van Zwol. 2008. Flickr tag recommendation based on collective knowledge. In Proc. Int. Conf. on World Wide Web, Beijing, China. pp. 327–336. New York, NY, USA: ACM Press. Sood, S., K. Hammond, S. Owsley, and L. Birnbaum. 2007. TagAssist: Automatic tag suggestion for blog posts. of Int. Conf. on Weblogs and Social Media, Boulder, Colorado. Menlo Park, CA. Turney, P. D. 2003. Coherent keyphrase extraction via web mining. In Proc. of the 18th Int. Joint Conf. on Artificial Intelligence, Acapulco, Mexico. pp. 434–439. San Francisco, CA: Morgan Kauf￾mann Publishers. Xu, Z., Fu, Y., Mao, J., and D. Su. 2006. Towards the Semantic Web: Collaborative tag suggestions. In Proc. Collaborative Web Tagging Workshop at the Int. Joint Conf. on Artificial Intelligence, Stock￾holm, Sweden. 1327