500 - PDF

Document Sample
500 - PDF Powered By Docstoc
					                  Entity Disambiguation for Knowledge Base Population
 †Mark Dredze and †Paul McNamee and †Delip Rao and †Adam Gerber and Tim Finin
†Human Language Technology Center of Excellence, Center for Language and Speech Processing
                               Johns Hopkins University
                        University of Maryland, Baltimore County
                           Abstract                                 cluster corresponds to a single real world entity.
  The integration of facts derived from information extraction
                                                                       The emergence of large scale publicly avail-
  systems into existing knowledge bases requires a system to        able KBs like Wikipedia and DBPedia has spurred
  disambiguate entity mentions in the text. This is challeng-       an interest in linking textual entity references to
  ing due to issues such as non-uniform variations in entity
  names, mention ambiguity, and entities absent from a knowl-       their entries in these public KBs. Bunescu and
  edge base. We present a state of the art system for entity dis-   Pasca (2006) and Cucerzan (2007) presented im-
  ambiguation that not only addresses these challenges but also     portant pioneering work in this area, but suffer
  scales to knowledge bases with several million entries using
  very little resources. Further, our approach achieves perfor-     from several limitations including Wikipedia spe-
  mance of up to 95% on entities mentioned from newswire            cific dependencies, scale, and the assumption of
  and 80% on a public test set that was designed to include         a KB entry for each entity. In this work we in-
  challenging queries.
                                                                    troduce an entity disambiguation system for link-
                                                                    ing entities to corresponding Wikipedia pages de-
  1    Introduction                                                 signed for open domains, where a large percent-
  The ability to identify entities like people, orga-               age of entities will not be linkable. Further, our
  nizations and geographic locations (Tjong Kim                     method and some of our features readily general-
  Sang and De Meulder, 2003), extract their at-                     ize to other curated KB. We adopt a supervised
  tributes (Pasca, 2008), and identify entity rela-                 approach, where each of the possible entities con-
  tions (Banko and Etzioni, 2008) is useful for sev-                tained within Wikipedia are scored for a match to
  eral applications in natural language processing                  the query entity. We also describe techniques to
  and knowledge acquisition tasks like populating                   deal with large knowledge bases, like Wikipedia,
  structured knowledge bases (KB).                                  which contain millions of entries. Furthermore,
     However, inserting extracted knowledge into a                  our system learns when to withhold a link when
  KB is fraught with challenges arising from nat-                   an entity has no matching KB entry, a task that
  ural language ambiguity, textual inconsistencies,                 has largely been neglected in prior research in
  and lack of world knowledge. To the discern-                      cross-document entity coreference. Our system
  ing human eye, the “Bush” in “Mr. Bush left                       produces high quality predictions compared with
  for the Zurich environment summit in Air Force                    recent work on this task.
  One.” is clearly the US president. Further con-
                                                                    2   Related Work
  text may reveal it to be the 43rd president, George
  W. Bush, and not the 41st president, George H.                    The information extraction oeuvre has a gamut of
  W. Bush. The ability to disambiguate a polyse-                    relation extraction methods for entities like per-
  mous entity mention or infer that two orthograph-                 sons, organizations, and locations, which can be
  ically different mentions are the same entity is                  classified as open- or closed-domain depending
  crucial in updating an entity’s KB record. This                   on the restrictions on extractable relations (Banko
  task has been variously called entity disambigua-                 and Etzioni, 2008). Closed domain systems ex-
  tion, record linkage, or entity linking. When per-                tract a fixed set of relations while in open-domain
  formed without a KB, entity disambiguation is                     systems, the number and type of relations are un-
  called coreference resolution: entity mentions ei-                bounded. Extracted relations still require process-
  ther within the same document or across multi-                    ing before they can populate a KB with facts:
  ple documents are clustered together, where each                  namely, entity linking and disambiguation.
   Motivated by ambiguity in personal name              whether their top prediction is correct, or whether
search, Mann and Yarowsky (2003) disambiguate           NIL should be output. We believe relying on fea-
person names using biographic facts, like birth         tures that are designed to inform whether absence
year, occupation and affiliation. When present           is correct is the better alternative.
in text, biographic facts extracted using regular
expressions help disambiguation. More recently,         3     Entity Linking
the Web People Search Task (Artiles et al., 2008)
                                                        We define entity linking as matching a textual en-
clustered web pages for entity disambiguation.
                                                        tity mention, possibly identified by a named en-
   The related task of cross document corefer-          tity recognizer, to a KB entry, such as a Wikipedia
ence resolution has been addressed by several           page that is a canonical entry for that entity. An
researchers starting from Bagga and Baldwin             entity linking query is a request to link a textual
(1998). Poesio et al. (2008) built a cross document     entity mention in a given document to an entry in
coreference system using features from encyclo-         a KB. The system can either return a matching en-
pedic sources like Wikipedia. However, success-         try or NIL to indicate there is no matching entry.
ful coreference resolution is insufficient for cor-      In this work we focus on linking organizations,
rect entity linking, as the coreference chain must      geo-political entities and persons to a Wikipedia
still be correctly mapped to the proper KB entry.       derived KB.
   Previous work by Bunescu and Pasca (2006)
and Cucerzan (2007) aims to link entity men-            3.1    Key Issues
tions to their corresponding topic pages in             There are 3 challenges to entity linking:
Wikipedia but the authors differ in their ap-
proaches. Cucerzan uses heuristic rules and             Name Variations. An entity often has multiple
Wikipedia disambiguation markup to derive map-          mention forms, including abbreviations (Boston
pings from surface forms of entities to their           Symphony Orchestra vs. BSO), shortened forms
Wikipedia entries. For each entity in Wikipedia,        (Osama Bin Laden vs. Bin Laden), alternate
a context vector is derived as a prototype for the      spellings (Osama vs. Ussamah vs. Oussama),
entity and these vectors are compared (via dot-         and aliases (Osama Bin Laden vs. Sheikh Al-
product) with the context vectors of unknown en-        Mujahid). Entity linking must find an entry de-
tity mentions. His work assumes that all entities       spite changes in the mention string.
have a corresponding Wikipedia entry, but this as-      Entity Ambiguity. A single mention, like
sumption fails for a significant number of entities      Springfield, can match multiple KB entries, as
in news articles and even more for other genres,        many entity names, like people and organizations,
like blogs. Bunescu and Pasca on the other hand         tend to be polysemous.
suggest a simple method to handle entities not in       Absence. Processing large text collections vir-
Wikipedia by learning a threshold to decide if the      tually guarantees that many entities will not ap-
entity is not in Wikipedia. Both works mentioned        pear in the KB (NIL), even for large KBs.
rely on Wikipedia-specific annotations, such as             The combination of these challenges makes
category hierarchies and disambiguation links.          entity linking especially challenging. Consider
   We just recently became aware of a system            an example of “William Clinton.” Most read-
fielded by Li et al. at the TAC-KBP 2009 eval-           ers will immediately think of the 42nd US pres-
uation (2009). Their approach bears a number            ident. However, the only two William Clintons in
of similarities to ours; both systems create candi-     Wikipedia are “William de Clinton” the 1st Earl
date sets and then rank possibilities using differing   of Huntingdon, and “William Henry Clinton” the
learning methods, but the principal difference is in    British general. The page for the 42nd US pres-
our approach to NIL prediction. Where we simply         ident is actually “Bill Clinton”. An entity link-
consider absence (i.e., the NIL candidate) as an-       ing system must decide if either of the William
other entry to rank, and select the top-ranked op-      Clintons are correct, even though neither are ex-
tion, they use a separate binary classifier to decide    act matches. If the system determines neither
matches, should it return NIL or the variant “Bill    (Bunescu and Pasca, 2006) – which is limited to
Clinton”? If variants are acceptable, then perhaps    Wikipedia and does not work for general KBs.
“Clinton, Iowa” or “DeWitt Clinton” should be         We consider a KB independent approach to selec-
acceptable answers?                                   tion that also allows for tuning candidate set size.
                                                      This involves a linear pass over KB entry names
3.2       Contributions                               (Wikipedia page titles): a naive implementation
We address these entity linking challenges.           took two minutes per query. The following sec-
Robust Candidate Selection. Our system is             tion reduces this to under two seconds per query.
flexible enough to find name variants but suffi-            For a given query, the system selects KB entries
ciently restrictive to produce a manageable can-      using the following approach:
didate list despite a large-scale KB.
Features for Entity Disambiguation. We de-
                                                      • Titles that are exact matches for the mention.
veloped a rich and extensible set of features based
on the entity mention, the source document, and       • Titles that are wholly contained in or contain
the KB entry. We use a machine learning ranker        the mention (e.g., Nationwide and Nationwide In-
to score each candidate.                              surance).
Learning NILs. We modify the ranker to learn
                                                      • The first letters of the entity mention match the
NIL predictions, which obviates hand tuning and
                                                      KB entry title (e.g., OA and Olympic Airlines).
importantly, admits use of additional features that
are indicative of NIL.                                • The title matches a known alias for the entity
   Our contributions differ from previous efforts     (aliases described in Section 5.2).
(Bunescu and Pasca, 2006; Cucerzan, 2007) in          • The title has a strong string similarity score
several important ways. First, previous efforts de-   with the entity mention. We include several mea-
pend on Wikipedia markup for significant perfor-       sures of string similarity, including: character
mance gains. We make no such assumptions, al-         Dice score > 0.9, skip bigram Dice score > 0.6,
though we show that optional Wikipedia features       and Hamming distance <= 2.
lead to a slight improvement. Second, Cucerzan
does not handle NILs while Bunescu and Pasca
                                                         We did not optimize the thresholds for string
address them by learning a threshold. Our ap-
                                                      similarity, but these could obviously be tuned to
proach learns to predict NIL in a more general
                                                      minimize the candidate sets and maximize recall.
and direct way. Third, we develop a rich fea-
                                                         All of the above features are general for any
ture set for entity linking that can work with any
                                                      KB. However, since our evaluation used a KB
KB. Finally, we apply a novel finite state machine
                                                      derived from Wikipedia, we included a few
method for learning name variations. 1
                                                      Wikipedia specific features. We added an entry if
   The remaining sections describe the candidate
                                                      its Wikipedia page appeared in the top 20 Google
selection system, features and ranking, and our
                                                      results for a query.
novel approach learning NILs, followed by an
empirical evaluation.                                    On the training dataset (Section 7) the selection
                                                      system attained a recall of 98.8% and produced
4       Candidate Selection for Name Variants         candidate lists that were three to four orders of
                                                      magnitude smaller than the KB. Some recall er-
The first system component addresses the chal-         rors were due to inexact acronyms: ABC (Arab
lenge of name variants. As the KB contains a large    Banking; ‘Corporation’ is missing), ASG (Abu
number of entries (818,000 entities, of which 35%     Sayyaf; ‘Group’ is missing), and PCF (French
are PER, ORG or GPE), we require an efficient se-      Communist Party; French reverses the order of the
lection of the relevant candidates for a query.       pre-nominal adjectives). We also missed Interna-
   Previous approaches used Wikipedia markup          tional Police (Interpol) and Becks (David Beck-
for filtering – only using the top-k page categories   ham; Mr. Beckham and his wife are collectively
    1 markus/fstrain       referred to as ‘Posh and Becks’).
4.1      Scaling Candidate Selection                        constraint is equivalent to the ranking SVM algo-
                                                            rithm of Joachims (2002), where we define an or-
Our previously described candidate selection re-
                                                            dered pair constraint for each of the incorrect KB
lied on a linear pass over the KB, but we seek
                                                            entries y and the correct entry y. Training sets pa-
more efficient methods. We observed that the
                                                            rameters such that score(y) ≥ score(ˆ) + γ. We
above non-string similarity filters can be pre-
                                                            used the library SVMrank to solve this optimiza-
computed and stored in an index, and that the skip
                                                            tion problem.3 We used a linear kernel, set the
bigram Dice score can be computed by indexing
                                                            slack parameter C as 0.01 times the number of
the skip bigrams for each KB title. We omitted
                                                            training examples, and take the loss function as
the other string similarity scores, and collectively
                                                            the total number of swapped pairs summed over
these changes enable us to avoid a linear pass over
                                                            all training examples. While previous work used
the KB. Finally we obtained speedups by serving
                                                            a custom kernel, we found a linear kernel just as
the KB concurrently2 . Recall was nearly identical
                                                            effective with our features. This has the advan-
to the full system described above: only two more
                                                            tage of efficiency in both training and prediction 4
queries failed. Additionally, more than 95% of
                                                            – important considerations in a system meant to
the processing time was consumed by Dice score
                                                            scale to millions of KB entries.
computation, which was only required to cor-
rectly retrieve less than 4% of the training queries.       5.1      Features for Entity Disambiguation
Omitting the Dice computation yielded results in
                                                            200 atomic features represent x based on each
a few milliseconds. A related approach is that of
                                                            candidate query/KB pair. Since we used a lin-
canopies for scaling clustering for large amounts
                                                            ear kernel, we explicitly combined certain fea-
of bibliographic citations (McCallum et al., 2000).
                                                            tures (e.g., acroynym-match AND known-alias) to
In contrast, our setting focuses on alignment vs.
                                                            model correlations. This included combining each
clustering mentions, for which overlapping parti-
                                                            feature with the predicted type of the entity, al-
tioning approaches like canopies are applicable.
                                                            lowing the algorithm to learn prediction functions
5       Entity Linking as Ranking                           specific to each entity type. With feature combina-
                                                            tions, the total number of features grew to 26,569.
We select a single correct candidate for a query            The next sections provide an overview; for a de-
using a supervised machine learning ranker. We              tailed list see McNamee et al. (2009).
represent each query by a D dimensional vector
x, where x ∈ RD , and we aim to select a sin-               5.2      Features for Name Variants
gle KB entry y, where y ∈ Y, a set of possible              Variation in entity name has long been recog-
KB entries for this query produced by the selec-            nized as a bane for information extraction sys-
tion system above, which ensures that Y is small.           tems. Poor handling of entity name variants re-
The ith query is given by the pair {xi , yi }, where        sults in low recall. We describe several features
we assume at most one correct KB entry.                     ranging from simple string match to finite state
   To evaluate each candidate KB entry in Y we              transducer matching.
create feature functions of the form f (x, y), de-          String Equality. If the query name and KB en-
pendent on both the example x (document and en-             try name are identical, this is a strong indication of
tity mention) and the KB entry y. The features              a match, and in our KB entry names are distinct.
address name variants and entity disambiguation.            However, similar or identical entry names that
   We take a maximum margin approach to learn-              refer to distinct entities are often qualified with
ing: the correct KB entry y should receive a                parenthetical expressions or short clauses. As
higher score than all other possible KB entries             an example, “London, Kentucky” is distinguished
y ∈ Y, y = y plus some margin γ. This learning
ˆ       ˆ                                                      3
                                                          light/svm rank.html
                                                                 Bunescu and Pasca (2006) report learning tens of thou-
    Our Python implementation with indexing features and    sands of support vectors with their “taxonomy” kernel while
four threads achieved up to 80× speedup compared to naive   a linear kernel represents all support vectors with a single
implementation.                                             weight vector, enabling faster training and prediction.
from “London, Ontario”, “London, Arkansas”,                         contained character n-grams; we used n-grams of
“London (novel)”, and “London”. Therefore,                          length 3 and less. The scores are combined using a
other string equality features were used, such as                   global log-linear model. Since different spellings
whether names are equivalent after some transfor-                   of a name may vary considerably in length (e.g.,
mation. For example, “Baltimore” and “Baltimore                     J Miller vs. Jennifer Miller) we eliminated the
City” are exact matches after removing a common                     limit on consecutive insertions used in previous
GPE word like city; “University of Vermont” and                     applications.6
“University of VT” match if VT is expanded.
Approximate String Matching. Many entity                            5.3    Wikipedia Features
mentions will not match full names exactly. We                      Most of our features do not depend on Wikipedia
added features for character Dice, skip bigram                      markup, but it is reasonable to include features
Dice, and left and right Hamming distance scores.                   from KB properties. Our feature ablation study
Features were set based on quantized scores.                        shows that dropping these features causes a small
These were useful for detecting minor spelling                      but statistically significant performance drop.
variations or mistakes. Features were also added if
the query was wholly contained in the entry name,                   WikiGraph statistics. We added features de-
or vice-versa, which was useful for handling ellip-                 rived from the Wikipedia graph structure for an
sis (e.g., “United States Department of Agricul-                    entry, like indegree of a node, outdegree of a node,
ture” vs. “Department of Agriculture”). We also                     and Wikipedia page length in bytes. These statis-
included the ratio of the recursive longest com-                    tics favor common entity mentions over rare ones.
mon subsequence (Christen, 2006) to the shorter                     Wikitology. KB entries can be indexed with hu-
of the mention or entry name, which is effective at                 man or machine generated metadata consisting of
handling some deletions or word reorderings (e.g.,                  keywords or categories in a domain-appropriate
“Li Gong” and “Gong Li”). Finally, we checked                       taxonomy. Using a system called Wikitology,
whether all of the letters of the query are found in                Syed et al. (2008) investigated use of ontology
the same order in the entry name (e.g., “Univ Wis-                  terms obtained from the explicit category system
consin” would match “University of Wisconsin”).                     in Wikipedia as well as relationships induced from
Acronyms. Features for acronyms, using dic-                         the hyperlink graph between related Wikipedia
tionaries and partial character matches, enable                     pages. Following this approach we computed top-
matches between “MIT” and “Madras Institute of                      ranked categories for the query documents and
Technology” or “Ministry of Industry and Trade.”                    used this information as features. If none of the
Aliases. Many aliases or nicknames are non-                         candidate KB entries had corresponding highly-
trivial to guess.      For example JAVA is the                      ranked Wikitology pages, we used this as a NIL
stock symbol for Sun Microsystems, and “Gin-                        feature (Section 6.1).
ger Spice” is a stage name of Geri Halliwell. A
reasonable way to do this is to employ a dictio-                    5.4    Popularity
nary and alias lists that are commonly available                    Although it may be an unsafe bias to give prefer-
for many domains5 .                                                 ence to common entities, we find it helpful to pro-
FST Name Matching. Another measure of sur-                          vide estimates of entity popularity to our ranker
face similarity between a query and a candidate                     as others have done (Fader et al., 2009). Apart
was computed by training finite-state transducers                    from the graph-theoretic features derived from the
similar to those described in Dreyer et al. (2008).                 Wikipedia graph, we used Google’s PageRank to
These transducers assign a score to any string pair                 by adding features indicating the rank of the KB
by summing over all alignments and scoring all                      entry’s corresponding Wikipedia page in a Google
     We used multiple lists, including class-specific lists (i.e.,   query for the target entity mention.
for PER, ORG, and GPE) lists extracted from Freebase (Bol-
lacker et al., 2008) and Wikipedia redirects. PER, ORG, and              Without such a limit, the objective function may diverge
GPE are the commonly used terms for entity types for peo-           for certain parameters of the model; we detect such cases and
ple, organizations and geo-political regions respectively.          learn to avoid them during training.
5.5   Document Features                              cascaded to allow arbitrary feature conjunctions.
The mention document and text associated with a      Thus it is possible to end up with a feature kbtype-
KB entry contain context for resolving ambiguity.    is-ORG AND high-TFIDF-score AND low-name-
                                                     similarity. The combined features increased the
Entity Mentions. Some features were based on
                                                     number of features from roughly 200 to 26,000.
presence of names in the text: whether the query
appeared in the KB text and the entry name in the
                                                     6     Predicting NIL Mentions
document. Additionally, we used a named-entity
tagger and relation finder, SERIF (Boschee et al.,    So far we have assumed that each example has a
2005), identified name and nominal mentions that      correct KB entry; however, when run over a large
were deemed co-referent with the entity mention      corpus, such as news articles, we expect a signifi-
in the document, and tested whether these nouns      cant number of entities will not appear in the KB.
were present in the KB text. Without the NE anal-    Hence it will be useful to predict NILs.
ysis, accuracy on non-NIL entities dropped 4.5%.        We learn when to predict NIL using the SVM
KB Facts. KB nodes contain infobox attributes        ranker by augmenting Y to include NIL, which
(or facts); we tested whether the fact text was      then has a single feature unique to NIL answers.
present in the query document, both locally to a     It can be shown that (modulo slack variables) this
mention, or anywhere in the text. Although these     is equivalent to learning a single threshold τ for
facts were derived from Wikipedia infoboxes,         NIL predictions as in Bunescu and Pasca (2006).
they could be obtained from other sources as well.      Incorporating NIL into the ranker has several
Document Similarity We measured similarity           advantages. First, the ranker can set the thresh-
between the query document and the KB text in        old optimally without hand tuning. Second, since
two ways: cosine similarity with TF/IDF weight-      the SVM scores are relative within a single exam-
ing (Salton and McGill, 1983); and using the Dice    ple and cannot be compared across examples, set-
coefficient over bags of words. IDF values were       ting a single threshold is difficult. Third, a thresh-
approximated using counts from the Google 5-         old sets a uniform standard across all examples,
gram dataset as by Klein and Nelson (2008).          whereas in practice we may have reasons to favor
Entity Types. Since the KB contained types           a NIL prediction in a given example. We design
for entries, we used these as features as well as    features for NIL prediction that cannot be cap-
the predicted NE type for the entity mention in      tured in a single parameter.
the document text. Additionally, since only a
small number of KB entries had PER, ORG, or          6.1    NIL Features
GPE types, we also inferred types from Infobox       Integrating NIL prediction into learning means
class information to attain 87% coverage in the      we can define arbitrary features indicative of NIL
KB. This was helpful for discouraging selection      predictions in the feature vector corresponding to
of eponymous entries named after famous enti-        NIL. For example, if many candidates have good
ties (e.g., the former U.S. president vs. “John F.   name matches, it is likely that one of them is cor-
Kennedy International Airport”).                     rect. Conversely, if no candidate has high entry-
                                                     text/article similarity, or overlap between facts
5.6   Feature Combinations                           and the article text, it is likely that the entity is
To take into account feature dependencies we cre-    absent from the KB. We included several features,
ated combination features by taking the cross-       such as a) the max, mean, and difference between
product of a small set of diverse features. The      max and mean for 7 atomic features for all KB
attributes used as combination features included     candidates considered, b) whether any of the can-
entity type; a popularity based on Google’s rank-    didate entries have matching names (exact and
ings; document comparison using TF/IDF; cov-         fuzzy string matching), c) whether any KB en-
erage of co-referential nouns in the KB node         try was a top Wikitology match, and d) if the top
text; and name similarity. The combinations were     Google match was not a candidate.
                                   Micro-Averaged                                     Macro-Averaged
                    Best     Median All Features     Best Features      Best     Median All Features        Best Features
      All          0.8217    0.7108      0.7984         0.7941         0.7704    0.6861     0.7695             0.7704
      non-NIL      0.7725    0.6352      0.7063         0.6639         0.6696    0.5335     0.6097             0.5593
      NIL          0.8919    0.7891      0.8677         0.8919         0.8789    0.7446     0.8464             0.8721

Table 1: Micro and macro-averaged accuracy for TAC-KBP data compared to best and median reported performance.
Results are shown for all features as well as removing a small number of features using feature selection on development data.

7       Evaluation                                              PERs and ORGs these percentages were signifi-
                                                                cantly lower, 19% and 30% respectively.
We evaluated our system on two datasets: the                       Table 1 shows results on TAC-KBP data us-
Text Analysis Conference (TAC) track on Knowl-                  ing all of our features as well a subset of features
edge Base Population (TAC-KBP) (McNamee and                     based on feature selection experiments on devel-
Dang, 2009) and the newswire data used by                       opment data. We include scores for both micro-
Cucerzan (2007) (Microsoft News Data).                          averaged accuracy – averaged over all queries
   Since our approach relies on supervised learn-               – and macro-averaged accuracy – averaged over
ing, we begin by constructing our own training                  each unique entity – as well as the best and me-
corpus.7 We highlighted 1496 named entity men-                  dian reported results for these data (McNamee
tions in news documents (from the TAC-KBP doc-                  and Dang, 2009). We obtained the best reported
ument collection) and linked these to entries in                results for macro-averaged accuracy, as well as
a KB derived from Wikipedia infoboxes. 8 We                     the best results for NIL detection with micro-
added to this collection 119 sample queries from                averaged accuracy, which shows the advantage of
the TAC-KBP data. The total of 1615 training ex-                our approach to learning NIL. See McNamee et
amples included 539 (33.4%) PER, 618 (38.3%)                    al. (2009) for additional experiments.
ORG, and 458 (28.4%) GPE entity mentions. Of
                                                                   The candidate selection phase obtained a re-
the training examples, 80.5% were found in the
                                                                call of 98.6%, similar to that of development data.
KB, matching 300 unique entities. This set has a
                                                                Missed candidates included Iron Lady, which
higher number of NIL entities than did Bunescu
                                                                refers metaphorically to Yulia Tymoshenko, PCC,
and Pasca (2006) (10%) but lower than the TAC-
                                                                the Spanish-origin acronym for the Cuban Com-
KBP test set (43%).
                                                                munist Party, and Queen City, a former nickname
   All system development was done using a train                for the city of Seattle, Washington. The system re-
(908 examples) and development (707 examples)                   turned a mean of 76 candidates per query, but the
split. The TAC-KBP and Microsoft News data                      median was 15 and the maximum 2772 (Texas). In
sets were held out for final tests. A model trained              about 10% of cases there were four or fewer can-
on all 1615 examples was used for experiments.                  didates and in 10% of cases there were more than
                                                                100 candidate KB nodes. We observed that ORGs
7.1       TAC-KBP 2009 Experiments
                                                                were more difficult, due to the greater variation
The KB is derived from English Wikipedia pages                  and complexity in their naming, and that they can
that contained an infobox. Entries contain basic                be named after persons or locations.
descriptions (article text) and attributes. The TAC-
KBP query set contains 3904 entity mentions for                 7.2    Feature Effectiveness
560 distinct entities; entity type was only provided
                                                                We performed two feature analyses on the TAC-
for evaluation. The majority of queries were for
                                                                KBP data: an additive study – starting from a
organizations (69%). Most queries were missing
                                                                small baseline feature set used in candidate selec-
from the KB (57%). 77% of the distinct GPEs
                                                                tion we add feature groups and measure perfor-
in the queries were present in the KB, but for
                                                                mance changes (omitting feature combinations),
        Data available from                      and an ablative study – starting from all features,
    8               remove a feature group and measure performance.
 Class                                      All          non-NIL           NIL                   Num. Queries               Accuracy
 Baseline                                 0.7264          0.4621          0.9251                 Total    Nil       All     non-NIL       NIL
 Acronyms                                 0.7316          0.4860          0.9161         NIL      452    187      0.4137       0.0         1.0
 NE Analysis                              0.7661          0.7181          0.8022         GPE      132     20      0.9696      1.00       0.8000
 Google                                   0.7597          0.7421          0.7730         ORG      115     45      0.8348     0.7286       1.00
 Doc/KB Text Similarity                   0.7313          0.6699          0.7775         PER      205    122      0.9951     0.9880       1.00
 Wikitology                               0.7318          0.4549          0.9399         All      452    187      0.9469     0.9245      0.9786
 All                                      0.7984          0.7063          0.8677            Cucerzan (2007)       0.914         -           -

  Table 2: Additive analysis: micro-averaged accuracy.                                    Table 3: Micro-average results for Microsoft data.

   Table 2 shows the most significant features in                                     the number of candidates considered. We selected
the feature addition experiments. The baseline                                       a median of 13 candidates for PER, 12 for ORG
includes only features based on string similarity                                    and 102 for GPE. Accuracy results are in Table
or aliases and is not effective at finding correct                                    3. The high results reported for this dataset over
entries and strongly favors NIL predictions. In-                                     TAC-KBP is primarily because we perform very
clusion of features based on analysis of named-                                      well in predicting popular and rare entries – both
entities, popularity measures (e.g., Google rank-                                    of which are common in newswire text.
ings), and text comparisons provided the largest                                         One issue with our KB was that it was derived
gains. The overall changes are fairly small,                                         from infoboxes in Wikipedia’s Oct 2008 version
roughly ±1%; however changes in non-NIL pre-                                         which has both new entities, 12 and is missing en-
cision are larger.                                                                   tities.13 Therefore, we manually confirmed NIL
   The ablation study showed considerable redun-                                     answers and new answers for queries marked as
dancy across feature groupings. In several cases,                                    NIL in the data. While an exact comparison is not
performance could have been slightly improved                                        possible (as described above), our results (94.7%)
by removing features. Removing all feature com-                                      appear to be at least on par with Cucerzan’s sys-
binations would have improved overall perfor-                                        tem (91.4% overall accuracy).With the strong re-
mance to 81.05% by gaining on non-NIL for a                                          sults on TAC-KBP, we believe that this is strong
small decline on NIL detection.                                                      confirmation of the effectiveness of our approach.

7.3       Experiments on Microsoft News Data                                         8     Conclusion
We downloaded the evaluation data used in                                            We presented a state of the art system to disam-
Cucerzan (2007)9 : 20 news stories from MSNBC                                        biguate entity mentions in text and link them to
with 642 entity mentions manually linked to                                          a knowledge base. Unlike previous approaches,
Wikipedia and another 113 mentions not having                                        our approach readily ports to KBs other than
any corresponding link to Wikipedia.10 A sig-                                        Wikipedia. We described several important chal-
nificant percentage of queries were not of type                                       lenges in the entity linking task including han-
PER, ORG, or GPE (e.g., “Christmas”). SERIF                                          dling variations in entity names, ambiguity in en-
assigned entity types and we removed 297 queries                                     tity mentions, and missing entities in the KB, and
not recognized as entities (counts in Table 3).                                      we showed how to each of these can be addressed.
   We learned a new model on the training data                                       We described a comprehensive feature set to ac-
above using a reduced feature set to increase                                        complish this task in a supervised setting. Impor-
speed.11 Using our fast candidate selection sys-                                     tantly, our method discriminately learns when not
tem, we resolved each query in 1.98 seconds (me-                                     to link with high accuracy. To spur further re-
dian). Query processing time was proportional to                                     search in these areas we are releasing our entity
                                                                                     linking system.
  10                                                                                   12
      One of the MSNBC news articles is no longer available                               2008 vs. 2006 version used in Cucerzan (2007) We
so we used 759 total entities.                                                       could not get the 2006 version from the author or the Internet.
   11                                                                                  13
      We removed Google, FST and conjunction features                                     Since our KB was derived from infoboxes, entities not
which reduced system accuracy but increased performance.                             having an infobox were left out.
References                                               Gideon S. Mann and David Yarowsky. 2003. Unsuper-
                                                           vised personal name disambiguation. In Conference
Javier Artiles, Satoshi Sekine, and Julio Gonzalo.         on Natural Language Learning (CONLL).
   2008. Web people search: results of the first evalu-
   ation and the plan for the second. In WWW.            Andrew McCallum, Kamal Nigam, and Lyle Ungar.
                                                           2000. Efficient clustering of high-dimensional data
Amit Bagga and Breck Baldwin. 1998. Entity-
                                                           sets with application to reference matching. In
 based cross-document coreferencing using the vec-
                                                           Knowledge Discovery and Data Mining (KDD).
 tor space model. In Conference on Computational
 Linguistics (COLING).                                   Paul McNamee and Hoa Trang Dang. 2009. Overview
Michele Banko and Oren Etzioni. 2008. The tradeoffs        of the TAC 2009 knowledge base population track.
  between open and traditional relation extraction. In     In Text Analysis Conference (TAC).
  Association for Computational Linguistics.             Paul McNamee, Mark Dredze, Adam Gerber, Nikesh
K. Bollacker, C. Evans, P. Paritosh, T. Sturge, and        Garera, Tim Finin, James Mayfield, Christine Pi-
  J. Taylor. 2008. Freebase: a collaboratively cre-        atko, Delip Rao, David Yarowsky, and Markus
  ated graph database for structuring human knowl-         Dreyer. 2009. HLTCOE approaches to knowledge
  edge. In SIGMOD Management of Data.                      base population at TAC 2009. In Text Analysis Con-
                                                           ference (TAC).
E. Boschee, R. Weischedel, and A. Zamanian. 2005.
   Automatic information extraction. In Conference       Marius Pasca. 2008. Turning web text and search
   on Intelligence Analysis.                              queries into factual knowledge: hierarchical class
                                                          attribute extraction. In National Conference on Ar-
Razvan C. Bunescu and Marius Pasca. 2006. Using           tificial Intelligence (AAAI).
  encyclopedic knowledge for named entity disam-
  biguation. In European Chapter of the Assocation       Massimo Poesio, David Day, Ron Artstein, Jason Dun-
  for Computational Linguistics (EACL).                   can, Vladimir Eidelman, Claudio Giuliano, Rob
                                                          Hall, Janet Hitzeman, Alan Jern, Mijail Kabadjov,
Peter Christen. 2006. A comparison of personal name       Stanley Yong, Wai Keong, Gideon Mann, Alessan-
  matching: Techniques and practical issues. Techni-      dro Moschitti, Simone Ponzetto, Jason Smith, Josef
  cal Report TR-CS-06-02, Australian National Uni-        Steinberger, Michael Strube, Jian Su, Yannick Ver-
  versity.                                                sley, Xiaofeng Yang, and Michael Wick. 2008. Ex-
                                                          ploiting lexical and encyclopedic resources for en-
Silviu Cucerzan. 2007. Large-scale named entity           tity disambiguation: Final report. Technical report,
   disambiguation based on wikipedia data. In Em-         JHU CLSP 2007 Summer Workshop.
   pirical Methods in Natural Language Processing
   (EMNLP).                                              Gerard Salton and Michael McGill. 1983. Introduc-
                                                           tion to Modern Information Retrieval. McGraw-
Markus Dreyer, Jason Smith, and Jason Eisner. 2008.        Hill Book Company.
 Latent-variable modeling of string transductions
 with finite-state methods. In Empirical Methods in       Erik Tjong Kim Sang and Fien De Meulder. 2003. In-
 Natural Language Processing (EMNLP).                       troduction to the conll-2003 shared task: Language-
                                                            independent named entity recognition. In Confer-
Anthony Fader, Stephen Soderland, and Oren Etzioni.
                                                            ence on Natural Language Learning (CONLL).
  2009. Scaling Wikipedia-based named entity dis-
  ambiguation to arbitrary web text. In WikiAI09         Zareen Syed, Tim Finin, and Anupam Joshi. 2008.
  Workshop at IJCAI 2009.                                  Wikipedia as an ontology for describing documents.
                                                           In Proceedings of the Second International Confer-
Thorsten Joachims. 2002. Optimizing search engines
                                                           ence on Weblogs and Social Media. AAAI Press.
  using clickthrough data. In Knowledge Discovery
  and Data Mining (KDD).
Martin Klein and Michael L. Nelson. 2008. A com-
 parison of techniques for estimating IDF values to
 generate lexical signatures for the web. In Work-
 shop on Web Information and Data Management
Fangtao Li, Zhicheng Zhang, Fan Bu, Yang Tang,
  Xiaoyan Zhu, and Minlie Huang. 2009. THU
  QUANTA at TAC 2009 KBP and RTE track. In Text
  Analysis Conference (TAC).

Shared By: