% final results in comparison to other systems.  A typical TREC KBA

% poster presentation or talk explains the system pipeline and reports

% the final results. The systems may employ similar (even the same)

% steps  but the choices they make at every step are usually

% different. 

In such a situation, it becomes hard to identify the factors that

result in improved performance. There is  a lack of insight across

different approaches. This makes  it hard to know whether the

improvement in performance of a particular approach is due to

preprocessing, filtering, classification, scoring  or any of the

sub-components of the pipeline.

In this paper, we therefore fix the subsequent steps of the pipeline,

and zoom in on \emph{only} the filtering step; and conduct an in-depth analysis of its

main components.  In particular, we study the effect of cleansing,

entity profiling, type of entity filtered for (Wikipedia or Twitter), and

document category (social, news, etc) on the filtering components'

performance. The main contribution of the

paper are an in-depth analysis of the factors that affect entity-based

stream filtering, identifying optimal entity profiles without

compromising precision, describing and classifying relevant documents

that are not amenable to filtering , and estimating the upper-bound

of recall on entity-based filtering.

We base this analysis on the TREC-KBA 2013 dataset%

\footnote{\url{http://trec-kba.org/trec-kba-2013.shtml}}

that consists of three main parts: a time-stamped stream corpus, a set of

KB entities to be curated, and a set of relevance judgments. A CCR

system now has to identify for each KB entity which documents in the

stream corpus are to be considered by the human curator.

\subsection{Stream corpus} The stream corpus comes in two versions:

raw and cleaned. The raw and cleansed versions are 6.45TB and 4.5TB

respectively,  after xz-compression and GPG encryption. The raw data

is a  dump of  raw HTML pages. The cleansed version is the raw data

after its HTML tags are stripped off and only English documents

identified with Chromium Compact Language Detector

\footnote{\url{https://code.google.com/p/chromium-compact-language-detector/}}

are included.  The stream corpus is organized in hourly folders each

of which contains many  chunk files. Each chunk file contains between

hundreds and hundreds of thousands of serialized  thrift objects. One

thrift object is one document. A document could be a blog article, a

news article, or a social media post (including tweet).  The stream

corpus comes from three sources: TREC KBA 2012 (social, news and

linking) \footnote{\url{http://trec-kba.org/kba-stream-corpus-2012.shtml}},

arxiv\footnote{\url{http://arxiv.org/}}, and

spinn3r\footnote{\url{http://spinn3r.com/}}.

Table \ref{tab:streams} shows the sources, the number of hourly

%The results of the different entity profiles on the raw corpus are

%broken down by source categories and relevance rank% (vital, or

%relevant).  

In total, the dataset contains 24162 unique entity-document

pairs, vital or relevant; 9521 of these have been labelled as vital,

and the remaining 17424 as relevant.

All documents are categorized into 8 source categories: 0.98\%

arxiv(a), 0.034\% classified(c), 0.34\% forum(f), 5.65\% linking(l),

11.53\% mainstream-news(m-n), 18.40\% news(n), 12.93\% social(s) and

50.2\% weblog(w). We have regrouped these source categories into three

groups ``news'', ``social'', and ``other'', for two reasons: 1) some groups

are very similar to each other. Mainstream-news and news are

similar. The reason they exist separately, in the first place,  is

because they were collected from two different sources, by different

groups and at different times. we call them news from now on.  The

same is true with weblog and social, and we call them social from now

on.   2) some groups have so small number of annotations that treating

them independently does not make much sense. Majority of vital or

relevant annotations are social (social and weblog) (63.13\%). News

(mainstream +news) make up 30\%. Thus, news and social make up about

93\% of all annotations.  The rest make up about 7\% and are all

grouped as others.

 The TREC Filtering track defines filtering as a ``system that sifts

 through stream of incoming information to find documents that are

 relevant to a set of user needs represented by profiles''

 \cite{robertson2002trec}. Its information needs are long-term and are

 represented by persistent profiles, unlike the traditional search system

 whose adhoc information need is represented by a search

 query. Adaptive Filtering, one task of the filtering track,  starts

 with  a persistent user profile and a very small number of positive

 examples. A filtering system can improve its user profiles with a

 feedback obtained from interaction with users, and thereby improve

 its performance. The  filtering stage of entity-based stream

 filtering and ranking can be likened to the adaptive filtering task

 of the filtering track. The persistent information needs are the KB

 entities, and the relevance judgments are the small number of postive

 examples.

Stream filtering is then the task to, given a stream of documents of news items, blogs

 and social media on one hand and a set of KB entities on the other,

 to filter the stream for  potentially relevant documents  such that

 the relevance classifier(ranker) achieves as maximum performance as

 possible.  Specifically, we conduct in-depth analysis on the choices

 and factors affecting the cleansing step, the entity-profile

 construction, the document category of the stream items, and the type

\end{enumerate}

The TREC filtering and the filtering as part of the entity-centric

stream filtering and ranking pipepline have different purposes. The

TREC filtering track's goal is the binary classification of documents:

for each incoming docuemnt, it decides whether the incoming document

is relevant or not for a given profile. The docuemnts are either

relevant or not. In our case, the documents have relevance ranking and

the goal of the filtering stage is to filter as many potentially

relevant documents as possible, but less  irrelevant documents as

possible not to obfuscate the later stages of the piepline.  Filtering

as part of the pipeline needs that delicate balance between retrieving

relavant documents and irrrelevant documensts. Bcause of this,

filtering in this case can only be studied by binding it to the later

stages of the entity-centric pipeline. This bond influnces how we do

evaluation.

To achieve this, we use recall percentages in the filtering stage for

the different choices of entity profiles. However, we use the overall

performance to select the best entity profiles.To generate the overall

pipeline performance we use the official TREC KBA evaluation metric

and scripts \cite{frank2013stream} to report max-F, the maximum

F-score obtained over all relevance cut-offs.

There has been a great deal of interest  as of late on entity-based filtering and ranking. One manifestation of that is the introduction of TREC KBA in 2012. Following that, there have been a number of research works done on the topic \cite{frank2012building, ceccarelli2013learning, taneva2013gem, wang2013bit, balog2013multi}.  These works are based on KBA 2012 task and dataset  and they address the whole problem of entity filtering and ranking.  TREC KBA continued in 2013, but the task underwent some changes. The main change between  the 2012 and 2013 are in the number of entities, the type of entities, the corpus and the relevance rankings.

The number of entities increased from 29 to 141, and it included 20 Twitter entities. The TREC KBA 2012 corpus is 1.9TB after xz-compression and has  400M documents. By contrast, the KBA 2013 corpus is 6.45 after XZ-compression and GPG encryption. A version with all-non English documented removed  is 4.5 TB and consists of 1 Billion documents. The 2013 corpus subsumed the 2012 corpus and added others from spinn3r, namely main-stream news, forum, arxiv, classified, reviews and meme-tracker.  A more important difference is, however, a change in the definitions of relevance ratings vital and relevant. While in KBA 2012, a document was judged vital if it has citation-worthy content for a given entity, in 2013 it must have the freshliness, that is the content must trigger an editing of the given entity's KB entry. 

While the tasks of 2012 and 2013 are fundamentally the same, the approaches  varied due  to the size of the corpus. In 2013, all participants used filtering to reduce the size of the big corpus.   They used different ways of filtering: many of them used two or more of different name variants from DBpedia such as labels, names, redirects, birth names, alias, nicknames, same-as and alternative names \cite{wang2013bit,dietzumass,liu2013related, zhangpris}.  Although most of the participants used DBpedia name variants none of them used all the name variants.  A few other participants used bold words in the first paragraph of the Wikipedia entity's profiles and anchor texts from other Wikipedia pages  \cite{bouvierfiltering, niauniversity}. One participant used Boolean \emph{and} built from the tokens of the canonical names \cite{illiotrec2013}.  

All of the studies used filtering as their first step to generate a smaller set of documents. And many systems suffered from poor recall and their system performances were highly affected \cite{frank2012building}. Although  systems  used different entity profiles to filter the stream, and achieved different performance levels, there is no study on and the factors and choices that affect the filtering step itself. Of course filtering has been extensively examined in TREC Filtering \cite{robertson2002trec}. However, those studies were isolated in the sense that they were intended to optimize recall. What we have here is a different scenario. Documents have relevance rating. Thus we want to study filtering in connection to  relevance to the entities and thus can be done by coupling filtering to the later stages of the pipeline. This is new to the best of our knowledge and the TREC KBA problem setting and data-sets offer a good opportunity to examine this aspect of filtering. 

Moreover, there has not been a chance to study at this scale and/or a study into what type of documents defy filtering and why? In this paper, we conduct a manual examination of the documents that are missing and classify them into different categories. We also estimate the general upper bound of recall using the different entities profiles and choose the best profile that results in an increased over all performance as measured by F-measure. 

All analyses in this paper are carried out on the documents that have

relevance assessments associated to them. For this purpose, we

extracted those documents from the big corpus. We experiment with all

KB entities. For each KB entity, we extract different name variants

from DBpedia and Twitter.

\

\subsection{Entity Profiling}

We build entity profiles for the KB entities of interest. We have two

types: Twitter and Wikipedia. Both entities have been selected, on

purpose by the track organisers, to occur only sparsely and be less-documented.

For the Wikipedia entities, we fetch different name variants

from DBpedia: name, label, birth name, alternative names,

redirects, nickname, or alias. 

These extraction results are summarized in Table

\ref{tab:sources}.

For the Twitter entities, we visit

their respective Twitter pages and fetch their display names. 

\begin{table}

\caption{Number of different DBpedia name variants}

\begin{center}

 \begin{tabular}{l*{4}{c}l}

 Name variant& No. of strings  \\

             \hline 

\multirow{2}{*}{Total} & Wikipedia       &8071   &14426&19832  \\

                       &Twitter  &1450  &2998&4330  \\

                       &All Entities&9521   &17424&24162 \\

\hline

\end{tabular}

\end{center}

\label{tab:breakdown}

\end{table}

%Most (more than 80\%) of the annotation documents are in the test set.

The 2013 training and test data contain 68405

annotations, of which 50688 are unique document-entity pairs.   Out of

these, 24162 unique document-entity pairs are vital (9521) or relevant

(17424).

 We conducted experiments to study  the effect of cleansing, different entity profiles, types of entities, category of documents, relevance ranks (vital or relevant), and the impact on classification.  In the following subsections, we present the results in different categories, and describe them.

 \subsection{Cleansing: raw or cleansed}

\begin{table}

\caption{Percentage of vital or relevant documents retrieved under different name variants (upper part from cleansed, lower part from raw)}

\begin{center}

\begin{tabular}{l@{\quad}rrrrrrr}

\hline

&cano&cano-part  &all &all-part  \\

\hline

   Wikipedia      &61.8  &74.8  &71.5  &77.9\\

   Twitter        &1.9   &1.9   &41.7  &80.4\\

   All Entities   &51.0  &61.7  &66.2  &78.4 \\	

\hline

\hline

   Wikipedia      &70.0  &86.1  &82.4  &90.7\\

   Twitter        & 8.7  &8.7   &67.9  &88.2\\

  All Entities    &59.0  &72.2  &79.8  &90.2\\

\hline

``roryscovel`` is the canonical name and its partial is identical.}

%The low recall is because the canonical names of Twitter entities are

%not really names; they are usually arbitrarily created user names. It

%shows that  documents  refer to them by their display names, rarely

%by their user name, which is reflected in the name-variant recall

%(67.9\%). The use of name-variant partial increases the recall to

%88.2\%.

The tables in \ref{tab:name} and \ref{tab:source-delta} show a higher recall

for Wikipedia than for Twitter entities. Generally, at both

aggregate and document category levels, we observe that recall

increases as we move from canonicals to canonical partial, to

name-variant, and to name-variant partial. The only case where this

does not hold is in the transition from Wikipedia's canonical partial

to name-variant. At the aggregate level (as can be inferred from Table

\ref{tab:name}), the difference in performance between  canonical  and

name-variant partial is 31.9\% on all entities, 20.7\% on Wikipedia

entities, and 79.5\% on Twitter entities. 

Section \ref{sec:analysis} discusses the most plausible explanations for these findings.

%% TODO: PERHAPS SUMMARY OF DISCUSSION HERE

In the overall experimental setup, classification, ranking, and

evaluation are kept constant. Following \cite{balog2013multi}

settings, we use

WEKA's\footnote{\url{http://www.cs.waikato.ac.nz/~ml/weka/}} Classification

Random Forest. However, we use fewer numbers of features which we

found to be more effective. We determined the effectiveness of the

features by running the classification algorithm using the fewer

features we implemented and their features. Our feature

implementations achieved better results.  The total numbers of

features we used are 13 and are listed below.

\paragraph*{Google's Cross Lingual Dictionary (GCLD)}

This is a mapping of strings to Wikipedia concepts and vice versa

\cite{spitkovsky2012cross}. 

(1) the probability with which a string is used as anchor text to

a Wikipedia entity 

\paragraph*{jac} 

  Jaccard similarity between the document and the entity's Wikipedia page

\paragraph*{cos} 

  Cosine similarity between the document and the entity's Wikipedia page

\paragraph*{kl} 

  KL-divergence between the document and the entity's Wikipedia page

2) Taking both performance (recall at filtering and overall F-score

during evaluation) into account, there is a clear trade-off between using a richer entity-profile and retrieval of irrelevant documents. The richer the profile, the more relevant documents it retrieves, but also the more irrelevant documents. To put it into perspective, lets compare the number of documents that are retrieved with  canonical partial and with name-variant partial. Using the raw corpus, the former retrieves a total of 2547487 documents and achieves a recall of 72.2\%. By contrast, the later retrieves a total of 4735318 documents and achieves a recall of 90.2\%. The total number of documents extracted increases by 85.9\% for a recall gain of 18\%. The rest of the documents, that is 67.9\%, are newly introduced irrelevant documents. 

Wikipedia's canonical partial is the best entity profile for Wikipedia entities. This is interesting  to see that the retrieval of of  thousands vital-relevant document-entity pairs by name-variant partial does not translate to an increase in over all performance. It is even more interesting since canonical partial was not considered as contending profile for stream filtering by any of participant to the best of our knowledge. With this understanding, there  is actually no need to go and fetch different names variants from DBpedia, a saving of time and computational resources.

%%%%%%%%%%%%

The deltas between entity profiles, relevance ratings, and document categories reveal four differences between Wikipedia and Twitter entities. 1) For Wikipedia entities, the difference between canonical partial and canonical is higher(16.1\%) than between name-variant partial and  name-variant(8.3\%).  This can be explained by saturation. This is to mean that documents have already been extracted by  name-variants and thus using their partials does not bring in many new relevant documents.  2) Twitter entities are mentioned by name-variant or name-variant partial and that is seen in the high recall achieved  compared to the low recall achieved by canonical(or their partial). This indicates that documents (specially news and others) almost never use user names to refer to Twitter entities. Name-variant partials are the best entity profiles for Twitter entities. 3) However, comparatively speaking, social documents refer to Twitter entities by their user names than news and others suggesting a difference in 

adherence to standard in names and naming. 4) Wikipedia entities achieve higher recall and higher overall performance. 

The high recall and subsequent higher overall performance of Wikipedia entities can  be due to two reasons. 1) Wikipedia entities are relatively well described than Twitter entities. The fact that we can retrieve different name variants from DBpedia is a measure of relatively rich description. Rich description plays a role in both filtering and computation of features such as similarity measures in later stages of the pipeline.   By contrast, we have only two names for Twitter entities: their user names and their display names which we collect from their Twitter pages. 2) There is not DBpedia-like resource for Twitter entities from which alternative names cane be collected.   

In the experimental results, we also observed that recall scores in the vital category are higher than in the relevant category. This observation  confirms one commonly held assumption:(frequency) mention is related to relevance.  this is the assumption why term frequency is used an indicator of document relevance in many information retrieval systems. The more  a document mentions an entity explicitly by name, the more likely the document is vital to the entity.

Across document categories, we observe a pattern in recall of others, followed by news, and then by social. Social documents are the hardest to retrieve. This can be explained by the fact that social documents (tweets and  blogs) are more likely to point to a resource where the entity is mentioned, mention the entities with some short abbreviation, or talk without mentioning the entities, but with some context in mind. By contrast news documents mention the entities they talk about using the common name variants more than social documents do. However, the greater difference in percentage recall between the different entity profiles in the news category indicates news refer to a given entity with different names, rather than by one standard name. By contrast others show least variation in referring to news. Social documents falls in between the two.  The deltas, for Wikipedia entities, between canonical partials and canonicals,  and name-variants and canonicals are high, an indication that canonical partials 

and name-variants bring in new relevant documents that can not be retrieved by canonicals. The rest of the two deltas are very small,  suggesting that partial names of name variants do not bring in new relevant documents. 

\subsection{Missing vital-relevant documents \label{miss}}

% 

 The use of name-variant partial for filtering is an aggressive attempt to retrieve as many relevant documents as possible at the cost of retrieving irrelevant documents. However, we still miss about  2363(10\%) of the vital-relevant documents.  Why are these documents missed? If they are not mentioned by partial names of name variants, what are they mentioned by? Table \ref{tab:miss} shows the documents that we miss with respect to cleansed and raw corpus.  The upper part shows the number of documents missing from cleansed and raw versions of the corpus. The lower part of the table shows the intersections and exclusions in each corpus.  

\begin{table}

\caption{The number of documents missing  from raw and cleansed extractions. }

\begin{center}

\begin{tabular}{l@{\quad}llllll}

\hline

\multicolumn{1}{l}{\rule{0pt}{12pt}category}&\multicolumn{1}{l}{\rule{0pt}{12pt}Vital }&\multicolumn{1}{l}{\rule{0pt}{12pt}Relevant }&\multicolumn{1}{l}{\rule{0pt}{12pt}Total }\\[5pt]

\hline

Cleansed &1284 & 1079 & 2363 \\

Raw & 276 & 4951 & 5227 \\

\hline

 missing only from cleansed &1065&2016&3081\\

  missing only from raw  &57 &160 &217 \\

  Missing from both &219 &1927&2146\\

\hline

%%%%%%%%%%%%%%%%%%%%%%

We observed that there are vital-relevant documents that we miss from raw only, and similarly from cleansed only. The reason for this is transformation from one format to another. The most interesting documents are those that we miss from both raw and cleansed corpus. We first identified the number of KB entities who have a vital relevance judgment and  whose documents can not be retrieved (they were 35 in total) and conducted a manual examination into their content to find out why they are missing. 

 We  observed  that among the missing documents, different document ids can have the same content, and be judged multiple times for a given entity.  %In the vital annotation, there are 88 news, and 409 weblog. 

 Avoiding duplicates, we randomly selected 35 documents, one for each entity.   The documents are 13 news and  22  social. Here below we have classified the situation under which a document can be vital for an entity without mentioning the entities with the different entity  profiles we used for filtering. 

\paragraph*{Outgoing link mentions} A post (tweet) with an outgoing link which mentions the entity.

\paragraph*{Event place - Event} A document that talks about an event is vital to the location entity where it takes place.  For example Maha Music Festival takes place in Lewis and Clark\_Landing, and a document talking about the festival is vital for the park. There are also cases where an event's address places the event in a park and due to that the document becomes vital to the park. This is basically being mentioned by address which belongs to alarger space. 

\paragraph*{Entity -related entity} A document about an important figure such as artist, athlete  can be vital to another. This is specially true if the two are contending for the same title, one has snatched a title, or award from the other. 

\paragraph*{Organization - main activity} A document that talks about about an area on which the company is active is vital for the organization. For example, Atacocha is a mining company  and a news item on mining waste was annotated vital. 

\paragraph*{Entity - group} If an entity belongs to a certain group (class),  a news item about the group can be vital for the individual members. FrankandOak is  named innovative company and a news item that talks about the group  of innovative companies is relevant for a  it. Other examples are: a  big event  of which an entity is related such an Film awards for actors. 

\paragraph*{Artist - work} Documents that discuss the work of artists can be relevant to the artists. Such cases include  books or films being vital for the book author or the director (actor) of the film. Robocop is film whose screenplay is by Joshua Zetumer. A blog that talks about the film was judged vital for Joshua Zetumer. 

\paragraph*{Politician - constituency} A major political event in a certain constituency is vital for the politician from that constituency. 

 A good example is a weblog that talks about two north Dakota counties being drought disasters. The news is vital for Joshua Boschee, a politician, a member of North Dakota democratic party.  

\paragraph*{head - organization} A document that talks about an organization of which the entity is the head can be vital for the entity.  Jasper\_Schneider is USDA Rural Development state director for North Dakota and an article about problems of primary health centers in North Dakota is judged vital for him. 

\paragraph*{World Knowledge} Some things are impossible to know without your world knowledge. For example ''refreshments, treats, gift shop specials, "bountiful, fresh and fabulous holiday decor," a demonstration of simple ways to create unique holiday arrangements for any home; free and open to the public`` is judged relevant to Hjemkomst\_Center. This is a social media post, and unless one knows the person posting it, there is no way that this text shows that. Similarly ''learn about the gray wolf's hunting and feeding behaviors and watch the wolves have their evening meal of a full deer carcass; $15 for members, $20 for nonmembers`` is judged vital to Red\_River\_Zoo.  

\paragraph*{No document content} A small number of documents were found to have no content.

\paragraph*{Disagreement} For a few remaining documents, the authors disagree with the assessors as to why these are vital to the entity.

In this paper, we examined the filtering stage of the entity-centric stream filtering and ranking  by holding the later stages of fixed. In particular, we studied the cleansing step, different entity profiles, type of entities(Wikipedia or Twitter), categories of documents(news, social, or others) and the relevance ratings. We attempted to address the following research questions: 1) does cleansing affect filtering and subsequent performance? 2) what is the most effective way of entity profiling? 3) is filtering different for Wikipedia and Twitter entities? 4) are some type of documents easily filterable and others not? 5) does a gain in recall at filtering step translate to a gain in F-measure at the end of the pipeline? and 6) what are the circumstances under which vital documents can not be retrieved? 

Cleansing does remove parts or entire contents of documents making them irretrievable. However, because of the introduction of false positives, recall gains by  raw corpus and some  richer entity profiles do not necessarily translate to overall performance gain. The results conclusion on this is mixed in the sense that cleansing helps improve the recall on vital documents and Wikipedia entities, but reduces the recall on Twitter entities and the relative category of relevance ranking. Vital and relevant documents show a difference in retrieval nonperformance documents are easier to filter than relevant.  

Despite an aggressive attempt to filter as many vital-relevant documents as possible,  we observe that there are still documents that we miss. While some are possible to retrieve with some modifications, some others are not. There are some document that indicate that an information filtering system does not seem to get them no matter how rich representation of entities they use. These circumstances under which this happens are many. We found that some documents have no content at all, subjectivity(it is not clear why some are judged vital). However, the main circumstances under which vital  documents can defy filtering is: outgoing link mentions, 

venue-event, entity - related entity, organization - main area of operation, entity - group, artist - artist's work,  party-politician, and world knowledge.  

%ACKNOWLEDGMENTS are optional

%\section{Acknowledgments}

%

% The following two commands are all you need in the

% initial runs of your .tex file to

% produce the bibliography for the citations in your paper.

\bibliographystyle{abbrv}

\bibliography{sigproc}  % sigproc.bib is the name of the Bibliography in this case

% You must have a proper ".bib" file

%  and remember to run:

% latex bibtex latex latex

% to resolve all references

%

% ACM needs 'a single self-contained file'!