Tweets are short messages that do not exceed 140 characters. Since they must be written respecting this limitation, a particular vocabulary is used. To make them understandable to a reader, it is therefore necessary to know their context. In this paper, we describe our approach submitted for the tweet contextualization track in CLEF 2014 (Conference and Labs of Evaluation Forums). This approach allows the extension of the tweet's vocabulary by a set of thematically related words using mining association rules between terms.
Web 2.0 is the term associated with the transition of the World Wide Web
from a collection of individual web sites to an emerging platform in its own
right. This emergence is due largely to users collaborations, these users have
been the driving force for the emergence of new services [
In contrast to traditional blogs, media-sharing and social networks services,
microblogs (tweets) are textual messages submitted in real-time to report an
idea, an actual interest, or an opinion [
In this respect, we will focus on the Tweet Contextualization track. The participants of INEX 20141 are required to perform the task of contextualizing tweets, i.e., given a tweet and a related entity, they try to answer questions of the form "why this tweet concerns this entity? should it be an alert?".These questions can be answered by several sentences or by an aggregation of texts from di erent articles of Wikipedia.
The purpose of this task is to allow the reader a better understanding of the tweet. This task can be divided into two subtasks. The rst is to nd the most relevant Wikipedia articles using an IR system, and the second is to extract the most representative passages of the tweet from the relevant Wikipedia articles.
In this paper, we propose to use a statistical method based on association
rules mining [
The remainder of this paper is organized as follows: Section 2 cites some related works. Then, Section 3 describes the test data. Section 4 recalls the basic de nition for the derivation of association rules between terms and details our proposed approach for tweets contextualization based on these association rules. Next, Section 5 describes our di erent submitted runs for the tweet contextualization track as the experiment results. The conclusion and future work are nally presented in Section 6. 2
Though the idea to contextualize tweets is quite recent, there are several works
in this context. Recently, Morchid et al. [
1. A collection of articles, that has been rebuilt based on a recent dump of the English Wikipedia from November 2012, it is composed of 3 902 346 articles, all notes and bibliographic references that are di cult to handle are removed and only non-empty Wikipedia pages (pages having at least one section) are kept. Resulting documents are made of a title (title), an abstract (a) and sections (s). Each section has a sub-title (h). Abstract and sections are made of paragraphs (p) and each paragraph can have entities (t) that refer to Wikipedia pages. Each document is provided in XML format and respects the Document Type De nition (DTD) described in Table 1. <!ELEMENT xml (page)+> <!ELEMENT page (ID, title, a, s*)> <!ELEMENT ID (#PCDATA)> <!ELEMENT title (#PCDATA)> <!ELEMENT a (p+)> <!ELEMENT s (h, p+)> <!ATTLIST s o CDATA #REQUIRED> <!ELEMENT h (#PCDATA)> <!ELEMENT p (#PCDATA | t)*> <!ATTLIST p o CDATA #REQUIRED> <!ELEMENT t (#PCDATA)> <!ATTLIST t e CDATA #IMPLIED> 2. A collection of English tweets, composed of 240 tweets selected from the CLEF RepLab 2013, to focus on content analysis alone. URLs are removed from the tweets. An example of a tweet representation is given in Table 2. Each tweet is composed of: { An identi er (Id), { A category, the tweet collection is divide into four (4) distinct categories: automotive, music, banking and university. { An entity, there are 64 distinct entities name from Wikipedia. { A topic, there are 235 distinct manual topic labels. { A text content, which must not exceed 140 characters. { A timestamp.
Given a user query, the number of retrieved documents can be overwhelmingly
large, hampering their e cient exploitation by the user. In this situation, the
query expansion technique o ers an interesting solution for obtaining a complete
answer while preserving the quality of retained documents. In [
An association rule R is an implication of the form R: T1 ) T2, where T1 and T2 are subsets of T , and T1 \ T2 = ;. The termsets T1 and T2 are, respectively, called the premise and the conclusion of R. The rule R is said to be based on the termset T equal to T1 [ T2. The support of a rule R: T1 ) T2 is then de ned as 1,2: while its con dence is computed as:
Supp(R) = Supp(T ); Conf(R) =
Supp(T ) Supp(T1) : (1) (2)
An association R is said to be valid if its con dence value, i.e., Conf(R), is greater than or equal to a user-de ned threshold denoted minconf. This con dence threshold is used to exclude non valid rules.
Interestingly enough, to address tweet contextualization in an e cient and
e ective manner, we claim that a synergy with some advanced text mining
methods, especially association rules [
The tweet contextualization system serves to expand a given tweet and to
elaborate the corresponding query, which is sent in order to retrieve its related context.
Our proposed approach based on association rules between terms is depicted in
Figure 1. The contextualization tweet process is performed on the following steps:
1. Selecting of a sub-set of articles, according to the tweet's subject, from the
documents collection.
2. Annotating the selected Wikipedia articles with part-of-speech and lemma
information using TreeTagger 3.
3. Extracting of only nouns from the annotated Wikipedia articles, and
removing the most frequents nouns.
4. Generating the association rules using an e cient algorithm CHARM 4 for
mining all the closed frequent termsets [
We released three runs that are based on the approach described in Section 4.2, namely : Run1 (ID = 359): We selected a sub-set of 50 000 articles, according to the entities associated to tweets, using an algorithm that calculates the similarity between the entities associated to tweets and the articles from the documents collection. Then, we proceeded as in the proposed approach by applying the algorithm CHARM with the following parameters : minsupp = 16, and minconf = 0:7.
Run2 (ID = 360): In the second run, we used the entities associated to tweets to extract their context, then we exploited these contexts considering them as articles, we proceeded, afterwards, as in the proposed approach. We applied the CHARM algorithm taking minsupp = 5 and minconf = 0:7 as parameters. Run3 (ID = 359): In this run, we used the recommended documents list for query, provided by INEX. As in runs 1 and 2, we proceeded, afterwards, as in the proposed approach. We applied the CHARM algorithm taking minsupp = 5 and minconf = 0:7 as parameters. 5.2
Informativeness: Informativeness aims at measuring how well the summary helps a user to understand the tweet content. Therefore, for each tweet, each passage will be evaluated independently from the others, even in the same summary. the results are based on a thorough manual run on 1/5 of the 2014 topics using the baseline system. From this run two types of references were extracted, namely: { a list of relevant sentences per topic. { extraction of Noun Phrases from these sentences together with the corresponding Wikipedia entry.
Two reference runs using 2013 and 2012 corpus were also generated to see if the evolution of the Wikipedia had an impact.
The dissimilarity between a reference text and the proposed summary is given P = fT (t) + 1
fT Q = fS(t) + 1
fS T , a set of query terms present in reference summary and for each t 2 T . fT (t), the frequency of term t in reference summary.
S, a set of query terms present in a submitted summary and for each t 2 S. fS (t), the frequency of term t in a submitted summary.
Organizers used three di erent distributions for the reference summaries in the 2014 tracks, namely: { Unigrams made of single lemmas (after removing stop-words). { Bigrams made of pairs of consecutive lemmas (in the same sentence). { Bigrams with 2-gaps also made of pairs of consecutive lemmas but allowing the insertion between them of a maximum of two lemmas (also referred to as skip distribution).
INEX evaluated 12 runs from four participants. Our result runs (IDs are 361, 360 and 359) ranked rst, second and fth, respectively, in the informativeness evaluation based on sentences (cf. Table3). Table 4 highlights informativeness by: where :
Dis(T; S) = 1)
1 X(P t2T min(log(P ); log(Q)) max(log(P ); log(Q)) (3) quality -based on noun phrases- di erence between the original queries (Ref Runs) and the expanded ones using association rules mined with the algorithm CHARM. Our result runs 361, 360 and 359 are ranked rst, fourth and sixth, respectively. Readability: readability aims at measuring how clear and easy it is to understand the summary. Each summary has been evaluated by considering the following four parameters: { Readable: Percentage of passages considered as readable (non trash). { Syntax: Percentage of passages without syntax or grammatical errors. { Diversity: Percentage of non redundant passages.
{ Structure: Percentage of non breaking anaphora passages.
Evaluation was carried out by one reader over a pool of 12 summaries per run. Our runs, 359,360 and 361, ranked eighth, fth and eleventh respectively on the o cial evaluation. Table 5) shows the runs from the 2014 readability evaluations.
The obtained informativeness evaluation results shed light that our proposed approach, based on association rules, o ers interesting results and helps ensure that context summaries contain adequate correlating information with the tweets and avoid inclusion of non-similar information in them as much as possible. Run Id Rank Readable Syntax Diversity Structure Avg Furthermore, our tweets contextualization approach based on association rules leads to the improvement of the context informativeness ; we note also that the selection of articles impacts the quality of the contexts. The contexts are less readable, it may be that they contain some unexpected noises which need to be cleaner. 6
In this paper, we have described a new method for tweet contextualization based on association rules between sets of terms. The experimental study was conducted on INEX 2014 collections. The results con rmed that the synergy between association rules and query expansion is fruitful. Further work investigates whether the combination of the con dence measure with other measures in the weighting process can be of bene t to the tweet contextualization process. Acknowledgments. This work is partially supported by the French-Tunisian project PHC-Utique RIMS-FD 14G 1404.