According to the organizers, the aim of the 2013 INEX Tweet Contextualization Track is: \...given a tweet, the system must provide some context about the subject of the tweet, in order to help the reader to understand it. This context should take the form of a readable (and short) summary, composed of passages from [...] Wikipedia." We present an automatic greedy summarizer named REG applied to the INEX 2013 task. REG summarizer uses a greedy optimization algorithm to weigh the sentences. The summary is obtained by concatenating the relevant sentences, weighed in the optimization step. The results show that the REG system (using original tweets with a manual processing) do not perform very well on INEX 2013 contextualization track.
Automatic text summarization is indispensable to cope with ever increasing
volumes of valuable information. An abstract is by far the most concrete and
most recognized kind of text condensation [
An open domain Question-Answering system (QA) has to precisely answer a question expressed in natural language. QA systems are confronted with a ne and di cult task because they are expected to supply speci c information and not whole documents. At present there exists a strong demand for this kind of text processing systems on the Internet. A QA system comprises, a priori, the following stages: 1 REsumeur Glouton (Greedy summarizer). 1. Transform the questions into queries, then associate them to a set of documents; 2. Filter and sort these documents to calculate various degrees of similarity; 3. Identify the sentences which might contain the answers, then extract text fragments from them that constitute the answers. In this phase an analysis using Named Entities (NE) is essential to nd the expected answers.
Most research e orts in summarization emphasize generic summarization [
This paper is organized as follows. In Section 2 we explain the INEX 2013
Tweet Contextualization Track. In Section 3 we explain the methodology of our
work. Experimental settings and results obtained with REG are presented in
Section 4. Section 5 exposes the conclusions of the paper and the future work.
2 INEX 2013 Tweet Contextualization Track
The Initiative for the Evaluation of XML Retrieval (INEX) is an established
evaluation forum for XML information retrieval (IR) [
Like in Question Answering track of INEX 2011 and 2012, the present task is about contextualizing tweets, i.e. answering questions of the form "What is this tweet about?" using a recent cleaned dump of the Wikipedia3. As organizers claim, the general process involves three steps: 1. Tweet analysis.
3 See the o cial INEX 2013 Tweet Contextualization Track Website: https://inex. mmci.uni-saarland.de/tracks/qa/. 2. Passage and/or XML elements retrieval. 3. Construction of the answer.
Then, a relevant passage segment contains relevant information but as few nonrelevant information as possible (the result is speci c to the question). 2.1
598 tweets in English were collected by the organizers from Twitter4 for the Track 2013.. Tweets were selected and checked among informative accounts (for example, @CNN, @TennisTweets, @PeopleMag, @science...), in order to avoid purely personal tweets that could not be contextualized. Information such as the user name, tags or URLs will be provided. 3
The REG system includes three modules. The rst one is responsible for the text
vectorial processing (Cortex system [
Documents are pre-treated with conventional ltering algorithms of functional
words, normalization and stemming [
The presence of the word i is represented by its frequency T Fi (his absence by 0, respectively), and a sentence is an array of N occurrences. 3.2
A graph G = (V; E) is created from the vector representation of documents, where S vertices express sentences and A is the set of edges. An edge between two nodes is established if the corresponding phrases have at least one word in common.An adjacency matrix A[P P ] is constructed from the matrix S[sentences words]. The calculation is as follows: scan the line i, and for each element ai;j equal to 1, down by j column to identify other phrases that share the word.
The proposed algorithm works as follows: 4 www.tweeter.com 5 The system Cortex performs unsupervised relevant sentences using several metrics controlled by a decision algorithm extraction. 1. generate A, which has rows and columns of sentences considered; 2. calculate the weight of vertices (the sum of incoming edges of the vertex); 3. calculate the degree of each vertex (the number of sentences partitioned with other words); The solution adopted is based on a calculation of greedy search paths. 4
In this study, we used the document sets made available during the Initiative for the Evaluation of XML retrieval (INEX)6, in particular on the INEX 2013 Tweet Contextualization Track. We have performed a simpli cation of tweets provided by using a simple manual processing and this new list T of tweets was normalized before continue our experimentation protocol.
The strategy of REG system to deal multi-document summary problem is quite simple: rst, a long single document D is formed by concatenation of all i = 1; :::; n relevant documents provided by Indri engine: d1; d2; :::dn. The rst line of this multi-document D is the tweet T . The REG summarizer system extracts of D the most relevant sentences following T . Then, this subset of sentences is sorted by the date of documents di. The summarizers add sentences into the summary until the word limit is reached.
The strategy employed to generate 598 queries from tweets was very simple. The tweets not carrying information words were removed. Then, the summarizer used the query as a title of a big multi-document retrieved by Indri engine.
We show an example of our manual processing. Let's consider the tweet number 303260378618531840 that the summary should contextualize: <topic id="169231181181747200"> <tweet> Ibra booked for having a barney with the referee The only thing he has done noticeable tonight Into stoppage time now </tweet>
Then, query 169231181181747200 is simpli ed as show: q = "Ibra booked for having a barney with the referee The only thing he has done noticeable tonight" 4.2
The REG system used three methods to normalize the words: lemmatization,
stemming and 4-ultra stemming [
INEX had provided two evaluations: the informativity and readability of the candidates summaries (500 words). The tables 1 and 2 shows the o cial results of some participants of INEX 2013 contextualization task: runs 263 to 265, the two rst places and the two last ones. Table 1 uses the all overlapping reference values to compare the performance of the di erent runs. In particular, the run 265 performs better than 263 and 264 and 4-ultra stemming outperforms stemming. Unfortunately, the divergence results provided by REG are not very good.
The same process to compare the readability results presented on table 2. In this case, stemming outperforms 4-ultra stemming. In this paper we have presented the REG (REsumeur Glouton) summarization system applied on INEX 2013 Tweet Contextualization Track. REG is an automatic greedy summarization system which uses graph methods to spot the most important sentences in the document.
REG summarizer used a normalized list issued from a manual processing on the original tweets as inputs. It did not provide good results in the informativity evaluation, but in the readability context, it could be more competitive with some improvements on the queries sent to Indri engine. The manual process strategy was quite simple.