The task of automatic keyphrase extraction has been studied for several years. Firstly, as semantic metadata useful for tasks such as summarization (Barzilay and Elhadad, 1997; Lawrie et al., 2001; DAvanzo and Magnini, 2005), but later rec­ognizing the impact that good keyphrases would have on the quality of various Nat­ural Language Processing (NLP) applica­tions (Frank et al., 1999; Witten et al., 1999; Turney, 1999; Barker and Corrnacchia, 2000; Medelyan and Witten, 2008). Thus, the selection of important, topical phrases from within the body of a document may be used in order to improve the performance of systems dealing with different NLP problems such as, clustering, question-answering, named entity recognition, information retrieval, etc.

In general, a keyphrase may be considered as a sequence of one or more words that capture the main topic of the document, as that keyphrase is expected to represent one of the key ideas ex­pressed by the document author. Following the previously mentioned hypothesis, we may take ad­vantage of two different techniques of text analy­sis: maximal frequent sequences to extract a se­quence of one or more words from a given text, and pageranking, expecting to extract those word sequences that represent the key ideas of the au­thor.

The interest on extracting high quality keyphrases from raw text has motivated forums, such as SemEval, where different systems may evaluate their performances. The purpose of SemEval is to evaluate semantic analysis systems. In particular, in this paper we are reporting the results obtained in Task #5 of SemEval-2 2010, which has been named: "Automatic Keyphrase Extraction from Scientific Articles". We focused this paper on the description of our approach and, therefore, we do not describe into detail the task nor the dataset used. For more information about this information read the "Task #5 Description paper", also published in this proceedings volume (Nam Kim et al, 2010).

The rest of this paper is structured as follows. Section 2 describes into detail the components of the proposed approach. In Section 3 it is shown the performance of the presented system. Finally, in Section 4 a discussion of findings and further work is given.

The approach presented in this paper relies on the combination of two different techniques for select­ing the most prominent terms of a given text: max­imal frequent sequences and pageranking. In Fig­ure 1 we may see this two step approach, where we are considering a sequence to be equivalent to an n-gram. The complete description of the pro­cedure is given as follows.

We select maximal frequent sequences which we consider to be candidate keyphrases and, there­after, we ranking them in order to determine which ones are the most importants (according to the pageranking algorithm). In the following subsec­tions we give a brief description of these two tech­niques. Afterwards, we provide an algorithm of the presented approach.

Scientific articles

n-grams extraction

Pageranking

Keyphrases

Figure 1 : Two step approach of BUAP Team at the Task #5 of SemEval-2

4 MF S = Maximal_Freq_Sequences(d|); s CK = PageRanking(d!, M F S);

7 CT = Top_Three_Trigrams(CK); Ki = CT; NU = 0; Acronyms = 0; foreach unigram G CU do

if unigram G AcronymSet then if Acronyms < 3 then Ki = Ki[j {unigram}; EA = MultiTermf/umgram); K% = K% U {EA}; Acronyms++;

In this edition of the Task #5 of SemEval-2 2010, we tested three different runs, which were named: BUAP - 1, BUAP - 2 and BUAP - 3. Definition and differences among the three runs are given in Table 3.

The results obtained with each run, together with three different baselines are given in the fol­lowing section.

In all tables, P, R, F mean micro-averaged pre­cision, recall and F-scores. For baselines, there were provided 1,2,-3 grams as candidates and TFIDF as features. In Table 2, TFIDF is an unsupervised method to rank the candidates based on TFIDF scores. NB and ME are super­vised methods using Naïve Bayes and maximum entropy in WEKA. In second column, R means to use the reader-assigned keyword set as gold-standard data and C means to use both author-assigned and reader-assigned keyword sets as an­swers.

Notice from Tables 2 and 3 that we outper­formed all the baselines for the Top 15 candidates. However, the Top 10 candidates were only outper­formed by the .Reader-Assigned keyphrases found. This implies that the Writer keyphrases we ob­tained were not of as good as the Reader ones. As we mentioned, readers and writers assign different keywords. The former write keyphrases based on the lecture done, by the latter has a wider context and their keyphrases used to be more complex. We plan to investigate this issue in the future.

We have presented an approach based on the ex­traction of maximal frequent sequences which are then ranked by using the pageranking algorithm. Three different runs were tested, modifying the preprocessing stage and the number of bigrams given as output. We did not see an improve­ment when we used lemmatization of the docu­ments. The run which obtained the best results was ranking by the organizer according to the top 15 best keyphrases, however, we may see that our runs need to be analysed more into detail in order to provide a re-ranking procedure for the best 15 keyphrases found. This procedure may improve the top 5 candidates precision.

This work has been partially supported by CONA-CYT (Project #106625) and PROMEP (Grant #103.5/09/4213).