Keyphrases are sequences of words which cap­ture the main topics discussed in a document. Keyphrases are very useful in many natural lan­guage processing (NLP) applications such as doc­ument summarization, classification and cluster­ing. But it is an expensive and time-consuming job for users to tag keyphrases of a document. These needs motivate methods for automatic keyphrase extraction.

Most existing algorithms for keyphrase extrac­tion treat this task as a supervised classifica­tion task. The KEA algorithm (Gordon et al., 1999) identifies candidate keyphrases using lex­ical methods, calculates feature values for each candidate, and uses a machine-learning algorithm to predict which candidates are good keyphrases. A domain-specific method (Frank et al., 1999) was proposed based on the Naive Bayes learn­ing scheme. Turney (Turney, 2000) treated a document as a set of phrases, which the learn­ing algorithm must learn to classify as positive or negative examples of keyphrases. Turney (Tur­ney, 2003) also presented enhancements to the

Department of Computer Science & Engineering Shanghai Jiao Tong University Shanghai, China

KEA keyphrase extraction algorithm that are de­signed to increase the coherence of the extracted keyphrases. Nguyen and yen Kan (Nguyen and yen Kan, 2007) presented a keyphrase extraction algorithm for scientific publications. They also in­troduced two features that capture the positions of phrases and salient morphological phenom­ena. Wu and Agogino (Wu and Agogino, 2004) proposed an automated keyphrase extraction al­gorithm using a nondominated sorting multi-objective genetic algorithm. Kumar and Srinathan (Kumar and Srinathan, 2008) used n-gram filtra­tion technique and weight of words for keyphrase extraction from scientific articles.

For this evaluation task, Kim and Kan (Kim and Kan, 2009) tackled two major issues in au­tomatic keyphrase extraction using scientific ar­ticles: candidate selection and feature engineer­ing. They also re-examined the existing features broadly used for the supervised approach.

Different from previous systems, our system uses a chunk based method to extract keyphrases from scientific articles. Domain-specific informa­tion is used to find out useful parts in a document. The chunk based method is used to extract candi­dates of keyphrases in a document. Keywords of a document are used to select keyphrases from can­didates.

In the following, Section 2 will describe the ar­chitecture of the system. Section 3 will introduce functions and implementation of each part in the system. Experiment results will be showed in Sec­tion 4. The conclusion will be given in Section 5.

Figure 1 shows the architecture of our system. The system accepts a document as input (go through arrows with solid lines), then does the preprocess­ing job and identifies the structure of the docu­ment. After these two steps, the formatted doc­ument is sent to the candidate selection module

Figure 1 : System architecture which first extracts chunks from the document, then uses some rules to filter the extracted chunks. After candidate selection, the system will choose top fifteen (ordered by the position of the first oc­currence in the original document) chunks from the candidates as the keyphrases and output the result ("Outputl" in Figure 1) which is our sub­mitted result. The candidates will also be sent to keyphrase selection module which first extracts keywords from the formatted document, then uses keywords to choose keyphrases from the candi­dates. Keywords extraction needs some training data (go through arrows with dotted lines) which also needs first two steps of our system. The result of keywords selection module will be sent to "Out-put2" as the final result after choosing top fifteen chunks.

OpenNLP and KEA are used in chunk extrac­tion and keywords extraction respectively.

In preprocessing, our system first deletes line breaks between each broken lines to reconnect the broken sentences while line breaks after title and section titles will be reserved. Title and section titles are recognized through some heuristic rules that title occupies first few lines of a document and section titles are started with numbers except abstract and reference. The system then deletes brackets blocks in the documents to make sure no keyphrases will be splitted by brackets blocks (e.g., the brackets in "natural language processing (NLP) applications" could be an obstacle to ex­tracting phrase "natural language processing ap­plications").

Table 1 shows the F-score of two outputs of our system and some baseline systems. The first three methods are the baselines provided by the task owner. TFIDF is an unsupervised method to rank the candidates based on TFIDF scores. NB and ME are supervised methods using Navie Bayes and maximum entropy in WEKA. KEA refers to the KEA system with the parameters that can out­put the best results. OP1 is our system with the "Outputl" as result and OP 2 is our system with the "Output2" as result (see Figure 1). In second column, "R" means to use the reader-assigned keyphrases set as gold-standard data and "C" means to use both author-assigned and reader-assigned keyphrases sets as answers.

Table 1 : The comparison of F-score of our system with other systems.

From the table, we can see that, both two out­puts of our system made an improvement over the baseline systems and got better results than the well known KEA system.

We submitted both results of OP 1 and OP 2 to the evaluation task. Because of some misunder­standing over the result upload system, only the result of OP 1 (with bold style) was successfully submitted.

We proposed a chunk based method for keyphrase extraction in this paper. In our system, document structure information of scientific articles is used to pick up significant contents, chunk based candi­date selection is used to reduce the quantity of can­didates and reserve their original meanings, key­words are used to select keyphrases from a docu­ment. All these factors contribute to the result of our system.