Key phrases are a set of words to capture the main topic of the document. Since key phrases con­tain the substance of the document, they are used in the large spectrum of areas; from applications which explicitly use key phrases such as automatic indexing, documents classification and search en­gine optimization in information retrieval, to ap­plications which implicitly use key phrases such as summarization and question-answering systems. During the last decade, many previous works have dealt with the various methods for automatically extracting key phrases (e.g., Frank et al., 1999; Barker and Corrnacchia, 2000; Turney, 2003; Medelyan and Witten, 2006; Nguyen and Kan, 2007; Wan and Xiao, 2008).

The task of extracting key phrases would be considered as a subtask of extracting terminology if key phrases are a kind of terms. Typical ap­proaches for automatically extracting terms use linguistic preprocessing which involves morpho-syntactic analysis such as part-of-speech tagging and phrase chunking, and statistical postprocess­ing such as log likelihood which compares the term frequencies in a document against their ex­pected frequencies derived in a bigger text. Be­sides, extracting terms prefers syntactically plau­sible noun phrases (NPs) which are mainly multi-words terms. Kim and Kan (2009) report that most of key phrases are often simple words than less of­ten compound words.

The task for extracting key phrases tend to in­clude analyzing the document structure. Espe­cially, extracting key phrases from well-structured scientific articles should consider cross-section in­formation (Nguyen and Kan, 2007). This informa­tion has been explored to assess the suitability of features during learning in Kim and Kan (2009).

Extracting key phrases, however, is more than to extracting terminology or analyzing the document structure. While terms are words which appear in specific contexts and analyse concept structures in domains of human activity, key phrases are words that capture the key idea of documents. In addi­tion, while terms usually occur in the given doc­ument more often than we would expect to occur, key phrases do not necessarily occur frequently or key phrases do not occur at all in the document. Consequently, the task for extracting key phrases should not be considered as the subtask of extract­ing terminology and we are not able to directly ap­ply general-purpose term extractors to extract key phrases.

In this paper, we describe our method for "Au­tomatic Keyphrase Extraction from Scientific Artides", the shared task of SemEval-2 Evalua­tion Exercise which we participated in. Al­though term extractors along with linguistically-motivated analysis allow us to extract even elab­orated morpho-syntactic variation forms of terms, the naive statistic approach proposed in this pa­per is very simple and quite efficient for extracting keyphrases especially from well-structured scien­tific articles. In a nutshell, our method is based on empirical rules without any linguistically-motivated preprocessing. Empirical rules are ob­tained from the analysis of the characteristics of keyphrases by observing training data.

The remaining of this paper is organized as fol­lows: Section 2 explains the characteristics of keyphrases in scientific articles. Section 3 and 4 detail our naive statistic approach and experiment, respectively. We conclude this paper and discuss a further improvement in Section 6.

In this section, we investigate the characteristics of keyphrases in training data. Table 1 shows statis­tics of training data. In Table 1, D-author means the keyphrases assigned by authors, D-reader the keyphrases assigned by readers, and D-combined the combined keyphrases assigned by both of au­thors and readers.

Table 1 : Statistics of training data

This section shows the experiment results with training and test data.

In this paper, we described our simple method for extracting keyphrases from scientific arti­cles which we participate in the shared task of SemEval-2 Evaluation Exercise. The naive ap­proach was proposed. This approach turned out very simple and quite efficient for extracting keyphrases from well-structured scientific articles. Based on learning the distribution of keyphrases with section information, we obtain 18.34% for f-measure using top 15 candidates.

Our naive approach still has much room for improvement. For example, we are able to im­prove the result for same test data up to 20.71% and 25.55% for f-measure using top 15 candidates simply by adding the rest sections and normaliz­ing the number of occurrences of terms from each section.

Moreover, our n-word terms based extraction can be benefited by linguistic preprocessing such as normalizing surface forms. Handcrafted regu­lar expression rules along with part-of-speech tag­ging and phrase chunking would be also intro­duced to improve candidate selection. We have not explored thoroughly feature engineering, nei­ther. For example, more fine-grained section infor­mation and weight re-assignment might help filter out irrelevant candidates. We leave these possibil­ities for future work.