The increasing amount of documents in elec­tronic form makes imperative the need for docu­ment content classification and semantic labelling. Keyphrase extraction contributes to this goal by the identification of important and discriminative concepts expressed as keyphrases. Keyphrases as reduced document content representations may find applications in document retrieval, classifica­tion and summarisation (D'Avanzo and Magnini, 2005). The literature distinguishes between two principal processes: keyphrase extraction and keyphrase assignment. In the case of keyphrase assignment, suitable keyphrases from an exist­ing knowledge resource, such as a controlled vo­cabulary, or a thesaurus are assigned to docu­ments based on classification of their content. In keyphrase extraction, the phrases are mined from the document itself. Supervised approaches to the problem of keyphrase extraction include the Naive Bayes-based KEA algorithms (Gordon et al., 1999) (Medelyan and Witten, 2006), deci­sion tree-based and the genetic algorithm-based GenEx (Turney, 1999), and the probabilistic KL divergence-based language model (Tomokiyo and Hurst, 2003). Research in keyphrase extrac­tion proposes the detection of keyphrases based on various statistics-based, or pattern-based features. Statistical measures investigated focus pri­marily on keyphrase frequency measures, whereas pattern-features include noun phrase pattern filter­ing, identification of keyphrase head and respec­tive frequencies (Barker and Cornacchia, 2000), document section position of the keyphrase (e.g., (Medelyan and Witten, 2006)) and keyphrase coherence (Turney, 2003). In this paper, we present an unsupervised approach which combines pattern-based morphosyntactic rules with a statis­tical measure, the C-value measure (Frantzi et al., 2000) which originates from research in the field of automatic term recognition and was initially de­signed for specialised domain terminology acqui­sition.

The input documents in the Keyphrase Extrac­tion task were scientific articles converted from their originally published form to plain text. Due to this process, some compound hyphen­ated words are erroneously converted into a single word (e.g., "resourcemanagement" vs. "resource-management"). Moreover, document sections such as tables, figures, footnotes, headers and foot­ers, often intercept sentence and paragraph text. Finally, due to the particularity of the scientific ar­ticles domain, input documents often contain ir­regular text, such as URLs, inline bibliographic references, mathematical formulas and symbols. In our approach, we attempted to address some of these issues by document structuring, treatment of orthographic variation and filtering of irregular text.

The approach adopted first applies part-of-speech tagging and basic document structuring (sec. 2.1 and 2.2). Subsequently, keyphrase can­didates conforming to pre-defined morphosyntac­tic rule patterns are identified (sec. 2.3). In the next stage, orthographic, morphological and abbreviation variation phenomena are addressed (sec. 2.4) and, finally, candidate keyphrases are selected based on C-value statistical measure (sec. 2.5).

The overall official results of the UvT system are shown in Table 1, where P, R and F correspond to micro-averaged precision, recall and F-score for the respective sets of candidate keyphrases, based on reader-assigned and combined author-and reader-assigned gold standards. Table 1 also illustrates the reported performance of the task baseline systems (i.e., TF IDF, Naive Bayes (NB) and maximum entropy (ME) ) and the UvT sys­tem performance variance based on document sec­tion candidates (-A: Abstract, -I: Introduction, -M: Main, -IC: Introduction and Conclusion combina­tion). In these system variants, rather than select­ing the top 15 C-value candidates from the sys­tem output, we also apply restrictions based on the candidate keyphrase document section infor­mation, thus skipping candidates which do not ap­pear in the respective document section.

Overall, the UvT system performance is close to the baseline systems results. We observe that the system exhibits higher performance for its top 5 candidate set and this performance drops rapidly as we include more terms in the answer set. One possible reason for its average performance could be attributed to increased "noise" in the results set. In particular, our text filtering method failed to ac­curately remove a large amount of irregular text in form of mathematical formulas and symbols which were erroneously tagged as proper nouns. As indicated in Table 1, the improved results of system variants based on document sections, such as Abstract, Introduction and Conclusion, where these symbols and formulas are rather uncommon, could be partly attributed to "noise" reduction.

Interestingly, the best system performance in these document section results is demon­strated by the Introduction-Conclusion com­bination (UvT-IC). Other tested combinations (not illustrated in Table 1), such as abstract-intro, abstract-intro-conclusions, abstract-intro-conclusions-references, display similar results on the reader-assigned set and a performance rang­ing between 15,6-16% for the 15 candidates on the combined set, while the inclusion of the Main section candidates reduces the performance to the overall system output (i.e., UvT results). Further experiments are required for refining the criteria for document section information, when the text filtering process for "noise" is improved.

Finally, another reason that contributes to the system's average performance lies in its inherent limitation for the detection of multi-word phrases, rather than both single and multi-word. In partic­ular, single word keyphrases account for approx. 20% of the correct keyphrases in the gold standard sets.

We have presented an approach to keyphrase ex­traction mainly based on adaptation and imple­mentation of the C-value method. This method was originally proposed for the detection of ter­minological phrases and although domain terms may express the principal informational content of a scientific article document, a method designed for their exhaustive identification (including both nested and longer multi-word terms) has not been proven more effective than baseline methods in the keyphrase detection task. Potential improve­ments in performance could be investigated by (1) improving document structure detection, so as to reduce irregular text, (2) refinement of document section information in keyphrase selection, (3) adaptation of the C-value measure, so as to possibly combine keyphrase frequency with a dis­criminative measure, such as idf.