Fraunhofer ISI Discussion Papers Innovation Systems and Policy Analysis 42
[Conclusion] We presented an approach for building clusters of scientific publications according to the information given by their title, authors and reference list and for selecting a list of ET candidates from these clusters. The approach was trained with information from the years 2000 and 2001 and evaluated for the years 2002 to 2007, which correspond to the years in our dataset in which ETs actually emerge. The information that would have to be processed manually could be reduced to less than 0.14 % of the dataset. We experienced problems because the actual relevance of the hay documents in the train and test set was still unknown. A hay document still could represent an ET that was not included in our list when creating these artificial haystacks. It could also be an outlier due to a deviant title, reference list etc. For instance, in the ET candidate list for dataset 3, there were documents (or 1-instance clusters) that dealt with automatically recognizing traffic signs and extracting validation rules from microbiological data. Such documents might indeed represent a topic that might have been established as a new scientific topic. Or they could simply represent a novel application for established methodologies. A clear trend in the ET candidate list for dataset 3 could be identified for grid systems which was then a newly established scientific topic. Our evaluation showed that even-though we included citations as a second non-textual feature for LDA, the approach was still prone to disruptions based on insufficient or misleading information in both titles and reference lists. An improvement of the clustering step in future work with further extensions of LDA would also influence and probably improve the connection step. Nonetheless, this step could be further improved with a more specific analysis of the individual connections made. Since our results were worse for more recent years, the connection step could be influenced by the sheer number of connection candidates, since for each year approximately 500 cluster candidates are added. Another factor for future work would be to investigate upon the fact that the performance for dataset 3 was significantly better than for the other datasets. All in all, the results suggest that the composition of our approach and its implementation are promising, but further investigations have to be made in terms of extensibility and appropriate evaluation environments.