Generating research websites using summarisation techniques
Advaith Siddharthan & Ann Copestake Natural Language and Information Processing Group Computer Laboratory, University of Cambridge {as372,aac10}@cl.cam.ac.uk

Abstract
We describe an application that generates web pages for research institutions by summarising terms extracted from individual researchers' publication titles. Our online demo covers all researchers and research groups in the Computer Laboratory, University of Cambridge. We also present a novel visualisation interface for browsing collaborations.

1 Introduction
Many research organisations organise their websites as a tree (e.g., department pages  research group pages  researcher pages). Individual researchers take responsibility for maintaining their own web pages and, in addition, researchers are organised into research groups that also maintain a web page. In this framework, information easily gets outdated, and publications lists generally stay more up-to-date than research summaries. Also, as individuals maintain their own web pages, connections between researchers in the organisation are often hard to find; a surfer then needs to move up and down the tree hierarchy to browse the profiles of different people. Browsing is also diffcult because individual web pages are organised differently, since standardised stylesheets are often considered inappropriate for diverse organisations. Research summary pages using stylesheets can offer alternative methods of information access and browsing, aiding navigation and providing different views for different user needs, but these are timeconsuming to create and maintain by hand. We are exploring the idea of automatically generated and updated web pages that accurately reflect the research interests being pursued within a research institution. We take as input existing personal pages 5

from the Computer Laboratory, University of Cambridge, that contain publication lists in html. In our automatically generated pages, content (a research summary) is extracted from publication titles, and hence stays up-to-date provided individual researchers maintain their publication lists. Note that publication information is increasingly available through other sources, such as Google Scholar. We aim to format information in a way that facilitates browsing; a screen shot is shown in Figure 1 for the researcher Frank Stajano, who is a member of the Security and DTG research groups. The left of the page contains links to researchers of the same research groups and the middle contains a research profile in the form of lists of key phrases presented in five year intervals (by publication date). In addition, the right of the page contains a list of recommendations: other researchers with similar research interests. Web pages for research groups are created by summarising the research profiles of individual members. In addition, we present a novel interactive visualisation that we have developed for displaying collaborations with the rest of the world. In this paper we describe our methodology for identifying terms, clustering them and then creating research summaries (§2) and a generative summariser of collaborations (§4) that plugs into a novel visualisation (§3). An online demo is available at:
http://www.cl.cam.ac.uk/research/nl/webpage-demo/NLIP.html

2 Summarising research output
Our program starts with a list of publications extracted from researcher web pages; for example:
· S. Teufel. 2007. An Overview of evaluation methods in TREC Ad-hoc Information Retrieval and TREC Question Answering. In Evaluation of Text and Speech Systems. L. Dybkjaer, H. Hemsen, W. Minker (Eds.) Springer, Dordrecht (The Netherlands).

Proceedings of the ACL-08: HLT Demo Session (Companion Volume), pages 5­8, Columbus, June 2008. c 2008 Association for Computational Linguistics


From each publication entry such as that above, the program extracts author names, title and year of publication. This is the only information used. We do not use the full paper, as pdfs are not available for all papers in publication pages (due to copyright and other issues). The titles are then parsed using the RASP parser (Briscoe and Carroll, 2002) and keyphrases are extracted by pattern matching. From the publication entry above, the extracted title:
"An overview of evaluation methods in TREC ad-hoc information retrieval and TREC question answering"

2.1 Individual researcher summaries To create a web page for an individual researcher, the key-phrases extracted from all the paper titles authored by that researcher are clustered together based on similarity - an example cluster is shown below (from Karen Sparck Jones' profile):
`automatic classification for information retrieval', `intelligent automatic information retrieval', `information retrieval test collections', `information retrieval system', `automatic classification', `intelligent retrieval', `information retrieval', `information science', `test collections', `mail retrieval', `trec ad-hoc information retrieval'

produces five key-phrases:
`evaluation methods', `evaluation methods in TREC ad-hoc information retrieval', `TREC ad-hoc information retrieval', `TREC question answering', `information retrieval'

A representative phrase (most similar to others in the cluster) is selected from each cluster (`information retrieval' from the above) and this phrase is linked with all the publication dates for papers the terms in the cluster come from. These extracted keyphrases are enumerated as lists in five year intervals; for example (from Karen Sparck Jones' profile):
1990­1994: `information retrieval'; `document retrieval'; `video mail retrieval'; `automatic summarisation'; `belief revision'; `discourse structure'; `cambridge/olivetti retrieval system'; `system architecture'; `agent interaction'; `better NLP system evaluation'; `early classification work'; `text retrieval'; `discourse modelling'...;

Figure 1: Screenshot: researcher web page.
http://www.cl.cam.ac.uk/research/nl/webpage-demo/Frank Stajano.html

2.2 Recommendations (related people) Recommendations for related people are generated by comparing the terms extracted between 2000 and 2008 for each researcher in the Computer Laboratory. The (at most) seven most similar researchers are shown in tabular form along with a list of terms from their profiles that are relevant to the researcher being viewed. These term lists inform the user as to why they might find the related people relevant. 2.3 Research Group Pages

Figure 2: Screenshot: research group web page.
http://www.cl.cam.ac.uk/research/nl/webpage-demo/DTG.html

Group pages are produced by summarising the pages of members of the group. Terms from individual research profiles are clustered according to who is working on them (gleaned from the author lists of the the associated paper title). The group page is presented as a list of clusters. This presentation shows how group members collaborate, and for each term shows the relevant researchers, making navigation

6


easier. Two clusters for the Graphics and Interaction (Rainbow) Group are show below to illustrate:
`histogram warping'; `non-uniform b-spline subdivision'; `stylised rendering'; `multiresolution image representation'; `human behaviour'; `subdivision schemes'; `minimising gaussian curvature variation near extraordinary vertices'; `sampled cp surfaces'; `bounded curvature variants': Neil Dodgson; Thomas Cashman; Ursula Augsdorfer; `text for multiprojector tiled displays'; `tabletop interface'; `high-resolution tabletop applications'; `distributed tabletops'; `remote review meetings'; `rapid prototyping': Peter Robinson; Philip Tuddenham;

Figure 3: Screenshot: Visualisation of Collaboration between the NLIP Group and the rest of the world

3 Visualisation
Scalable Vector Graphics (SVG)1 is a language for describing two-dimensional graphics and graphical applications in XML. Interactive images such as those in Figure 3 are produced by an XSLT script that transforms an input XML data file containing information about collaborations and latitudes and longitudes of cities and countries into an SVG representation2 . This can be viewed through an Adobe Browser Plugin3 . In the map, circles indicate the locations of co-authors of members of the NLIP research group, their size being proportional to the number of co-authors at that location. The map can be zoomed into, and at sufficient zoom, place names are made visible. Clicking on a location (circle) provides a summary of the collaboration (the summarisation is described in §4), while clicking on a country (oval) provides a contrywise overview such as:
In the Netherlands, the NLIP Group has collaborators in Philips Research (Eindhoven), University of Twente (Enschede), Vrije Universiteit (VU) (Amsterdam) and University of Nijmegen. Figure 4: Screenshot: Visualisation of Collaborations of ARG Group; zoomed into Europe and having clicked on Catonia (Italy) for a popup summary

4 Summarising collaborations
Our summarisation module slots into the visualisation interface; an example is shown in Figure 4. The aim is to summarise the topics that members of the research group collaborate with the researchers in
http://www.w3.org/Graphics/SVG/ Author Affiliations and Latitudes/Longitudes are semiautomatically extracted from the internet and hand corrected. The visualisation is only available for some research groups. 3 http://www.adobe.com/svg/viewer/install/main.html
2 1

each location on. The space constraints are dictated by the interface. To keep the visualisation clean, we enforce a four sentence limit for the summaries. There are four elements that each sentence contains-- names of researchers in research group, names of researchers at location, terms that summarise the collaboration, and years of collaboration. Our summaries are produced by an iterative process of clustering and summarising. In the first step, terms (key phrases) are extracted from all the papers that have co-authors in the location. Each term is tagged with the year(s) of publication and the names of researchers involved. These terms are then clustered based on the similarity of words in the terms and the similarity of their authors. Each such cluster contributes one sentence to the summary. The clustering process is pragmatic; the four sentence per summary limit means that at most four clusters should be formed. This means coarser clustering (fewer and larger clusters) for locations with many collaborations and finer-grained (more and smaller clusters) for locations with fewer collaborations. The next step is to generate a sentence from each cluster. In this step, the terms in a sentence cluster are reclustered according to their date tag. then each time period is realised separately within the sentence, for example:

7


Lawrence C Paulson collaborated with Cristiano Longo and Giampaolo Bella from 1997 to 2003 on `formal verification', `industrial payment and nonrepudiation protocol', `kerberos authentication system' and `secrecy goals' and in 2006 on `cardholder registration in Set' and `accountability protocols'.

To make the summaries more readable, lists of conjunctions are restricted to a maximum length of four. Terms are incorporated into the list in decreasing order of frequency of occurrence. Splitting the sentence above into two time periods allows for the inclusion of more terms, without violating the restriction on list length. This form of sentence splitting is also pragmatic and is performed more aggressively in summaries with fewer sentences, having the effect of making short summaries slightly longer. Another method for increasing the number of terms is by aggregating similar terms. In the example below, three terms (video mail retrieval, information retrieval and document retrieval) are aggregated into one term. Thus six terms have made it to the clause, while keeping to the four terms per list limit.
In the mid 1990s, K Sparck Jones, S J Young and M G Brown collaborated with J T Foote on `video mail, information and document retrieval', `cambridge/olivetti retrieval system', `multimedia documents' and `broadcast news'.

sentence extraction followed by some rewrite operation (e.g., sentence shortening (Vanderwende et al., 2007; Zajic et al., 2006; Conroy et al., 2004), aggregation (Barzilay, 2003) or reference regeneration (Siddharthan et al., 2004; Nenkova and McKeown, 2003)). In contrast, our system does not extract sentences at all; rather, it extracts terms from paper titles and our summaries are produced by clustering, summarising, aggregating and generalising over sets of terms and people. Our space constraints are dictated by by our visualisation interface, and our program employs pragmatic clustering and generalisation based on the amount of information it needs to summarise.

Acknowledgements
This work was funded by the Computer Laboratory, University of Cambridge, and the EPSRC (EP/C010035/1 and EP/F012950/1).

References
R. Barzilay. 2003. Information Fusion for Multidocument Summarization: Paraphrasing & Generation. Ph.D. thesis, Columbia University. E.J. Briscoe and J. Carroll. 2002. Robust accurate statistical annotation of general text. In Proceedings of the 3rd International Conference on Language Resources and Evaluation, pages 1499­1504, Las Palmas, Gran Canaria. J.M. Conroy, J.D. Schlesinger, J. Goldstein, and D.P. O'Leary. 2004. Left-brain/right-brain multidocument summarization. Proceedings of DUC 2004. A. Nenkova and K. McKeown. 2003. References to named entities: a corpus study. Companion proceedings of HLT-NAACL 2003­short papers-Volume 2, p ag es 7 0 ­ 7 2 . A. Siddharthan, A. Nenkova, and K. McKeown. 2004. Syntactic simplification for improving content selection in multi-document summarization. In Proceedings of the 20th International Conference on Computational Linguistics (COLING 2004), pages 896­902, Geneva, Switzerland. L. Vanderwende, H. Suzuki, C. Brockett, and A. Nenkova. 2007. Beyond SumBasic: Taskfocused summarization with sentence simplification and lexical expansion. Information Processing and Management, 43(6):1606­1618. D. Zajic, B. Dorr, J. Lin, and R. Schwartz. 2006. Sentence Compression as a Component of a MultiDocument Summarization System. Proceedings of the 2006 Document Understanding Workshop, New York.

The four word limit is also enforced on lists of people. If there are too many people, the program refers to them by affiliation; for example:
Joe Hurd collaborated with University of Utah on `theorem proving', `encryption algorithms', `functional correctness proofs' and `Arm verification'.

5 Discussion and Conclusions
Our summarisation strategy mirrors the multidocument summarisation strategy of Barzilay (2003), where sentences in the input documents are clustered according to their similarity. Larger clusters represent information that is repeated more often; hence the size of a cluster is indicative of importance. The novelty of our application is that this strategy has been used at a sub-sentential level, to summarise terms that are then used to generate sentences. While there has been research on generative summarisation, much of this has been focused on 8