Making histoGraph open source

As many of you already know, histoGraph is a web platform designed to help researchers to explore large multimedia archives akin to what we have here at the CVCE. The histoGraph tool has two main goals:

  1. To enable users to find and identify the most relevant documents for research
  2. To discover  document connections between persons.

In the last few months our designer/developer Daniele Guido, has started the redesign and development of the tool: this screencast shows some of the new features of Histograph as well as identifying the benefits and limitations of the new design.

We can now explore the “neighborhood” of a person in terms of other co-occurring people and documents, follow the paths that connect one person to one another; or simply search for resources and pose questions related to the person or the document.

Still under development, the open source version of Histograph starts from this assumption and basically serves as recommendation system. The original database has been enriched with CVCE text documents and been transposed to a graph database in order to speed up all network related computation. Moreover, a simple and powerful text extraction chain has been added:  thanks to the powerful Yago disambiguation engine, all text captions, titles and abstracts have been annotated with Named Entities in at least two languages.

Additionally, dates of both pictures and text documents have been reconciled to ISO standard; geographical places to latitude and longitude by using Geonames and Google Geocoding API; finally, person entities have been enriched with related dbpedia information. We will shortly be updating the data curation workflow to  assess and validate the quality of the automatic extraction of these entities and we plan further develop our concept of expert crowdsourcing techniques.

We will use the blog to update you on progress so watch this space.

CVCE at Talk of Europe Creative Camp with CRP Lippmann & Lab1100

The Talk of Europe Creative Camp was organised by Max Kemman and Astrid van Aggelen and featured research projects and presentations by 16 researchers in Computer Science and the Humanities – a big thanks for all their hard work!

What do you do when you have all the debates in the EU parliament from 1992 onwards digitized, in RDF and you are just one SPARQL query away from them? If you have one week to work with specialists in network data visualization and management? First of all you need to bridge the gap between domain experts and ask how this data can be used to answer research questions in European studies and Computer Science. This turned out to be as hard as expected.

This week nevertheless yielded a method which helps us to detect unexpected speaker appointments by Members of the European Parliament (created by Fintan McGee of our neighbouring research institute CRP Lippmann) and an import functionality for Nodegoat which makes it easy to pull and visualise data from the Talk of Europe’s SPARQL endpoint (created by Pim van Bree and Geert Kessels of Lab1100). For CVCE this was an excellent opportunity to experiment with new ways of extracting and visualising information from structured data repositories.

For starters: A little of bit on networks

To explain our work a short intro is needed: Graphs like the one below are so-called 1-mode networks. This means that there is only one type of nodes (in this case people) connected to each other. Imagine that the people in this network are connected to each other because they all are members of the same sports club.

Mac OS X:Users:marten.duering:Desktop:1mode.png

In this example, a 1-mode network visualisations represents group affiliations by showing links between all members of a group. This works fine here but very quickly becomes cluttered when a group has a large number of members.

You can also represent this kind of information using a 2-mode network; a network which has two types of actors (people, sports club). It is important to mention that in 2-mode networks ties can only exist between actors of different types (person -> sport club), never within one type of actors (person -> person).

Mac OS X:Users:marten.duering:Desktop:2mode.png

2-mode networks are also sometimes called affiliation networks or bimodal networks and are particularly well-suited to identify overlaps between groups as illustrated in the graph above. In both examples it is clear that Laura is the person who is part of both groups. 2-mode visualisations are however often leaner than 1-mode networks since they represent an affiliation of a person by one tie alone and merely imply the links between members of a group.

Both 1-mode and 2-mode networks can be described mathematically as well. There are numerous ways to describe how central a node is in a network and how different nodes cluster together. Fintan’s work focuses on algorithms which are used for 2-mode networks and networks with more than 2 types of actors, so-called multimodal networks. 2-mode networks can be projected into 1-mode networks. This means that instead of 2 types of nodes we end up with only one. A tie between these nodes is added when – in our example – two people are members of the same club. This works just fine in the above examples. But as as Fintan highlights, projections can lead to a loss of information: One actor may be a member of three clubs and end up with the same connections as someone who only attends two clubs. Redundancy is one way to measure this loss of information.

Redundancy may reveal irregularities

Getting the opportunity to speak is a privilege in the EU parliament and very often the same people are chosen to speak on their area of expertise. But sometimes they don’t. Domain experts Frédéric Allemand (CVCE) and Bjørn Høyland (Oslo University) showed interest to detect the latter. It turned out that Redundancy, a concept developed in graph theory might do just that, as Fintan discovered during the week.

Drastically simplified, Redundancy describes how irreplaceable a node is in a 2-mode graph. In other words: Nodes which can be removed from a graph without changing its overall structure when it is projected into a single mode are considered to be highly redundant.

The graph below was created by Fintan and shows connections between Members of the European Parliament who spoke on certain agenda items in the year 2010 (527 nodes, 595 edges). An interactive version of this graph, realized with the sigmajs plugin for Gephi, is also available.

                Mac OS X:Users:marten.duering:Desktop:toe-fig1.png

Click here to launch an interactive version of this Graph

Only those MEPs are shown who have a low level of redundancy (<0.5) which means that speakers with very atypical combinations of topics on which they speak will be highlighted. An example is Mike Nattrass (United_Kingdom) who spoke on contrasting subjects such as:

  • Ban on commercial whaling (debate)

  • Implementation of the first railway package directives (debate)

  • Progress made on resettling Guantanamo detainees and on closing Guantanamo (debate)

  • Protection of animals used for scientific purposes (debate)

  • Welfare of laying hens (debate)

This is where our journey ends and domain experts need to take a close look at the data to identify the meaning of the graph and link back their observations to the context knowledge which is and remains essential for the humanist analysis of the parliamentary debates.

From SPARQL to Network via Nodegoat

For our second project, Geert Kessels and Pim van Bree of Lab1100 have developed a new feature for Nodegoat which allows users to load data directly from the Talk of Europe SPARQL endpoint (or any other). This will make it significantly easier to explore the debate data for anyone who can write a query. Nodegoat might well be the to-date easiest way to set up a relational database, fill it with data and visualise the data as a social or spatial network and to observe changes over time.

That said, it is no trivial task to write such a query. But its advantage lies in the precision with which one can download a very particular constellation of data points.

User can define custom import templates which define how the data will be organized and linked to each other in Nodegoat.

Once these two steps are sorted, the data is ready to be visualised. Nodegoat supports the visualisation of social graphs but also combinations of social graphs and maps. You can check out a few examples on their website and use this User Guide to setup your own project. Sadly we ran out of time to test this new feature and develop queries together with our domain expert.

All in all, this was a great opportunity to start a collaboration between the CVCE DH Lab, CVCE EIS and Fintan of CRP Lippmann and Geert and Pim of Lab1100.

Talk of Europe – Travelling CLARIN Campus

In October we will be participating in a one week creative camp, organised by the Talk of Europe – Travelling CLARIN Campus project (ToE-TCC). Funded by Dutch agencies NWO and OCW, the project follows up on the PoliMedia project, which linked entities in Dutch parliamentary records to corresponding media sources archived by the Netherlands Institute for Sound and Vision.

The CVCE was successful with two proposals submitted to ToE-TC, both focused on exploring social graphs.

  1. Visualising Co-occurrence networks of the plenary debates of the European Parliament
  2. Exploring the potential of multimodal graphs for modelling European Plenary Debates

We are excited to participate in travelling campus, as we are looking forward to finding new ways of structuring parliamentary data. We look forward to exploring how we can use such tools, techniques and data structures to shape future developments of the CVCE research infrastructure and develop new ways of exploring our collections.

There are three aims for the one week event, scheduled for 6-10 October 2014:

First Aim: “To translate the proceedings of the European Parliament debates (europarl) to CLARIN standards. In other words, the europarl data is curated to linked data so that it can be linked to and reused by other datasets and services. An interesting aspect of this dataset is that its available in 21 languages, enabling to link to other datasets in as many different languages. (…)

Second Aim: To research how this linked data can be exploited for use by humanities scholars. The combination of linked datasets with the base available in 21 languages should allow research questions not yet feasible. As such, interlingual comparative research through digital tools becomes possible. In comparison, where PoliMedia focused on how Dutch debates were covered in the Dutch media, the europarl data could be analysed to compare how European Parliament debates are covered in Dutch versus Polish media. Political scientists could analyse how subjects from EP debates relate to subjects in debates of national parliaments. In order to build tools helpful, we will investigate user requirements and possible research questions from humanities scholars. This second aim will be undertaken by Max Kemman, Martijn Kleppe and Henri Beunders from the History department of EUR.

Third Aim: To spark a transnational, European collaboration to create tools for scholars to analyse linked datasets. To this end, three meetings of a week will be organised in which teams from all CLARIN countries are invited to participate. Think hackathon, but a week long and with a European focus.”

The European parliamentary debates have been processed and are now available in RDF format. RDF, commonly associated with the buzz-word “Linked data“, “is based upon the idea of making statements about resources (in particular web resources) in the form of subject–predicate–object expressions. These expressions are known as triples in RDF terminology” (Wikipedia article on RDF). RDF is highly flexible and allows users to formulate very precise queries on their datasets.

A simple social network graph

The first proposal, in cooperation with Dr Fintan McGee of our neighbouring research centre CRP Lippmann, will explore the extent to which multimodal network analysis can be fruitfully applied to the study of the parliamentary debates. Where one-mode network visualisations can only represent one type of actor in a network (e.g. people), multimodal networks are capable of representing multiple types of actors, e.g. people, institutions, texts and places. This approach bears the potential to represent more closely the complexities of real-world networks.

An example of a geographic social network
(source: Nodegoat).

The second proposal was submitted with Lab1100, a Dutch company which develops Nodegoat. Nodegoat offers its users with simple yet powerful instruments to create relational databases online and visualise their data in form of maps and network graphs. Lab1100 will develop a way to link ToE-TCC data directly to Nodegoat using the formers SPARQL endpoint. This will ensure very quick and highly flexible data import and visualisations of the data.

CVCE researcher Marten Düring and Frédéric Allemand, Coordinator of the European Integration Studies Department at CVCE will participate in the project.