Legal Search and Recommendation

This Friday we'll have two talks followed by drinks. The topic of this meetup is Legal Search and Recommendation. We have two great talks:

The industrial talk will be given by Legal Intelligence (https://www.legalintelligence.com/), the largest legal search engine of the Netherlands.

Our academic speaker is Radboud Winkels (http://www.uva.nl/profiel/w/i/r.g.f.winkels/r.g.f.winkels.html), the dean of the PPLE (http://pple.uva.nl/) at the University of Amsterdam, and associate professor at the Leibniz Center for Law (http://www.leibnizcenter.org/). He will talk about his research into Legal Recommender Systems which he coordinated for openlaws.eu (https://info.openlaws.com/openlaws-eu/).

This edition of SEA will be held in at Science Park 904 in Room D1.115.

Program:
16:00 - 16:30 Radboud Winkels
16:30 - 17:00 Legal Intelligence
17:00 - 18:00 Drinks & Snacks

Details of the talks:

Legal Intelligence

Legal Intelligence is a Dutch legal search engine allowing legal professionals (e.g. lawyers, judges) to search though a large data repository containing content of various publishers. With a 70% market share and 100,000+ users today, the Legal Intelligence repository contains roughly 8 million relevant legal documents harvested using web-crawlers (public data), delivered by legal publishers or by the firms themselves (internal data). Legal Intelligence has been running on Solr since 2008. Over the years, Legal Intelligence has optimized search for our specific group of users by adding domain specific features.

One of Legal Intelligence’s characteristics is the application of thesauri while indexing and searching, using a combination of domain driven text analysis and query parsing/expanding. Legal Intelligence recognizes topics from the legal vocabulary, synonyms of courts, references to law articles and legal publications. This approach results in well-balanced recall and precision.

We have realized this using Solr by writing a custom query parser and analysis chain, including a token filter that takes care of advanced Finite State Transducer based thesaurus matching. The talk discusses how we have implemented this in Solr and share the challenges we encountered. We will also explain how we are using Solr’s function queries in order to give our users the best possible ranking of search result. Finally, we will give some insights what we did to create an automated test suite for these features. Lastly, we will address some of the next steps envisioned like meta-data enrichment via a learned classification model and legal relevancy using similarity based upon latent semantic analysis.

This talk will be given by Pieter van Boxtel and Jaco de Vroed. Pieter is a Java developer at Legal Intelligence mainly working on its search implementation and indexing pipeline. Jaco has been involved in Legal Intelligence since the start of the company, working on all aspects of software development with a strong focus on search.

----------------------------------------------------------------------------------------

Radboud Winkels -- Towards a Legal Recommender System

We aim to suggest relevant new sources of law to users of legal portals based on the documents they are focusing on at a certain moment in time, or those they have selected. In the future we attempt to do this both based on ‘objective’ features of the documents themselves and on ‘subjective’ information gathered from other users (‘crowdsourcing’). At this moment we concentrate on the first method. I will describe how we create the web of law if it is not available in machine readable form, or extend it when that is necessary. Next, I present results of experiments using analysis of the network of references or citations to suggest these new documents. We experiment with mixing network analysis with similarity based on the comparison of the actual text of documents. One experiment is based on simple bag-of-words and normalisation, the other uses Latent Dirichlet Allocation (LDA) with added n-grams. A small formative evaluation in both experiments suggests that text similarity alone works better than network analysis alone or a combination, at least for Dutch court decisions. In the discussion we may focus on where the legal domain differs from others and what this means for interpreting results of search and analysis techniques.

Radboud Winkels, the dean of the PPLE at the University of Amsterdam, and associate professor at the Leibniz Center for Law. He will talk about his research into Legal Recommender Systems which he coordinated for openlaws.eu.

----------------------------------------------------------------------------------------