skip navigation

Prosumption: shifting the barriers between information producers and consumers

One of the major revolutions of the Internet era has been the shifting of the frontiers between producers and consumers [1]. Prosumption refers to the emergence of a new category of actors who not only consume but also contribute to content creation and sharing. Under the umbrella of Web 2.0, many sites indeed enable users to share multimedia content, data, experiences [2], views and opinions on different issues, and even to act cooperatively to solve global problems [3]. Web 2.0 has become a fertile terrain for the proliferation of valuable user data enabling user profiling, opinion mining, trend and crisis detection, and collective problem solving [4].

The private sector has long understood the potentialities of user data and has used them for analysing customer preferences and satisfaction, for finding sales opportunities, for developing marketing strategies, and as a driver for innovation. Recently, corporations have relied on Web platforms for gathering new ideas from clients on the improvement or the development of new products and services (see for instance Dell’s Ideastorm; salesforce’s IdeaExchange; and My Starbucks Idea). Similarly, Lego’s Mindstorms encourages users to share online their projects on the creation of robots, by which the design becomes public knowledge and can be freely reused by Lego (and anyone else), as indicated by the Terms of Service. Furthermore, companies have been recently mining social network data to foresee future action of the Occupy Wall Street movement.

Even scientists have caught up and adopted collaborative methods that enable the participation of laymen in scientific projects [5].

Now, how far has government gone in taking up this opportunity?

Some recent initiatives indicate that the public sector is aware of the potential of the “wisdom of crowds.” In the domain of public health, MedWatcher is a mobile application that allows the general public to submit information about any experienced drug side effects directly to the US Food and Drug Administration. In other cases, governments have asked for general input and ideas from citizens, such as the brainstorming session organized by Obama government, the wiki launched by the New Zealand Police to get suggestions from citizens for the drafting of a new policing act to be presented to the parliament, or the Website of the Department of Transport and Main Roads of the State of Queensland, which encourages citizens to share their stories related to road tragedies.

Even in so crucial a task as the drafting of a constitution, government has relied on citizens’ input through crowdsourcing [6]. And more recently several other initiatives have fostered crowdsourcing for constitutional reform in Morocco and in Egypt .

It is thus undeniable that we are witnessing an accelerated redefinition of the frontiers between experts and non-experts, scientists and non-scientists, doctors and patients, public officers and citizens, professional journalists and street reporters. The ‘Net has provided the infrastructure and the platforms for enabling collaborative work. Network connection is hardly a problem anymore. The problem is data analysis.

In other words: how to make sense of the flood of data produced and distributed by heterogeneous users? And more importantly, how to make sense of user-generated data in the light of more institutional sets of data (e.g., scientific, medical, legal)? The efficient use of crowdsourced data in public decision making requires building an informational flow between user experiences and institutional datasets.

Similarly, enhancing user access to public data has to do with matching user case descriptions with institutional data repositories (“What are my rights and obligations in this case?”; “Which public office can help me”?; “What is the delay in the resolution of my case?”; “How many cases like mine have there been in this area in the last month?”).

From the point of view of data processing, we are clearly facing a problem of semantic mapping and data structuring. The challenge is thus to overcome the flood of isolated information while avoiding excessive management costs. There is still a long way to go before tools for content aggregation and semantic mapping are generally available. This is why private firms and governments still mostly rely on the manual processing of user input.

The new producers of legally relevant content: a taxonomy

Before digging deeper into the challenges of efficiently managing crowdsourced data, let us take a closer look at the types of user-generated data flowing through the Internet that have some kind of legal or institutional flavour.

One type of user data emerges spontaneously from citizens’ online activity, and can take the form of:

  • citizens’ forums
  • platforms gathering open public data and comments over them (see for instance data-publica )
  • legal expert blogs (blawgs)
  • or the journalistic coverage of the legal system.

User data can as well be prompted by institutions as a result of participatory governance initiatives, such as:

  • crowdsourcing (targeting a specific issue or proposal by government as an open brainstorming session)
  • comments and questions addressed by citizens to institutions through institutional Websites or through e-mail contact.

This variety of media supports and knowledge producers gives rise to a plurality of textual genres, semantically rich but difficult to manage given their heterogeneity and quick evolution.

Managing crowdsourcing

The goal of crowdsourcing in an institutional context is to extract and aggregate content relevant for the management of public issues and for public decision making. Knowledge management strategies vary considerably depending on the ways in which user data have been generated. We can think of three possible strategies for managing the flood of user data:

Pre-structuring: prompting the citizen narrative in a strategic way

A possible solution is to elicit user input in a structured way; that is to say, to impose some constraints on user input. This is the solution adopted by IdeaScale, a software application that was used by the Open Government Dialogue initiative of the Obama Administration. In IdeaScale, users are asked to check whether their idea has already been covered by other users, and alternatively to add a new idea. They are also invited to vote for the best ideas, so that it is the community itself that rates and thus indirectly filters the users’ input.

The MIT Deliberatorium, a technology aimed at supporting large-scale online deliberation, follows a similar strategy. Users are expected to follow a series of rules to enable the correct creation of a knowledge map of the discussion. Each post should be limited to a single idea, it should not be redundant, and it should be linked to a suitable part of the knowledge map. Furthermore, posts are validated by moderators, who should ensure that new posts follow the rules of the system. Other systems that implement the same idea are featurelist and Debategraph [7].

While these systems enhance the creation and visualization of structured argument maps and promote community engagement through rating systems, they present a series of limitations. The most important of these is the fact that human intervention is needed to manually check the correct structure of the posts. Semantic technologies can play an important role in bridging this gap.

Semantic analysis through ontologies and terminologies

Ontology-driven analysis of user-generated text implies finding a way to bridge Semantic Web data structures, such as formal ontologies expressed in RDF or OWL, with unstructured implicit ontologies emerging from user-generated content. Sometimes these emergent lightweight ontologies take the form of unstructured lists of terms used for tagging online content by users. Accordingly, some works have dealt with this issue, especially in the field of social tagging of Web resources in online communities. More concretely, different works have proposed models for making compatible the so-called top-down metadata structures (ontologies) with bottom-up tagging mechanisms (folksonomies).

The possibilities range from transforming folksonomies into lightly formalized semantic resources (Lux and Dsinger, 2007; Mika, 2005) to mapping folksonomy tags to the concepts and the instances of available formal ontologies (Specia and Motta, 2007; Passant, 2007). As the basis of these works we find the notion of emergent semantics (Mika, 2005), which questions the autonomy of engineered ontologies and emphasizes the value of meaning emerging from distributed communities working collaboratively through the Web.

We have recently worked on several case studies in which we have proposed a mapping between legal and lay terminologies. We followed the approach proposed by Passant (2007) and enriched the available ontologies with the terminology appearing in lay corpora. For this purpose, OWL classes were complemented with a has_lexicalization property linking them to lay terms.

The first case study that we conducted belongs to the domain of consumer justice, and was framed in the ONTOMEDIA project. We proposed to reuse the available Mediation-Core Ontology (MCO) and Consumer Mediation Ontology (COM) as anchors to legal, institutional, and expert knowledge, and therefore as entry points for the queries posed by consumers in common-sense language.

The user corpus contained around 10,000 consumer questions and 20,000 complaints addressed from 2007 to 2010 to the Catalan Consumer Agency. We applied a traditional terminology extraction methodology to identify candidate terms, which were subsequently validated by legal experts. We then manually mapped the lay terms to the ontological classes. The relations used for mapping lay terms with ontological classes are mostly has_lexicalisation and has_instance.

A second case study in the domain of consumer law was carried out with Italian corpora. In this case domain terminology was extracted from a normative corpus (the Code of Italian Consumer law) and from a lay corpus (around 4000 consumers’ questions).

In order to further explore the particularities of each corpus respecting the semantic coverage of the domain, terms were gathered together into a common taxonomic structure [8]. This task was performed with the aid of domain experts. When confronted with the two lists of terms, both laypersons and technical experts would link most of the validated lay terms to the technical list of terms through one of the following relations:

  • Subclass: the lay term denotes a particular type of legal concept. This is the most frequent case. For instance, in the class objects, telefono cellulare (cell phone) and linea telefonica (phone line) are subclasses of the legal terms prodotto (product) and servizio (service), respectively. Similarly, in the class actors agente immobiliare (estate agent) can be seen as subclass of venditore (seller). In other cases, the linguistic structures extracted from the consumers’ corpus denote conflictual situations in which the obligations have not been fulfilled by the seller and therefore the consumer is entitled to certain rights, such as diritto alla sostituzione (entitlement to a replacement). These types of phrases are subclasses of more general legal concepts such as consumer right.
  • Instance: the lay term denotes a concrete instance of a legal concept. In some cases, terms extracted from the consumer corpus are named entities that denote particular individuals, such as Vodafone, an instance of a domain actor, a seller.
  • Equivalent: a legal term is used in lay discourse. For instance, contratto (contract) or diritto di recessione (withdrawal right).
  • Lexicalisation: the lay term is a lexical variant of the legal concept. This is the case for instance of negoziante, used instead of the legal term venditore (seller) or professionista (professional).

The distribution of normative and lay terms per taxonomic level shows that, whereas normative terms populate mostly the upper levels of the taxonomy [9], deeper levels in the hierarchy are almost exclusively represented by lay terms.

Term distribution per taxonomic level

The result of this type of approach is a set of terminological-ontological resources that provide some insights on the nature of laypersons’ cognition of the law, such as the fact that citizens’ domain knowledge is mainly factual and therefore populates deeper levels of the taxonomy. Moreover, such resources can be used for the further processing of user input. However, this strategy presents some limitations as well. First, it is mainly driven by domain conceptual systems and, in a way, they might limit the potentialities of user-generated corpora. Second, they are not necessarily scalable. In other words, these terminological-ontological resources have to be rebuilt for each legal subdomain (such as consumer law, private law, or criminal law), and it is thus difficult to foresee mechanisms for performing an automated mapping between lay terms and legal terms.

Beyond domain ontologies: information extraction approaches

One of the most important limitations of ontology-driven approaches is the lack of scalability. In order to overcome this problem, a possible strategy is to rely on informational structures that occur generally in user-generated content. These informational structures go beyond domain conceptual models and identify mostly discursive, emotional, or event structures.

Discursive structures formalise the way users typically describe a legal case. It is possible to identify stereotypical situations appearing in the description of legal cases by citizens (i.e., the nature of the problem; the conflict resolution strategies, etc.). The core of those situations is usually predicates, so it is possible to formalize them as frame structures containing different frame elements. We followed such an approach for the mapping of the Spanish corpus of consumers’ questions to the classes of the domain ontology (Fernández-Barrera and Casanovas, 2011). And the same technique was applied for mapping a set of citizens’ complaints in the domain of acoustic nuisances to a legal domain ontology (Bourcier and Fernández-Barrera, 2011). By describing general structures of citizen description of legal cases we ensure scalability.

Emotional structures are extracted by current algorithms for opinion- and sentiment mining. User data in the legal domain often contain an important number of subjective elements (especially in the case of complaints and feedback on public services) that could be effectively mined and used in public decision making.

Finally, event structures, which have been deeply explored so far, could be useful for information extraction from user complaints and feedback, or for automatic classification into specific types of queries according to the described situation.

Crowdsourcing in e-government: next steps (and precautions?)

Legal prosumers’ input currently outstrips the capacity of government for extracting meaningful content in a cost-efficient way. Some developments are under way, among which are argument-mapping technologies and semantic matching between legal and lay corpora. The scalability of these methodologies is the main obstacle to overcome, in order to enable the matching of user data with open public data in several domains.

However, as technologies for the extraction of meaningful content from user-generated data develop and are used in public-decision making, a series of issues will have to be dealt with. For instance, should the system developer bear responsibility for the erroneous or biased analysis of data? Ethical questions arise as well: May governments legitimately analyse any type of user-generated content? Content-analysis systems might be used for trend- and crisis detection; but what if they are also used for restricting freedoms?

The “wisdom of crowds” can certainly be valuable in public decision making, but the fact that citizens’ online behaviour can be observed and analysed by governments without citizens’ acknowledgement poses serious ethical issues.

Thus, technical development in this domain will have to be coupled with the definition of ethical guidelines and standards, maybe in the form of a system of quality labels for content-analysis systems.

[Editor’s Note: For earlier VoxPopuLII commentary on the creation of legal ontologies, see Núria Casellas, Semantic Enhancement of Legal Information… Are We Up for the Challenge? For earlier VoxPopuLII commentary on Natural Language Processing and legal Semantic Web technology, see Adam Wyner, Weaving the Legal Semantic Web with Natural Language Processing. For earlier VoxPopuLII posts on user-generated content, crowdsourcing, and legal information, see Matt Baca and Olin Parker, Collaborative, Open Democracy with LexPop; Olivier Charbonneau, Collaboration and Open Access to Law; Nick Holmes, Accessible Law; and Staffan Malmgren, Crowdsourcing Legal Commentary.]

[1] The idea of prosumption existed actually long before the Internet, as highlighted by Ritzer and Jurgenson (2010): the consumer of a fast food restaurant is to some extent as well the producer of the meal since he is expected to be his own waiter, and so is the driver who pumps his own gasoline at the filling station.

[2] The experience project enables registered users to share life experiences, and it contained around 7 million stories as of January 2011:

[3] For instance, the United Nations Volunteers Online platform ( helps volunteers to cooperate virtually with non-governmental organizations and other volunteers around the world.

[4] See for instance the experiment run by mathematician Gowers on his blog: he posted a problem and asked a large number of mathematicians to work collaboratively to solve it. They eventually succeeded faster than if they had worked in isolation:

[5] The Galaxy Zoo project asks volunteers to classify images of galaxies according to their shapes: See as well Cornell’s projects Nestwatch ( and FeederWatch (, which invite people to introduce their observation data into a Website platform.


[7] See the description of Debategraph in Marta Poblet’s post, Argument mapping: visualizing large-scale deliberations (

[8] Terms have been organised in the form of a tree having as root nodes nine semantic classes previously identified. Terms have been added as branches and sub-branches, depending on their degree of abstraction.

[9] It should be noted that legal terms are mostly situated at the second level of the hierarchy rather than the first one. This is natural if we take into account the nature of the normative corpus (the Italian consumer code), which contains mostly domain specific concepts (for instance, withdrawal right) instead of general legal abstract categories (such as right and obligation).


Bourcier, D., and Fernández-Barrera, M. (2011). A frame-based representation of citizen’s queries for the Web 2.0. A case study on noise nuisances. E-challenges conference, Florence 2011.

Fernández-Barrera, M., and Casanovas, P. (2011). From user needs to expert knowledge: Mapping laymen queries with ontologies in the domain of consumer mediation. AICOL Workshop, Frankfurt 2011.

Lux, M., and Dsinger, G. (2007). From folksonomies to ontologies: Employing wisdom of the crowds to serve learning purposes. International Journal of Knowledge and Learning (IJKL), 3(4/5): 515-528.

Mika, P. (2005). Ontologies are us: A unified model of social networks and semantics. In Proc. of Int. Semantic Web Conf., volume 3729 of LNCS, pp. 522-536. Springer.

Passant, A. (2007). Using ontologies to strengthen folksonomies and enrich information retrieval in Weblogs. In Int. Conf. on Weblogs and Social Media, 2007.

Poblet, M., Casellas, N., Torralba, S., and Casanovas, P. (2009). Modeling expert knowledge in the mediation domain: A Mediation Core Ontology, in N. Casellas et al. (Eds.), LOAIT- 2009. 3rd Workshop on Legal Ontologies and Artificial Intelligence Techniques joint with 2nd Workshop on Semantic Processing of Legal Texts. Barcelona, IDT Series n. 2.

Ritzer, G., and Jurgenson, N. (2010). Production, consumption, prosumption: The nature of capitalism in the age of the digital “prosumer.” In Journal of Consumer Culture 10: 13-36.

Specia, L., and Motta, E. (2007). Integrating folksonomies with the Semantic Web. Proc. Euro. Semantic Web Conf., 2007.

Meritxell Fernández-Barrera is a researcher at the Cersa (Centre d’Études et de Recherches de Sciences Administratives et Politiques) -CNRS, Université Paris 2-. She works on the application of natural language processing (NLP) to legal discourse and legal communication, and on the potentialities of Web 2.0 for participatory democracy.

VoxPopuLII is edited by Judith Pratt. Editor-in-Chief is Robert Richards, to whom queries should be directed. The statements above are not legal advice or legal representation. If you require legal advice, consult a lawyer. Find a lawyer in the Cornell LII Lawyer Directory.


The progressive deployment of information and communication technologies (ICT) in the courtroom (audio and video recording, document scanning, courtroom management systems), jointly with the requirement for paperless judicial folders pushed by e-justice plans (Council of the European Union, 2009), are quickly transforming the traditional judicial folder into an integrated multimedia folder, where documents, audio recordings and video recordings can be accessed, usually via a Web-based platform. This trend is leading to a continuous increase in the number and the volume of case-related digital judicial libraries, where the full content of each single hearing is available for online consultation. A typical trial folder contains: audio hearing recordings, audio/video hearing recordings, transcriptions of hearing recordings, hearing reports, and attached documents (scanned text documents, photos, evidences, etc.). The ICT container is typically a dedicated judicial content management system (court management system), usually physically separated and independent from the case management system used in the investigative phase, but interacting with it.

Most of the present ICT deployment has been focused on the deployment of case management systems and ICT equipment in the courtrooms, with content management systems at different organisational levels (court or district). ICT deployment in the judiciary has reached different levels in the various EU countries, but the trend toward full e-justice is clearly in progress. Accessibility of the judicial information, both of case registries (more widely deployed), and of case e-folders, has been strongly enhanced by state-of-the-art ICT technologies. Usability of the electronic judicial folders is still affected by a traditional support toolset, such that an information search is limited to text search, transcription of audio recordings (indispensable for text search) is still a slow and fully manual process, template filling is a manual activity, etc. Part of the information available in the trial folder is not yet directly usable, but requires a time-consuming manual search. Information embedded in audio and video recordings, describing not only what was said in the courtroom, but also the specific trial context and the way in which it was said, still needs to be exploited. While the information is there, information extraction and semantically empowered judicial information retrieval still wait for proper exploitation tools. The growing amount of digital judicial information calls for the development of novel knowledge management techniques and their integration into case and court management systems. In this challenging context a novel case and court management system has been recently proposed.

The JUMAS project (JUdicial MAnagement by digital libraries Semantics) was started in February 2008, with the support of the Polish and Italian Ministries of Justice. JUMAS seeks to realize better usability of multimedia judicial folders — including transcriptions, information extraction, and semantic search –to provide to users a powerful toolset able to fully address the knowledge embedded in the multimedia judicial folder.

The JUMAS project has several objectives:

  • (1) direct searching of audio and video sources without a verbatim transcription of the proceedings;
  • (2) exploitation of the hidden semantics in audiovisual digital libraries in order to facilitate search and retrieval, intelligent processing, and effective presentation of multimedia information;
  • (3) fusing information from multimodal sources in order to improve accuracy during the automatic transcription and the annotation phases;
  • (4) optimizing the document workflow to allow the analysis of (un)structured information for document search and evidence-based assessment; and
  • (5) supporting a large scale, scalable, and interoperable audio/video retrieval system.

JUMAS is currently under validation in the Court of Wroclaw (Poland) and in the Court of Naples (Italy).


In order to explain the relevance of the JUMAS objectives, we report some volume data related to the judicial domain context. Consider, for instance, the Italian context, where there are 167 courts, grouped in 29 districts, with about 1400 courtrooms. In a law court of medium size (10 courtrooms), during a single legal year, about 150 hearings per court are held, with an average duration of 4 hours. Considering that in approximately 40% of them only audio is recorded, in 20% both audio and video, while the remaining 40% has no recording, the multimedia recording volume we are talking about is 2400 hours of audio and 1200 hours of audio/video per year. The dimensioning related to the audio and audio/video documentation starts from the hypothesis that multimedia sources must be acquired at high quality in order to obtain good results in audio transcription and video annotation, which will affect the performance connected to the retrieval functionalities. Following these requirements, one can figure out a storage space of about 8.7 megabytes per minute (MB/min) for audio and 39 MB/min for audio/video. This means that during a legal year for a court of medium size we need to allocate 4 terabytes (TB) for audio/video material. Under these hypotheses, the overall size generated by all the courts in the justice system — for Italy only — in one year is about 800 TB. This shows how the justice sector is a major contributor to the data deluge (The Economist, 2010).

In order to manage such quantities of complex data, JUMAS aims to:

  • Optimize the workflow of information through search, consultation, and archiving procedures;
  • Introduce a higher degree of knowledge through the aggregation of different heterogeneous sources;
  • Speed up and improve decision processes by enabling discovery and exploitation of knowledge embedded in multimedia documents, in order to consequently reduce unnecessary costs;
  • Model audio-video proceedings in order to compare different instances; and
  • Allow traceability of proceedings during their evolution.


To achieve the above-mentioned goals, the JUMAS project has delivered the JUMAS system, whose main functionalities (depicted in Figure 1) are: automatic speech transcription, emotion recognition, human behaviour annotation, scene analysis, multimedia summarization, template-filling, and deception recognition.


Figure 1: Overview of the JUMAS functionalities

The architecture of JUMAS, depicted in Figure 2, is based on a set of key components: a central database, a user interface on a Web portal, a set of media analysis modules, and an orchestration module that allows the coordination of all system functionalities.

Figure 2: Overview of the JUMAS architecture

The media stream recorded in the courtroom includes both audio and video that are analyzed to extract semantic information used to populate the multimedia object database. The outputs of these processes are annotations: i.e., tags attached to media streams and stored in the database (Oracle 11g). The integration among modules is performed through a workflow engine and a module called JEX (JUMAS EXchange library). While the workflow engine is a service application that manages all the modules for audio and video analysis, JEX provides a set of services to upload and retrieve annotations to and from the JUMAS database.



Automatic Speech Transcription. For courtroom users, the primary sources of information are audio-recordings of hearings/proceedings. In light of this, JUMAS provides an Automatic Speech Recognition (ASR) system (Falavigna et al., 2009 and Rybach et al., 2009) trained on real judicial data coming from courtrooms. Currently two ASR systems have been developed: the first provided by Fondazione Bruno Kessler for the Italian language, and the second delivered by RWTH Aachen University for the Polish language. Currently, the ASR modules in the JUMAS system offer 61% accuracy over the generated automatic transcriptions, and represent the first contribution for populating the digital libraries with judicial trial information. In fact, the resulting transcriptions are the main information resource that are to be enriched by other modules, and then can be consulted by end users through the information retrieval system.

Emotion Recognition. Emotional states represent an aspect of knowledge embedded into courtroom media streams that may be used to enrich the content available in multimedia digital libraries. Enabling the end user to consult transcriptions by considering the associated semantics as well, represents an important achievement, one that allows the end user to retrieve an enriched written sentence instead of a “flat” one. Even if there is an open ethical discussion about the usability of this kind of information, this achievement radically changes the consultation process: sentences can assume different meanings according to the affective state of the speaker. To this purpose an emotion recognition module (Archetti et al., 2008), developed by the Consorzio Milano Ricerche jointly with the University of Milano-Bicocca, is part of the JUMAS system. A set of real-world human emotions obtained from courtroom audio recordings has been gathered for training the underlying supervised learning model.

Human Behavior Annotation. A further fundamental information resource is related to the video stream. In addition to emotional states identification, the recognition of relevant events that characterize judicial proceedings can be valuable for end users. Relevant events occurring during proceedings trigger meaningful gestures, which emphasize and anchor the words of witnesses, and highlight that a relevant concept has been explained. For this reason, the human behavior recognition modules (Briassouli et al., 2009, Kovacs et al., 2009), developed by CERTH-ITI and by MTA SZTAKI Research Institute, have been included in the JUMAS system. The video analysis captures relevant events that occur during the course of a trial in order to create semantic annotations that can be retrieved by judicial end users. The annotations are mainly concerned with the events related to the witness: change of posture, change of witness, hand gestures, gestures indicating conflict or disagreement.

Deception Detection. Discriminating between truthful and deceptive assertions is one of the most important activities performed by judges, lawyers, and prosecutors. In order to support these individuals’ reasoning activities, respecting corroborating/contradicting declarations (in the case of lawyers and prosecutors) and judging the accused (judges), a deception recognition module has been developed as a support tool. The deception detection module developed by the Heidelberg Institute for Theoretical Studies is based on the automatic classification of sentences performed by the ASR systems (Ganter and Strube, 2009). In particular, in order to train the deception detection module, a manual annotation of the output of the ASR module — with the help of the minutes of the transcribed sessions — has been performed. The knowledge extracted for training the classification module deals with lies, contradictory statements, quotations, and expressions of vagueness.

Information Extraction. The current amount of unstructured textual data available in the judicial domain, especially related to transcriptions of proceedings, highlights the necessity of automatically extracting structured data from unstructured material, to facilitate efficient consultation processes. In order to address the problem of structuring data coming from the automatic speech transcription system, Consorzio Milano Ricerche has defined an environment that combines regular expressions, probabilistic models, and background information available in each court database system. Thanks to this functionality, the judicial actors can view each individual hearing as a structured summary, where the main information extracted consists of the names of the judge, lawyers, defendant, victim, and witnesses; the names of the subjects cited during a deposition; the date cited during a deposition; and data about the verdict.


Information Retrieval. Currently, to retrieve audio/video materials acquired during a trial, the end user must manually consult all of the multimedia tracks. The identification of a particular position or segment of a multimedia stream, for purposes of looking at and/or listening to specific declarations, is possible either by remembering the time stamp when the events occurred, or by watching or hearing the whole recording. The amalgamation of automatic transcriptions, semantic annotations, and ontology representations allows us to build a flexible retrieval environment, based not only on simple textual queries, but also on broad and complex concepts. In order to define an integrated platform for cross-modal access to audio and video recordings and their automatic transcriptions, a retrieval module able to perform semantic multimedia indexing and retrieval has been developed by the Information Retrieval group at MTA SZTAKI. (Darczy et al., 2009)

Ontology as Support to Information Retrieval. An ontology is a formal representation of the knowledge that characterizes a given domain, through a set of concepts and a set of relationships that obtain among them. In the judicial domain, an ontology represents a key element that supports the retrieval process performed by end users. Text-based retrieval functionalities are not sufficient for finding and consulting transcriptions (and other documents) related to a given trial. A first contribution of the ontology component developed by the University of Milano-Bicocca (CSAI Research Center) for the JUMAS system provides query expansion functionality. Query expansion aims at extending the original query specified by end users with additional related terms. The whole set of keywords is then automatically submitted to the retrieval engine. The main objective is to narrow the search focus or to increase recall.

User Generated Semantic Annotations. Judicial users usually manually tag some documents for purposes of highlighting (and then remembering) significant portions of the proceedings. An important functionality, developed by the European Media Laboratory and offered by the JUMAS system, relates to the possibility of digitally annotating relevant arguments discussed during a proceeding. In this context, the user-generated annotations may aid judicial users in future retrieval and reasoning processes. The user-generated annotations module included in the JUMAS system allows end users to assign free tags to multimedia content in order to organize the trials according to their personal preferences. It also enables judges, prosecutors, lawyers, and court clerks to work collaboratively on a trial; e.g., a prosecutor who is taking over a trial can build on the notes of his or her predecessor.


Hyper Proceeding Views. The user interface of JUMAS — developed by ESA Projekt and Consorzio Milano Ricerche — is a Web portal, in which the contents of the database are presented in different views. The basic view allows browsing of the trial archive, as in a typical court management system, to view general information (dates of hearings, name of people involved) and documents attached to each trial. JUMAS’s distinguishing features include the automatic creation of a summary of the trial, the presentation of user-generated annotations, and the Hyper Proceeding View: i.e., an advanced presentation of media contents and annotations that allows the user to perform queries on contents, and jump directly to relevant parts of media files.


Multimedia Summarization. Digital videos represent a fundamental information resource about the events that occur during a trial: such videos can be stored, organized, and retrieved in a short time and at low cost. However, considering the dimensions that a video resource can assume during the recording of a trial, judicial actors have specified several requirements for digital trial videos: fast navigation of the stream, efficient access to data within the stream, and effective representation of relevant contents. One possible solution to these requirements lies in multimedia summarization, which derives a synthetic representation of audio/video contents with a minimal loss of meaningful information. In order to address the problem of defining a short and meaningful representation of a proceeding, a multimedia summarization environment based on an unsupervised learning approach has been developed (Fersini et al., 2010) by Consorzio Milano Ricerche jointly with University of Milano-Bicocca.


The JUMAS project demonstrates the feasibility of enriching a court management system with an advanced toolset for extracting and using the knowledge embedded in a multimedia judicial folder. Automatic transcription, template filling, and semantic enrichment help judicial actors not only to save time, but also to enhance the quality of their judicial decisions and performance. These improvements are mainly due to the ability to search not only text, but also events that occur in the courtroom. The initial results of the JUMAS project indicate that automatic transcription and audio/video annotations can provide additional information in an affordable way.

Elisabetta Fersini has a post-doctoral research fellow position at the University of Milano-Bicocca. She received her PhD with a thesis on “Probabilistic Classification and Clustering using Relational Models.” Her research interest is mainly focused on (Relational) Machine Learning in several domains, including Justice, Web, Multimedia, and Bioinformatics.

VoxPopuLII is edited by Judith Pratt.

Editor-in-Chief is Robert Richards, to whom queries should be directed.

In this post, I will describe how natural language processing can help in creating computer systems dealing with the law.

Law Books EmileA lot of computer systems are being designed to help users deal with legal texts — accessing, understanding, or applying them. [Editor’s Note: Michael Poulshock’s Jureeka is an example of a system that automates the application of legal texts.] Other systems — such as DALOS — are about creating legal texts, providing support for the writers, or simulating the effects of a text. Such systems are based on something more than “just” the legal text: there is XML mark-up, an OWL ontology, or a representation of the rules in SWRL or some programming language. This means that any piece of legislation that you want to use on your computer system needs to be translated into this computer representation.

We try to support this translation using natural language processing, so that (part of) the translation can be done by a computer. This automation should have a number of advantages. First of all, computers are cheaper than human experts, and automating the process should reduce the amount of resources needed for this task. Second, the models that are produced by automated processes are more consistent; human experts may treat two similar sentences differently, but a computer program will always behave the same. Finally, an approach employing structures ensures that there is a clear mapping between the elements of the computer model and the original text.

Natural Language Processing isn’t perfect yet: computers cannot understand human language. However, legal text is quite structured, and offers a lot more handholds for automated translation than, say, a novel.

Document Structure

The first step that we will have to undertake is to determine the structure of the document. Online services like and can make it easier to access legal documents because they can point you to the right part of the document (such as a chapter, paragraph, sentence, etc.). In most law texts, the structure has been made explicit using clear headings, like: Chapter 1 or Chapter 1. General Provisions. So, in order to detect structure, we need to detect these headings. This means we’ll need to search the document for lines starting with Chapter, followed by some designation (which we refer to as an index), and perhaps followed by some text – say, the title of the chapter. The index can Numbersbe a lot of things: Arabic numbers (1, 2, 3, …), Roman numbers (I, II, III, …) or letters (a, b, c, …). Sometimes the index is an ordinal appearing before the chapter label: First chapter. It may even be a combination of several numbers and letters (5.2a). This is not a great problem, as we can more or less assume that whatever follows the word Chapter is the index.

The main problem with this approach is that there are also regular sentences that start with the word Chapter, and we need to separate those out. To do so, we can use some heuristics: A title will not end with a full stop (.); a heading will always start on a new line; etc.

Queen BeatrixThis procedure to find the headings for chapters is repeated to find headings for sections, subsections, etc. Also, some sections (like numbered paragraphs or list items) will not have a full heading, but just a number, which we also need to recognise. Finally, some sections don’t have a heading but can be recognised because they start with a fixed language pattern. For example, a preamble in a (recent) Dutch Law — such as this — will start with: We, Beatrix, Queen of the Netherlands, Princess of Orange-Nassau, etc. etc. etc.

This procedure assumes that the input for the process is just text. Many documents will contain more information — such as textual markup — and headings may be more easily identified because they are marked as bold text, or even as headings. So, in situations where the input is made up of documents that are marked-up in a consistent way, it may be easier to recognise the patterns by taking layout into account in addition to text.

To actually find the patterns, we can use existing toolkits like GATE. After the patterns have been found, and the structure has been recognised, we can store it using a format such as MetaLex.


The second step is to detect the references from a portion of a law text to other portions of that text, or from a law text to other texts. References, like headings, follow a pattern. The simplest patterns are rather similar to headings; the text chapter 13 is probably a reference, unless it is part of a heading. ReferencesJust like headings, basic references consist of a label (section, chapter, article) and an index (13, 13.2.1, XIII,  m). And, just as with headings, we can find the references by looking for these patterns in the text.

However, this is only the simplest form of references. Besides references to a specific section, such as chapter 13, there are of course also references to a complete law. Some of these references follow a pattern as well, such as the law of October 1st, 2007. Most laws are cited by means of a citation title, though, such as the Railroad Act. Such titles can contain all kinds of words, and they don’t follow a strict pattern. Thus, such references cannot be detected using patterns. Instead, we use a list containing all (citation) titles to detect such references.

Other, more complex references contain multiple references in one statement, such as articles 13 and 14, or multiple levels: article 13, item e, of the Railroad Act, or even more complex combinations of the two: articles 13, item e, 14, item f, 15 and 16, items a and b, of the Railroad Act. Though more complex than the simple combination of label and index, these references still follow clear (sometimes recurring) patterns, and can be found in the text by searching for such patterns.

At the Leibniz Center for Law, we’ve created a parser based on these patterns, which had an accuracy of over 95%. For each reference found, we can construct some standardised name, and store it. With this technology, not only can we add hyperlinks to documents; we can also search for documents that refer to some specific document.


Now that we’ve got the structure and links in place, it’s time to start with the actual meaning of the text. Rather than tackling the entire text as a whole, we’ve selected sentences as the basic building blocks, and we attempt to create computer models for individual sentences first. Later, we can integrate those individual models to a complete model.

As a first step in creating the models, we start by assigning a broad meaning, or classification, to each sentence. Does the sentence give a definition for a concept, describe an obligation, or make a change in another law? In total, we distinguish fourteen different classes of sentences that appear in Dutch law texts. The next step in our automated approach is to assign a class to each sentence automatically.

To do so, we turn once again to language patterns. Legal language is rather strict, and legislative drafters don’t vary their language a lot — in a novel, variation may make for a more appealing text, but in a law, variation invites ambiguity. In fact, there are official Guidelines for Legislative Drafting that (among other things) reduce the variety of texts used. [Editor’s Note: For example, drafters of legislation in the U.S. House of Representatives Office of the Legislative Counsel have used Donald Hirsch’s Drafting Federal Law.] This means that for each of our classes, there’s a rather limited set of language patterns used. For example, definitions will look like one of these:

Under … is understood …

This law understands under … …

There are some variations in word order, but in the end, a small set of patterns is sufficient to describe all commonly used phrases. There is only one class of sentences where we cannot define a full set of patterns: obligations. In Dutch laws, obligations are often expressed without signal words like must or is obliged to. Instead, the obligations are presented as a fact:

No bodies are buried on a closed cemetery.

However, since the obligations are the only sentences lacking all-compassing patterns, we will assume that any sentence that does not mark a pattern is one of these obligations.

Based on the patterns found, we’ve created a classifier that attempts to sort sentences into these different classes. This classifier has an accuracy of 91%, and we expect that this can improved a bit further.

(As a side note: For classification tasks as these, a machine learning approach is often preferred; see, e.g., here. With such an approach, you provide the computer, not with patterns, but with a bunch of sample sentences. The computer will then extract its own patterns from those sentences, and use these to classify any new sentences. We’ve tried this approach as well (using the toolkit WEKA), and reached similarly accurate results.)


Having classified the sentences, we now want to create models of the sentences. In essence, this means breaking down each sentence into smaller components and defining relationships between them. In some cases, the patterns used to classify the sentence already give us sufficient information to break up the sentence. Suppose we have a sentence like:

In article 7.12, sub one, second sentence, «article 7.3b» is replaced by: article 7.3c.

We classify this sentence as a replacement because of the text is replaced by. We can then also conclude that the text between angle quotes is the text to be replaced, the text following the colon is the replacing text, and the reference preceding it (which we’ve already detected) is the location where the replacement should take place.

This works fine for sentences that are somehow “about” the law. But for sentences that deal with some other domain, such as taxes, traffic, or commerce, we cannot predict all the elements. These sentences could be about anything — and statutes are full of such sentences. For such sentences, we need to follow a generic method. The aim is to model rules as a situation or action that is allowed or not allowed, similar to the models created in the HARNESS system of the ESTRELLA project. For example, for an obligation, we assume that the sentence describes some action that must be done. warrantWe try to identify who should be doing the action, and what other elements are involved. Thus, for the sentence:

Our Minister issues a warrant to the negligent person.

we would like to extract the following information:

Action: Issue
Agent: Our Minister
Patient: Warrant
Recipient: Negligent person

(Such a table, or frame, is not the same as a computer model, but has all the elements needed to create one.)

Now, identifying these different elements of the sentence (agent, patient, recipient) is something that computer linguists have already worked on for a long time, which means we do not have to start from scratch. Instead, we can use existing parsers to do much of the work for us. For our Dutch laws, we use the Alpino parser. voxemiletree.jpgSuch a parser will create a parse tree of a sentence. In this parse tree, the sentence will be split up in parts. The parser can identify which part is the subject, the direct object, the indirect object, etc. Based on this information, we can determine the agent, patient, and recipient (so-called semantic roles). In a sentence with a verb in the active voice, the subject is the agent, the direct object is the patient, and the indirect object is the recipient. Furthermore, the parser will determine the relationship between words, such as an adjective that modifies a noun. This information, too, helps us to make more accurate models.

We start out with the output of these parsers, and then try to extract all terms that have some more significance. If we want an application to compute whether or not a situation is allowed, a word like car can be treated in a generic way, but terms like allowed and not some special attention.

To Be Continued…

We still need to refine the method for making these models, and evaluate the results. After that, the individual models will need to be merged. But even as things stand now, we think these tools will help with getting legal text from paper into your computer systems.

[Editor’s Note: For more information about this topic, please see Dr. Adam Wyner’s post, Weaving the Legal Semantic Web with Natural Language Processing.]

Emile_de_MaatEmile de Maat is a researcher at the Leibniz Center for Law (University of Amsterdam). His research focuses on the automatic extraction of metadata and meaning from legal sources.

VoxPopuLII is edited by Judith Pratt. Editor in chief is Robert Richards.

CornucopiaThe World Wide Web is a virtual cornucopia of legal information bearing on all manner of topics and in a spectrum of formats, much of it textual. However, to make use of this storehouse of textual information, it must be annotated and structured in such a way as to be meaningful to people and processable by computers. One of the visions of the Semantic Web has been to enrich information on the Web with annotation and structure. Yet, given that text is in a natural language (e.g., English, German, Japanese, etc.), which people can understand but machines cannot, some automated processing of the text itself is needed before further processing can be applied. In this article, we discuss one approach to legal information on the World Wide Web, the Semantic Web, and Natural Language Processing (NLP). Each of these are large, complex, and heterogeneous topics of research; in this short post, we can only hope to touch on a fragment and that heavily biased to our interests and knowledge. Other important approaches are mentioned at the end of the post. We give small working examples of legal textual input, the Semantic Web output, and how NLP can be used to process the input into the output.

Legal Information on the Web

For clients, legal professionals, and public administrators, the Web provides an unprecedented opportunity to search for, find, and reason with legal information such as case law, legislation, legal opinions, journal articles, and material relevant to discovery in a court procedure. With a search tool such as Google or indexed searches made available by Lexis-Nexis, Westlaw, or the World Legal Information Institute, the legal researcher can input key words into a search and get in return a (usually long) list of documents which contain, or are indexed by, those key words.

As useful as such searches are, they are also highly limited to the particular words or indexation provided, for the legal researcher must still manually examine the documents to find the substantive information. Moreover, current legal search mechanisms do not support more meaningful searches such as for properties or relationships, where, for example, a legal researcher searches for cases in which a company has the property of being in the role of plaintiff or where a lawyer is in the relationship of representing a client. Nor, by the same token, can searches be made with respect to more general (or more specific) concepts, such as “all cases in which a company has any role,” some particular fact pattern, legislation bearing on related topics, or decisions on topics related to a legal subject.

Binary MysteryThe underlying problem is that legal textual information is expressed in natural language. What literate people read as meaningful words and sentences appear to a computer as just strings of ones and zeros. Only by imposing some structure on the binary code is it converted to textual characters as we know them. Yet, there is no similar widespread system for converting the characters into higher levels of structure which correlate to our understanding of meaning. While a search can be made for the string plaintiff, there are no (widely available) searches for a string that represents an individual who bears the role of plaintiff. To make language on the Web more meaningful and structured, additional content must be added to the source material, which is where the Semantic Web and Natural Language Processing come into play.

Semantic Web

The Semantic Web is a complex of design principles and technologies which are intended to make information on the Web more meaningful and usable to people.Semantic Web Stack We focus on only a small portion of this structure, namely the syntactic XML (eXtensible Markup Language) level, where elements are annotated so as to indicate linguistically relevant information and structure. (Click here for more on these points.) While the XML level may be construed as a ‘lower’ level in the Semantic Web “stack” — i.e., the layers of interrelated technologies that make up the Semantic Web — the XML level is nonetheless crucial to providing information to higher levels where ontologies (and click here for more on this) and logic play a role. So as to be clear about the relation between the Semantic Web and NLP, we briefly review aspects of XML by example, and furnish motivations as we go.

Suppose one looks up a case where Harris Hill is the plaintiff and Jane Smith is the attorney for Harris Hill. In a document related to this case, we would see text such as the following portions:

Harris Hill, plaintiff.
Jane Smith, attorney for the plaintiff.

While it is relatively straightforward to structure the binary string into characters, adding further information is more difficult. Consider what we know about this small fragment: Harris and Jane are (very likely) first names, Hill and Smith are last names, Harris Hill and Jane Smith are full names of people, plaintiff and attorney are roles in a legal case, Harris Hill has the role of plaintiff, attorney for is a relationship between two entities, and Jane Smith is in the attorney for relationship to Harris Hill. It would be useful to encode this information into a standardised machine-readable and processable form.

XML helps to encode the information by specifying requirements for tags that can be used to annotate the text. It is a highly expressive language, allowing one to define tags that suit one’s purposes so long as the specification requirements are met. One requirement is that each tag has a beginning and an ending; the material in between is the data that is being tagged. For example, suppose tags such as the following, where … indicates the data:


Another requirement is that the tags have a tree structure, where each pair of tags in the document is included in another pair of tags and there is no crossing over:


is acceptable, but

</firstname> ...</lastname></fullname>

is unacceptable. Finally, XML tags can be organised into schemas to structure the tags.

With these points in mind, we could represent our fragment as:


We have added structured information — the tags — to the original text. While this is more difficult for us to read, it is very easy for a machine to read and process. In addition, the tagged text contains the content of the information, which can be presented in a range of alternative ways and formats using a transformation language such as XSLT (click here for more on this point) so that we have an easier-to-read format.

Why bother to include all this additional information in a legal text? Because these additions allow us to query the source text and submit the information to further processing such as inference. Given a query language, we could submit to the machine the query Who is the attorney in the case? and the answer would be Jane Smith. Given a rule language — such as RuleML or Semantic Web Rule Language (SWRL) — which has a rule such as If someone is an attorney for a client then that client has a privileged relationship with the attorney, it might follow from this rule that the attorney could not divulge the client’s secrets. Applying such a rule to our sample, we could infer that Jane Smith cannot divulge Harris Hill’s secrets.

Tower of BabelThough it may seem here like too much technology for such a small and obvious task, it is essential where we scale up our queries and inferences on large corpora of legal texts — hundreds of thousands if not millions of documents — which comprise vast storehouses of unstructured, yet meaningful data. Were all legal cases uniformly annotated, we could, in principle, find out every attorney for every plaintiff for every legal case. Where our tagging structure is very rich, our queries and inferences could also be very rich and detailed. Perhaps a more familiar way to view documents annotated with XML is as a database to which further processes can be applied over the Web.

Natural Language Processing

As we have presented it, we have an input, the corpus of texts, and an output, texts annotated with XML tags. The objective is to support a range of processes such as querying and inference. However, getting from a corpus of textual information to annotated output is a demanding task, generically referred to as the knowledge acquisition bottleneck. Not only is the task demanding on resources (time, money, manpower); it is also highly knowledge intensive since whoever is doing the annotation must know what to look for, and it is important that all of the annotators annotate the text in the same way (inter-annotator agreement) to support the processes. Thus, automation is central.

Yet processing language to support such richly annotated documents confronts a spectrum of difficult issues. Among them, natural language supports (1) implicit or presupposed information, (2) multiple forms with the same meaning, (3) the same form with different contextually dependent meanings, and (4) dispersed meanings. (Similar points can be made for sentences or other linguistic elements.) Here are examples of these four issues:

(1) “When did you stop taking drugs?” (presupposes that the person being questioned took drugs at sometime in the past);
(2) Jane Smith, Jane R. Smith, Smith, Attorney Smith… (different ways to refer to the same person);
(3) The individual referred to by the name “Jane Smith” in one case decision may not be the individual referred to by the name “Jane Smith” in another case decision;
(4) Jane Smith represented Jones Inc. She works for Dewey, Cheetum, and Howe. To contact her, write to .

When we search for information, a range of linguistic structures or relationships may be relevant to our query, such as:

People grasp relationships between words and phrases, such that Bill exercises daily contrasts with the meaning of Bill is a couch potato, or that if it is true that Bill used a knife to kill Phil, then Bill killed Phil. Finally, meaning tends to be sparse; that is, there are a few words and patterns that occur very regularly, while most words or patterns occur relatively rarely in the corpus.

Natural language processing (NLP) takes on this highly complex and daunting problem as an engineering problem, decomposing large problems into smaller problems and subdomains until it gets to those which it can begin to address. Having found a solution to smaller problems, NLP can then address other problems or larger scope problems. Some of the subtopics in NLP are:

  • Generation – converting information in a database into natural language.
  • Understanding – converting natural language into a machine-readable form.
  • Information Retrieval – gathering documents which contain key words or phrases. This is essentially what is done by Google.
  • Text Summarization – summarizing (in a paragraph) the main meaning of a text or corpus.
  • Question Answering – making queries and giving answers to them, in natural language, with respect to some corpus of texts.
  • Information Extraction — identifying, annotating, and extracting information from documents for reuse, representation, or reasoning.

In this article, we are primarily (here) interested in information extraction.

NLP Approaches: Knowledge Light v. Knowledge Heavy

There are a range of techniques that one can apply to analyse the linguistic data obtained from legal texts; each of these techniques has strengths and weaknesses with respect to different problems. Statistical and machine-learning techniques are considered “knowledge light.” With statistical approaches, the processing presumes very little knowledge by the system (or analyst). Rather, algorithms are applied that compare and contrast large bodies of textual data, and identify regularities and similarities. Such algorithms encounter problems with sparse data or patterns that are widely dispersed across the text. (See Turney and Pantel (2010) for an overview of this area.) Machine learning approaches apply learning algorithms to annotated material to extend results to unannotated material, thus introducing more knowledge into the processing pipeline. However, the results are somewhat of a black box in that we cannot really know the rules that are learned and use them further.

With a “knowledge-heavy” approach, we know, in a sense, what we are looking for, and make this knowledge explicit in lists and rules for processing. Yet, this is labour- and knowledge-intensive. In the legal domain it is crucial to have humanly understandable explanations and justifications for the analysis of a text, which to our thinking warrants a knowledge-heavy approach.

One open source text-mining package, General Architecture for Text Engineering (GATE), consists of multiple components in a cascade or pipeline, each component automatically processing some aspect of the text, and then feeding into the next process. The underlying strategy in all the components is to find a pattern (from either a list or a previous process) which matches a rule, and then to apply the rule which annotates the text. Each component performs a particular process on the text, such as:

  • Sentence segmentation – dividing text into sentences.
  • Tokenisation – words identified by spaces between them.
  • Part-of-speech tagging – noun, verb, adjective, etc., determined by look-up and relationships among words.
  • Shallow syntactic parsing/chunking – dividing the text by noun phrase, verb phrase, subordinate clause, etc.
  • Named entity recognition – the entities in the text such as organisations, people, and places.
  • Dependency analysis – subordinate clauses, pronominal anaphora [i.e., identifying what a pronoun refers to], etc.

The system can also be used to annotate more specifically to elements of interest. In one study, we annotated legal cases from a case base (a corpus of cases) in order to identify a range of particular pieces of information that would be relevant to legal professionals such as:

  • Case citation.
  • Names of parties.
  • Roles of parties, meaning plaintiff or defendant.
  • Type of court.
  • Names of judges.
  • Names of attorneys.
  • Roles of attorneys, meaning the side they represent.
  • Final decision.
  • Cases cited.
  • Nature of the case, meaning using keywords to classify the case in terms of subject (e.g., criminal assault, intellectual property, etc.)

Applying our lists and rules to a corpus of legal cases, a sample output is as follows, where the coloured highlights are annotated as per the key on the right; the colours are a visualisation of the sorts of tags discussed above (to see a larger version of the image, right click on the image, then click on “View Image” or a similar phrase; when finished viewing the image, use the browser’s back button to return to the text):

Annotation of a Criminal Case

The approach is very flexible and appears in similar systems. (See, for example, de Maat and Winkels, Automatic Classification of Sentences in Dutch Laws (2008).) While it is labour intensive to develop and maintain such list and rule systems, with a collaborative, Web-based approach, it may be feasible to construct rich systems to annotate large domains.


In this post, we have given a very brief overview of how the Semantic Web and Natural Language Processing (NLP) apply to legal textual information to support annotation which then enables querying and inference. Of course, this is but one take on a much larger domain. In our view, it holds great promise in making legal information more transparent and available to more legal professionals. Aside from GATE, some other resources on text analytics and NLP are textbooks and lecture notes (see, e.g., Wilcock), as well as workshops (such as SPLeT and LOAIT). While applications of Natural Language Processing to legal materials are largely lab studies, the use of NLP in conjunction with Semantic Web technology to annotate legal texts is a fast-developing, results-oriented area which targets meaningful applications for legal professionals. It is well worth watching.

Adam WynerDr. Adam Zachary Wyner is a Research Fellow at the University of Leeds, Institute of Communication Studies, Centre for Digital Citizenship. He currently works on the EU-funded project IMPACT: Integrated Method for Policy Making Using Argument Modelling and Computer Assisted Text Analysis. Dr. Wyner has a Ph.D. in Linguistics (Cornell, 1994) and a Ph.D. in Computer Science (King’s College London, 2008). His computer science Ph.D. dissertation is entitled Violations and Fulfillments in the Formal Representation of Contracts. He has published in the syntax and semantics of adverbs, deontic logic, legal ontologies, and argumentation theory with special reference to law. He is workshop co-chair of SPLeT 2010: Workshop on Semantic Processing of Legal Texts, to be held 23 May 2010 in Malta. He writes about his research at his blog, Language, Logic, Law, Software.

VoxPopuLII is edited by Judith Pratt. Editor in chief is Robert Richards.


To take the words of Walt Whitman, when it comes to improving legal information retrieval (IR), lawyers, legal librarians and informaticians are all to some extent, “Both in and out of the game, and watching and wondering at it“. The reason is that each group holds only a piece of the solution to the puzzle, and as pointed out in an earlier post, they’re talking past each other.

In addition, there appears to be a conservative contingent in each group who actively hinder the kind of cooperation that could generate real progress: lawyers do not always take up technical advances when they are made available, thus discouraging further research, legal librarians cling to indexing when all modern search technologies use free-text search, and informaticians are frequently impatient with, and misunderstand, the needs of legal professionals.

What’s holding progress back?

At root, what has held legal IR back may be the lack of cross-training of professionals in law and informatics, although I’m impressed with the open-mindedness I observe at law and artificial intelligence conferences, and there are some who are breaking out of their comfort zone and neat disciplinary boundaries to address issues in legal informatics.

I recently came back from a visit to the National Institute of Informatics in Japan where I met Ken Satoh, a logician who, late in his professional career, has just graduated from law school. This is not just hard work. I believe it takes a great deal of character for a seasoned academic to maintain students’ respect when they are his seniors in a secondary degree. But the end result is worth it: a lab with an exemplary balance of lawyers and computer scientists, experience and enthusiasm, pulling side-by-side.

Still, I occasionally get the feeling we’re all hoping for some sort of miracle to deliver us from the current predicament posed by the explosion of legal information. Legal professionals hope to be saved by technical wizardry, and informaticians like myself are waiting for data provision, methodical legal feedback on system requirements and performance, and in some cases research funding. In other words, we all want someone other than ourselves to get the ball rolling.

Miracle Occurs

The need to evaluate

Take for example, the lack of large corpora for study, which is one of the biggest stumbling blocks in informatics. Both IR and natural language processing (NLP) currently thrive on experimentation with vast amounts of data, which is used in statistical processing. More data means better statistical estimates and the fewer `guesses’ at relevant probabilities. Even commercial legal case retrieval systems, which give the appearance of being Boolean, use statistics and have done so for around 15 years. (They are based on inference networks that simulate Boolean retrieval with weighted indexing by reducing the rigidness associated with conditional probability estimates for Boolean operators `and’, `or’ and `not’. In this way, document ranking increasingly depends on the number of query constraints met).

The problem is that to evaluate new techniques in IR (and thus improve the field), you need not only a corpus of documents to search but also a sample of legal queries and a list of all the relevant documents in response to those queries that exist in your corpus, perhaps even with some indication of how relevant they are. This is not easy to come by. In theory a lot of case data is publicly available, but accumulating and cleaning legal text downloaded from the internet, making it amenable to search, is nothing short of tortuous. Then relevance judgments must be given by legal professionals, which is difficult given that we are talking about a community of people who charge their time by the hour.

Of course, the cooperation of commercial search providers, who own both the data and training sets with relevance judgments, would make everyone’s life much easier, but for obvious commercial reasons they keep their data to themselves.

To see the productive effects of a good data set we need only look at the research boom now occurring in e-discovery (discovery of electronic evidence, or DESI). In 2006 the TREC Legal Track, including a large evaluation corpus, was established in response to the number of trials requiring e-discovery: 75% of Fortune 500 company trials, with more than 90% of company information now stored electronically. This has generated so much interest that an annual DESI workshop has been established since 2007.

Qualitative evaluation of IR performance by legal professionals is an alternative to the quantitative evaluation usually applied in informatics. The development of new ways to visualize and browse results seems particularly well suited to this approach, where we want to know whether users perceive new interfaces to be genuine improvements. Considering the history of legal IR, qualitative evaluation may be as important as traditional IR evaluation metrics of precision and recall. (Precision is the number of relevant items retrieved out of the total number of items retrieved, and recall is the number of relevant items retrieved out of the total number of relevant items in a collection). However, it should not be the sole basis for evaluation.

A well-known study by Blair and Maron makes this point plain. The authors showed that expert legal researchers retrieve less than 20% of relevant documents when they believe they have found over 75%. In other words, even experts can be very poor judges of retrieval performance.

Context in legal retrieval


Setting this aside, where do we go from here? Dan Dabney has argued at the American Association of Law Libraries (AALL) 2005 Annual Meeting that free text search decontextualizes information, and he is right. One thing to notice about current methods in open domain IR, including vector space models, probabilistic models and language models, is that the only context they are taking into account is proximate terms (phrases). At heart, they treat all terms as independent.

However, it’s risky to conclude what was reported from the same meeting: “Using indexes improves accuracy, eliminates false positive results, and leads to completion in ways that full-text searching simply cannot.” I would be interested to know if this continues to be a general perception amongst legal librarians despite a more recent emphasis on innovating with technologies that don’t encroach upon the sacred ground of indexing. Perhaps there’s a misconception that capitalizing on full-text search methods would necessarily replace the use of index terms. This isn’t the case; inference networks used in commercial legal IR are not applied in the open domain, and one of their advantages is that they can incorporate any number of types of evidence.

Specifically, index numbers, terms, phrases, citations, topics and any other desired information are treated as representation nodes in a directed acyclic graph (the network). This graph is used to estimate the probability of a user’s information need being met given a document.

For the time being lawyers, unaware of technology under the hood, default to using inference networks in a way that is familiar, via a search interface that easily incorporates index terms and looks like a Boolean search. (Inference nets are not Boolean but they can be made to behave in the same way.) While Boolean search does tend to be more precise than other methods, the more data there is to search the less well the system performs. Further, it’s not all about precision. Recall of relevant documents is also important and this can be a weak point for Boolean retrieval. Eliminating false positives is no accolade when true positives are eliminated at the same time.

Since the current predicament is an explosion of data, arguing for indexing by contrasting it with full-text retrieval without considering how they might work together seems counterproductive.

Perhaps instead we should be looking at revamping legal reliance on a Boolean-style interface so that we can make better use of full-text search. This will be difficult. Lawyers who are charged, and charge, per search, must be able to demonstrate the value of each search to clients; they can’t afford the iterative nature of what characterizes open domain browsing. Further, if the intelligence is inside the retrieval system, rather than held by legal researchers in the form of knowledge about how to put complex queries together, how are search costs justified? Although Boolean queries are no longer well-adapted, at least value is easy to demonstrate. A push towards free-text search by either legal professionals or commercial search providers will demand a rethink of billing structures.

Given our current systems, are there incremental ways we can improve results from full-text search? Query expansion is a natural consideration and incidentally overlaps with much of the technology underlying graphical means of data exploration such as word clouds and wonderwheels; the difference is that query expansion goes on behind the scenes, whereas in graphical methods the user is allowed to control the process. Query expansion helps the user find terms they hadn’t thought of, but this doesn’t help with the decontextualization problem identified by Dabney; it simply adds more independent terms or phrases.

In order to contextualize information we can marry search using text terms and index numbers as is already applied. Even better would be to do some linguistic analysis of a query to really narrow down the situations in which we want terms to appear. In this way we might get at questions such as “What happened in a case?” or “Why did it happen?” rather than just, “What is this document about?”.

Language processing and IR

Use of linguistic information in IR isn’t a novel idea. In the 1980s, IR researchers started to think about incorporating NLP as an intrinsic part of retrieval. Many of the early approaches attempted to use syntactic information for improving term indexing or weighting. For example, Fagan improved performance by applying syntactic rules to extract similar phrases from queries and documents and then using them for direct matching, but it was held that this was comparable to a less complex, and therefore preferable, statistical approach to language analysis. In fact, Fagan’s work demonstrated early on what is now generally accepted: statistical methods that do not assume any knowledge of word meaning or linguistic role are surprisingly (some would say depressingly) hard to beat for retrieval performance.

Since then there have been a number of attempts to incorporate NLP in IR, but depending on the task involved, there can be a lack of highly accurate methods for automatic linguistic analysis of text that are also robust enough to handle unexpected and complex language constructions. (There are exceptions, for example, part-of-speech tagging is highly accurate.) The result is that improved retrieval performance is often offset by negative effects, resulting in a minimal positive, or even a negative impact on overall performance. This makes NLP techniques not worth the price of additional computational overheads in time and data storage.

However, just because the task isn’t easy doesn’t mean we should give up. Researchers, including myself, are looking afresh at ways to incorporate NLP into IR. This is being encouraged by organizations such as the NII Test Collection for IR Systems Project (NTCIR), who from 2003 to 2006 amassed excellent test and training data for patent retrieval with corpora in Japanese and English and queries in five languages. Their focus has recently shifted towards NLP tasks associated with retrieval, such as patent data mining and classification. Further, their corpora enable study of cross-language retrieval issues that become important in e-discovery since only a minority fraction of a global corporation’s electronic information will be in English.

We stand on the threshold of what will be a period of rapid innovation in legal search driven by the integration of existing knowledge bases with emerging full-text processing technologies. Let’s explore the options.

Tamsin MaxwellK. Tamsin Maxwell is a PhD candidate in informatics at the University of Edinburgh, focusing on information retrieval in law. She has a MSc in cognitive science and natural language engineering, and has given guest talks at the University of Bologna, UMass Amherst, and NAIST in Japan. Her areas of interest include text processing, machine learning, information retrieval, data mining and information extraction. Her passion outside academia is Arabic dance.

VoxPopuLII is edited by Judith Pratt.