=Paper=
{{Paper
|id=Vol-2471/paper2
|storemode=property
|title=AI-Assisted Message Processing for the Netherlands National Police
|pdfUrl=https://ceur-ws.org/Vol-2471/paper2.pdf
|volume=Vol-2471
|authors=Bas Testerink,Daphne Odekerken,Floris Bex
|dblpUrl=https://dblp.org/rec/conf/icail/TesterinkOB19
}}
==AI-Assisted Message Processing for the Netherlands National Police==
https://ceur-ws.org/Vol-2471/paper2.pdf
AI-assisted message processing for
the Netherlands National Police
Bas Testerink Daphne Odekerken Floris Bex
Police Lab AI Police Lab AI Police Lab AI
Netherlands National Police Netherlands National Police Utrecht University
Driebergen, The Netherlands Driebergen, The Netherlands Utrecht, The Netherlands
bas.testerink@politie.nl daphne.odekerken@politie.nl f.j.bex@uu.nl
ABSTRACT relevance of the messages? Which entities (persons, bank accounts,
The number of messages that the Netherlands National Police (NNP) countries, organisations) are mentioned in the message? What is
receives (e.g. from international partner institutes and citizens) the intent of the sender – are they asking for information, or do they
grows steadily every year. The NNP has initiated a number of expect the Dutch police to take action? Not all of these questions can
projects to develop artificial intelligence systems that assist in the be answered given just the message. For instance, the occurrence
processing of such messages. In this demo, we show a prototype of of a person in the police databases may determine the relevancy
one such system that will be used for supporting the processing of for The Netherlands. The Triage Agent thus reads the e-mails and
messages from international (Interpol) partners. supports the coordinator. If the coordinator agrees with the agent,
they forward the messages with the relevant annotations (entities,
1 INTRODUCTION intent, priority level, etc.).
The second type of agent is a Specialist Agent, which supports
The number of messages that the Netherlands National Police (NNP)
specialists to do routine work on their respective themes. The agents
receives grows steadily every year. Such messages range from noti-
that work for the specialists will formulate the task that is required
fications from citizens to requests for assistance from international
given the message, execute it, and then report their findings to the
partner institutes. The NNP has initiated a project to develop artifi-
specialist as an enhancement of the initial message. The idea is that
cial intelligence (AI) that assists in the processing of such messages,
the specialist ultimately receives messages as if a colleague already
creating autonomous software agents that support human opera-
processed it. For instance, consider a notification that during a
tors. The use of natural language processing tools is a cornerstone
routine border patrol a Dutch vehicle was found to contain illegal
of the agents, because incoming messages are typically free-text
drugs. The Triage Agent already determined the vehicle to be indeed
(e-mails, online forms). Furthermore, it is important that the agents
Dutch and the message is forwarded to the specialist agent for drug
are designed in such a way that every major decision is made trans-
related crime. This specialist agent has access to current drug-
parently, and that legal and ethical rules and regulations can be
related investigations such as which organizations are of special
enforced.
interest. It tries to match the notification to existing investigations
The demo system enhances the existing processing of messages
or otherwise initiates a new one. By the time the specialist receives
that are received through the Interpol channel. An overview of
the notification from agent he or she can immediately see how the
the goal system incorporated into the Interpol message process-
notification relates to past information and what course of action
ing workflow is shown in Figure 1. The pink components with
would be prudent. The main task of the specialist then becomes the
human icons are the human operators. The orange components
monitoring and training of the agent.
with the computer chips represent agent components. The yellow
A final piece of functionality will be to aggregate the messages.
components are components without agency.
At the NNP we are working on real-time monitoring of intelligence
Currently, a coordinator monitors all the incoming messages
data from international partners, combining it with open source
and categorizes them on priority, theme and relevancy for The
intelligence (news, Wikipedia, Twitter) and in-house intelligence
Netherlands. The coordinator may answer the message directly,
from the NNP.
forward the message to a specialist for further processing, or choose
to ignore the message, for example because it is not relevant for
the Netherlands. Specialists specialize in topics such as counter- 2 AGENT ARCHITECTURE
terrorism and child sex tourism. Usually they have access to domain The architecture we use for the individual (Triage and Specialist)
data and contacts that are relevant for their expertise. A specialist agents is the same architecture that we have used in our other
can forward a message internally or answer it directly. project Intelligence Amplification for Cybercrime (IAC) [1], in which
The first agent that is inserted into the workflow is the Triage we have designed an agent to assist the NNP in the assessment of
Agent, which supports the coordinator by performing classification crime reports submitted by civilians. In a nutshell, the agent ap-
and information extraction tasks. What is the theme, priority and plies information extraction techniques to understand a document,
In: Proceedings of the Workshop on Artificial Intelligence and the Administrative State applies legal reasoning to determine whether more information
(AIAS 2019), June 17, 2019, Montreal, QC, Canada. is required and applies a policy that is optimized for efficiency to
© 2019 Copyright for this paper by its authors. Use permitted under Creative Commons determine the next action.
License Attribution 4.0 International (CC BY 4.0).
Published at http://ceur-ws.org The agent’s goal is to produce some (information) product such
as a report, reply or analysis. We refer to a coherent sequence of
AIAS, June 17, 2019, Montreal, QC, Canada Testerink, Odekerken and Bex
Outbox Inbox Triage Coordinator
Agent
Suggestions
Incoming messages
Feedback
Annotated messages
Outgoing messages
Specialist 1 Specialist Specialist k Specialist
Agent 1 Agent k
Feedback Feedback
Suggestions Suggestions
Domain Domain
Data 1 Data k
Updates Data Updates Data
Message summaries
International Aggregation
Monitor Service
Overviews
Figure 1: AI-assisted message processing multi-agent system overview.
interactions with the environment as a session. For example, a 2.1 Deployment
session can be a sequence of database queries that were required The agent connects to its environment through an external interface.
to respond to a message. A session always ends with a terminal That interface differs per application. In the message processing sys-
action. For instance, terminal actions can be to ignore the message, tem, it will contain functionalities such as forwarding e-mails and
forward it or answer it directly. querying databases. Typically, the external interface is implemented
For many law-enforcement applications we need to be able to as a layer that calls different APIs of other systems and passes on
check why the agent suggests (or directly executes) some terminal the callback. The aim of the agent is to create some (information)
action. The basis for many decisions is legislation. Hence, we draw product such as a message which is annotated with analyses and
upon the field of computational legal argumentation (cf. [2]) to suggested actions. These products are stored in the product database.
ensure that the agent has an argument grounded in the relevant Such a product is typically built in two phases: first the agent tries
rules and regulations when it decides upon a terminal action. The to find enough information to make a final decision on the product,
agent architecture is designed for creating agents that efficiently and second the final decision results in a terminal action.
seek information in their environment and transparently decide on External information, such as a message and database results,
which terminal action ought to be executed [4]. are feedback which the external interface sends to the internals of
Figure 2 shows an overview of the agent architecture. The de- the agent. The feedback is put through a pipeline of classifiers and
ployment phase concerns the actual functionality of the agent. attribute extractors which turn the feedback into structured data
The training phase is required to configure the deployment phase (statements about the feedback which are attributes and Boolean
components. The deployment components are the top-half (blue) observations). For our earlier IAC intake agent, we use existing
components. The training components are the bottom-half (green) named entity recognition software [3] and bespoke classifiers (cf.
components. The monitor interface and argumentation engine are Section 2.2) to classify and extract entities (e.g. the suspect, the
used in both phases. victim, addresses) and relations (e.g. “the suspect received money
from the victim”). For the Interpol Triage agent we use Spacy as the
basis and apply pre- and post-processing to improve upon its base
AI-assisted message processing for
the Netherlands National Police AIAS, June 17, 2019, Montreal, QC, Canada
Product
Database Product, Actuator
Queries,
Terminal actions
Queries,
Terminal actions
Product
External Classifiers & Completeness &
Interface Attribute Extractors Consistency Control Policy
Feedback Observations, State
Attributes
Product,
Queries, Observations
Terminal actions, Classifiers,
Policy configuration Argumentation
Feedback Extractors
Monitor Auto- Policy Argumentation
Interface Experimenter Learner Engine
Observations
Argumentation
Queries
Updates
Observations
Training Data & Labeled data Environment
Configuration Model
User data
Figure 2: Agent architecture.
performance. Our choice for Spacy was based on its ease-of-use and returns the result as feedback to the classification and attribute
and available multi-language models for NLP. extraction pipeline.
The classification and extraction pipeline consists of many sepa-
rately constructed components which may result in inconsistent
results. Hence, we apply a consistency control mechanism which 2.2 Training
makes sure that the data is consistent. It also checks whether the The monitor interface allows a human operator to monitor the
data is complete. That check is mainly for fulfilling the precondi- agent’s activities and control its training phase. The human op-
tions of final actions, e.g., administrative requirements. The result erator uses the monitor interface also to create labelled data by
of this controller is what we consider to be the state of current approving or disapproving (part of) the agent’s activities. The mon-
session, where a session is a sequence of actions after the initial itor interface also shows the argumentation behind core decisions.
input until the information product is produced (i.e. the suggested This connection is not shown in the figure as showing the argu-
course of action for the user). mentation does not directly impact the agent’s decision-making.
The actual decision making of the agent is executed by a decision- However, it does help the human operator to understand the choices
making policy. Based on the state, it determines the next action; this of the agent and localize where potential corrections have to be
can be an information gathering action (query) or a terminal action. made. The training of the agent is based on example data and config-
The policy may draw upon the argumentation engine in order to uration settings of its different training tasks. For the observations
argue for or against an action. The actuator of the agent prepares and attributes, we apply supervised learning which is enabled by
the action for execution through the agent’s external interface. For the gathering of labelled data during deployment. An automated
instance, the action might be an information gathering action upon experimenter module tries different algorithms in order to deter-
a database. The actuator may then formulate an SQL query which mine for each observation and attribute what the best classifier or
the external interface ensures is sent to the appropriate database extractor is.
AIAS, June 17, 2019, Montreal, QC, Canada Testerink, Odekerken and Bex
The policy of the agent is shaped by reinforcement learning 4 CONCLUSION
(although other methods can be used). The policy learner tries to In this paper we briefly discuss a multi-agent architecture for han-
create a policy which efficiently interacts with the environment. dling messages from international Interpol partners to the NNP, as
For instance, it may try to minimize the amount of data that it well as the architecture of a single agent. In our live demo, we show
queried from databases. In order to practice these interactions, the the workings of a single Triage Agent with a realistic example. We
policy learner requires an environment model that is generated encourage interested programmers to contact the authors for the
from the training data. The model captures for instance probability source code.
distributions over random variables that the agent encounters. At
the moment the model’s implementation is a Bayesian network REFERENCES
where its nodes are observations that the argumentation module [1] F.J. Bex, J. Peters, and B. Testerink. 2016. AI for online criminal complaints:
may use to construct arguments with. For reinforcement learning, From natural dialogues to structured scenarios. In Artificial Intelligence for Justice
Workshop (ECAI 2016). 22–29.
we apply the argumentation engine as part of its reward function. [2] H. Prakken and G. Sartor. 1996. A dialectical model of assessing conflicting
The agent gets a positive reward when it achieves a state such arguments in legal reasoning. In Logical models of legal argumentation. Springer,
that more feedback from the external interface cannot change its 175–211.
[3] M.P. Schraagen, M.J.S. Brinkhuis, and F.J. Bex. 2017. Evaluation of Named Entity
opinion on the terminal action. When such a state is reached, it Recognition in Dutch online criminal complaints. Computational Linguistics in
is natural to opt for the terminal action that the agent can argue the Netherlands Journal 7 (2017), 3–16.
for [4]. [4] Marijn Schraagen, Bas Testerink, Daphne Odekerken, and Floris Bex. 2019.
Argumentation-driven information extraction for online crime reports. In First
Workshop on Legal Data Analytics and Mining (LeDAM 2018).
3 DESIGN CONSIDERATIONS
The application of autonomous A.I. systems requires careful con-
siderations with respect to their potential impact. We decided to
restrict the agent’s capabilities to reading messages, querying sys-
tems and presenting information to the human operator since we
cannot guarantee correct behaviour due to the agent’s reliance on
imperfect information extraction. The agent has no capability for
updating databases or sending messages without explicit approval
from the human operator.
During deployment, every decision outcome of the agent is doc-
umented in its trace for auditability and can be inspected by a
human operator. However, it should be noted that it is not always
possible to completely reproduce the behaviour of the agent: some
queried databases contain data that is forbidden by law to store
in the same environment. Hence, it is for instance not allowed to
store raw database query results. As a result, there are situations
in which it cannot be reproduced which information the agent ex-
actly had when it made a decision. This happens for example when
source databases are updated. The human operator can provide
feedback through the monitor interface which can be taken into
consideration when the system is retrained/adjusted.
Interpretability has been an import design influence from the
start. It was determined early on that extracting information from
data will be a hard to interpret exercise under most circumstances.
From this point of view, it was not desirable to design the appli-
cation as an end-to-end system. Instead, it was opted to create a
method where the granularity of extraction can be balanced with
interpretability and transparency. In short, we designed the system
in such a way that we can choose how much information is ob-
tained through extraction techniques and how much is inferred by
argumentation. Generally the trade-off is accuracy vs. transparency.
In order to increase and maintain the accuracy of the agent, we
rely on three pillars: A) human operators keep providing training
data, which is done not only for keeping the data up-to-date, but
also to comply with expiry dates of data; B) the auto-experimenter
rigorously searches for the best models; and C) collaborations with
academia ensure that the latest academic results are tried and tested.