=Paper=
{{Paper
|id=Vol-2151/Paper_P1
|storemode=property
|title=Data Quality in Reasoning
|pdfUrl=https://ceur-ws.org/Vol-2151/Paper_P1.pdf
|volume=Vol-2151
|authors=Stephanie Inglis,Ehud Reiter,Somayajulu Sripada
|dblpUrl=https://dblp.org/rec/conf/sicsa/InglisRS18
}}
==Data Quality in Reasoning==
https://ceur-ws.org/Vol-2151/Paper_P1.pdf
Data Quality in Reasoning
Stephanie Inglis, Ehud Reiter and Somayajulu Sripada
Department of Computing Science
University of Aberdeen, Aberdeen, Scotland
{r01si14, e.reiter, yaji.sripada}@abdn.ac.uk
Abstract. Decision making using data is dependent on the quality of the data
being used to make those decisions. Currently, data-to-text recommendation
systems do not take this into consideration. Unsatisfactory recommendations
are likely to cause further damage, which could have a detrimental effect eco-
nomically or from a health and safety perspective. Highlighting quality issues in
data-to-text systems will allow readers to consider this.
Keywords: Data quality, data-to-text, reasoning.
1 Introduction
In a world so reliant on big data, a large amount of decision making and reasoning is
based on techniques such as data mining which is especially prevalent in policy mak-
ing and resource allocation. Performance measuring variables are used to detect prob-
lematic areas in an organisation. Additional resources, such as a larger budget, can be
given to the problematic area in an attempt to alleviate issues and boost productivity.
This scenario can apply to a wide array of real world situations from increasing a
company’s profit, to improving road safety or creating prevention policies to limit
damage from natural disasters. Due to ignorance or lack of awareness, the data used
to make these decisions is frequently of lower quality than expected. Data-to-text
systems are often used to summarise data or produce recommendations. When incor-
rect or incomplete data is used to make recommendations, their output could be dam-
aging. Acting on poor recommendations may be detrimental, resulting in loss of mon-
ey, or perhaps lives. Currently, the awareness and reporting of data quality issues
within these systems is inadequate. This paper looks at the impact of using poor quali-
ty data to make decisions on the allocation of money to improve road safety issues.
2 Road Traffic Reports
Multiple studies have been carried out by collecting data from hospitals on traffic
accident casualties, then comparing it with police reports to observe how many of
these traffic casualties had been reported to the police. Casualties not reported to the
police do not feature in road traffic statistics, making this data “missing” from official
2
reports such as the World Health Organisation’s (WHO) Global Traffic Reports [2],
and from datasets modelling road traffic incidents. Such models simulate road crashes
and assist in decision making on how to mitigate dangerous scenarios.
For all studies investigated, less severe injuries were less likely to be reported to the
police. Therefore, any datasets using police data is highly likely to have missing inju-
ry data. All studies had discrepancies between the injury severity on the police report,
and the severity assigned by the hospital. These conflicts are not rectified. It is likely
the hospital report is more accurate since medical staff are better qualified to assess
this variable, but again it is the police data used in statistical analyses.
One study was conducted in Birmingham where data was collected at the Accident
Hospital for the first 100 people who had been in a traffic accident for each month in
1970 [1]. All fatalities were reported, but reporting lessened with injury severity level.
Ignoring entries with conflicts for simplicity, 255 serious injuries were reported while
64 (20%) were not, and 528 slight injuries were reported while 289 (35%) were not.
This is typical of findings across studies, mostly because reporting to the police is not
always legally required, such as for single vehicle crashes where no one else is affect-
ed by the incident (80% missing), or situations where only property was damaged.
The most striking figure in this study is that 98% of pedal cyclists who were injured
with no other vehicle involved did not report a road accident to the police. Therefore,
situations where police figures are used to make decisions on cyclist safety will al-
most surely be incorrect. The data implies that cycling is very safe, as only 3 people
had an isolated cyclist incident, however the reality is that 112 people were injured.
3 Conclusion
Studies comparing police and hospital data about road traffic incidents show clearly
that the data is not as accurate as it seems to be. The impact of using low quality data
to make decisions can be very high depending on the domain, especially where peo-
ple’s lives are at risk. Being more aware of the quality of data being used in reasoning
will allow more correct conclusions and will yield greater results in the longer term.
Future work by the authors will seek to improve mindfulness of data quality in data-
to-text systems by highlighting issues to readers, by dynamically describing data qual-
ity issues that are present in the outputs of textual systems alongside normal output.
References
1. Bull, J. P., & Roberts, B. J. (1973). Road accident statistics-A comparison of police and
hospital information. Accident Analysis and Prevention, 5(1), 45–53.
2. World Health Organisation Traffic Reports,
http://www.who.int/violence_injury_prevention/publications/road_traffic/en/, last accessed
2018/05/30.