Graham McDonald

Towards a Classifier for Digital Sensitivity Review

The sensitivity review of government records is essential before they can be released to the official government archives, to prevent sensitive information such as personal information, or that which is prejudicial to international relations from being released. As records are typically reviewed and released after a period of decades, sensitivity review practices are still based on paper records. The transition to digital records brings new challenges, e.g. increased volume of digital records, making current practices impractical to use. In this paper, we describe our current work towards developing a sensitivity review classifier that can identify and prioritise potentially sensitive digital records for review. Using a test collection built from government records with real sensitivities identified by government assessors, we show that considering the entities present in each record can markedly improve upon a text classification baseline.

Using Part-of-Speech N-grams for Sensitive-Text Classification

Freedom of Information legislations in many western democracies, including the United Kingdom (UK) and the United States of America (USA), state that citizens have typically the right to access government documents. However, certain sensitive information is exempt from release into the public domain. For example, in the UK, FOIA Exemption 27 (International Relations) excludes the release of Information that might damage the interests of the UK abroad. Therefore, the process of reviewing government documents for sensitivity is essential to determine if a document must be redacted before it is archived, or closed until the information is no longer sensitive. With the increased volume of digital government documents in recent years, there is a need for new tools to assist the digital sensitivity review process. Therefore, in this paper we propose an automatic approach for identifying sensitive text in documents by measuring the amount of sensitivity in sequences of text. Using government documents reviewed by trained sensitivity reviewers, we focus on an aspect of FOIA Exemption 27 which can have a major impact on international relations, namely, information supplied in confidence. We show that our approach leads to markedly increased recall of sensitive text, while achieving a very high level of precision, when compared to a baseline that has been shown to be effective at identifying sensitive text in other domains.

Enhancing Sensitivity Classification with Semantic Features using Word Embeddings

Government documents must be reviewed to identify any sensitive information they may contain, before they can be released to the public. However, traditional paper-based sensitivity review processes are not practical for reviewing born-digital documents. Therefore, there is a timely need for automatic sensitivity classification techniques, to assist the digital sensitivity review process. However, sensitivity is typically a product of the relations between combinations of terms, such as who said what about whom, therefore, automatic sensitivity classification is a difficult task. Vector representations of terms, such as word embeddings, have been shown to be effective at encoding latent term features that preserve semantic relations between terms, which can also be beneficial to sensitivity classification. In this work, we present a thorough evaluation of the effectiveness of semantic word embedding features, along with term and grammatical features, for sensitivity classification. On a test collection of government documents containing real sensitivities, we show that extending text classification with semantic features and additional term n-grams results in significant improvements in classification effectiveness, correctly classifying 9.99% more sensitive documents compared to the text classification baseline.

Towards Maximising Openness in Digital Sensitivity Review Using Reviewing Time Predictions

The adoption of born-digital documents, such as email, by governments, such as in the UK and USA, has resulted in a large backlog of born-digital documents that must be sensitivity reviewed before they can be opened to the public, to ensure that no sensitive information is released, e.g. personal or confidential information. However, it is not practical to review all of the backlog with the available reviewing resources and, therefore, there is a need for automatic techniques to increase the number of documents that can be opened within a fixed reviewing time budget. In this paper, we conduct a user study and use the log data to build models to predict reviewing times for an average sensitivity reviewer. Moreover, we show that using our reviewing time predictions to select the order that documents are reviewed can markedly increase the ratio of reviewed documents that are released to the public, e.g. +30% for collections with high levels of sensitivity, compared to reviewing by shortest document first. This, in turn, increases the total number of documents that are opened to the public within a fixed reviewing time budget, e.g. an extra 200 documents in 100 hours reviewing.

Active Learning Strategies for Technology Assisted Sensitivity Review

Government documents must be reviewed to identify and protect any sensitive information, such as personal information, before the documents can be released to the public. However, in the era of digital government documents, such as e-mail, traditional sensitivity review procedures are no longer practical, for example due to the volume of documents to be reviewed. Therefore, there is a need for new technology assisted review protocols to integrate automatic sensitivity classification into the sensitivity review process. Moreover, to effectively assist sensitivity review, such assistive technologies must incorporate reviewer feedback to enable sensitivity classifiers to quickly learn and adapt to the sensitivities within a collection, when the types of sensitivity are not known a priori. In this work, we present a thorough evaluation of active learning strategies for sensitivity review. Moreover, we present an active learning strategy that integrates reviewer feedback, from sensitive text annotations, to identify features of sensitivity that enable us to learn an effective sensitivity classifier (0.7 Balanced Accuracy) using significantly less reviewer effort, according to the sign test (\(p<0.01\)). Moreover, this approach results in a 51% reduction in the number of documents required to be reviewed to achieve the same level of classification accuracy, compared to when the approach is deployed without annotation features.

A Study of SVM Kernel Functions for Sensitivity Classification Ensembles with POS Sequences

Freedom of Information (FOI) laws legislate that government documents should be opened to the public. However, many government documents contain sensitive information, such as confidential information, that is exempt from release. Therefore, government documents must be sensitivity reviewed prior to release, to identify and close any sensitive information. With the adoption of born-digital documents, such as email, there is a need for automatic sensitivity classification to assist digital sensitivity review. SVM classifiers and Part-of-Speech sequences have separately been shown to be promising for sensitivity classification. However, sequence classification methodologies, and specifically SVM kernel functions, have not been fully investigated for sensitivity classification. Therefore, in this work, we present an evaluation of five SVM kernel functions for sensitivity classification using POS sequences. Moreover, we show that an ensemble classifier that combines POS sequence classification with text classification can significantly improve sensitivity classification effectiveness (+6.09% F2) compared with a text classification baseline, according to McNemars test of significance.