A systematic mapping study on automated analysis of privacy policies

All the following inclusion criteria must be met for a paper to pass to the manual filter:

In the manual stage, two screening phases were carried out: a title and abstract screening followed by a full-text screening, both performed through CADIMA (https://​www.​cadima.​info/​). We followed the decision tree shown in Fig. 1.

The list of inclusion-exclusion criteria used to evaluate the papers is included in Table 3. All criteria must hold for inclusion, but if any exclusion criterion holds then the paper is excluded.

Each paper was reviewed by two researchers and inconsistencies were resolved in daily meetings with all the team members. At the beginning of each stage, a pilot, divided into iterations, was conducted to align the criteria of all the researchers. In each iteration, five papers were analyzed by all of the team members and the Krippendorff’s alpha inter-coder reliability coefficient [20] was used to calculate the inter agreement. To finish a pilot, the agreement coefficients must be above the ‘good’ agreement threshold (0.8). Figure 2 shows the numbers of papers being considered in each step, distinguishing the ones retrieved from a database search (solid line) from the ones retrieved through snowballing (dotted line). The list of the 39 selected papers can be found in the “Appendix A”.

Most (nearly 90%) of the papers we have found identified specific content in the privacy policies (Table 4). In this table, the ‘Other’ group papers focus on contents related to specific privacy laws, i.e., CCPA (ID81) and GDPR (ID136). Remarkably, ID136 provides a GDPR conceptual model and a set of classifiers for identifying all these concepts in policy texts, including, for instance, information about automated decision making.

Only 15% of the papers analyzed focused on how the policy was written. We intentionally excluded papers that only report the use of existing readability scores, as their value is on the conclusions extracted from the application of the score to a given set of policies rather than on the novelty of the technique. However, we found novel approaches that might become useful to improve a privacy policy readability by detecting vagueness and inconsistencies.

Vagueness introduces ambiguous language that undermines the purpose and value of privacy policies as transparency elements. The system described in paper ID996 (2016) is aimed to detect if the words in a privacy policy are vague. Paper ID815 (2018) advances these results by classifying sentences in a privacy policy into different levels of vagueness, resulting in a more complete analysis.

An inconsistent privacy policy introduces contradictions between its contents, thus making it more difficult to understand. The authors of ID175 present a system capable of identifying contradictions in a given privacy policy. To this end, they perform an analysis of the privacy contents included in the text, model these contents, and then check if there is any contradiction between them. Other authors perform similar approaches but comparing the privacy policy of an application with the privacy policies of the software libraries it integrates (ID983) or comparing the privacy policy with the textual description provided by the developer (ID763).

As detailed in Table 6, contributions apply symbolic NLP at three different levels: morphological and lexical analysis of the words, syntactic and semantic analysis, and using ontologies to extract the meaning of the sentences.

The first levels of NLP (morphological and lexical analysis) have similar results to more complex techniques when targeting certain privacy practices. There are privacy practices (e.g., those related to encryption) that often use very specific distinctive terms (e.g., SSL). In such cases, a basic keyword-based analysis performs best (e.g., see ID72).

Most symbolic NLP techniques use some form of semantic analysis. The typical procedure followed in these cases consists of five phases: (1) splitting the policy into sentences, (2) parsing the words, e.g. through Part-of-Speech (PoS) tagging, (3) eliciting syntactic patterns related to a privacy practice, such as collection or disclosure, (4) detecting these patterns, and (5) deriving semantic meaning from them. The main differences between the authors are the proprietary tools or techniques implemented to carry out these tasks and the lexicons or taxonomies used.

We would like to highlight some of the most useful tools found in the study. The tool most used for carrying out syntactic analysis is the Stanford dependency parser, which is available in five different languages. One of the most critical stages is the creation of semantic patterns, which many authors manually create based on collections of privacy policies or taxonomies such as that created by Anton et al. [21]. By contrast, the authors of paper ID983 use a bootstrapping mechanism from Slankas et al. [22] to automatically find patterns from privacy policies according to a simple seed pattern. This process allows them to generate more inclusive patterns. Finally, there are some papers that report the use of specific programming languages to make annotations in the text and finding the patterns. This is the case of paper ID55 that uses Eddy [23], and paper ID1044 that uses Jape [24].

In our research, one in three papers using symbolic NLP techniques report the use of ontologies. Once a privacy policy text is represented as an ontology, information can be automatically extracted with query languages such as SPARQL. The most challenging part of the use of ontologies is the definition and the creation of the ontology. In most of the cases (50%), the creation of the ontology is a manual process conducted by a group of experts that annotate the privacy policies texts. Although this is the first step, what is really interesting is the automatic creation of the ontologies, which would allow researchers to analyze large amounts of policies without human intervention. Some examples of this are papers ID33, ID60, and ID175, each of which use a different approach. The authors of paper ID33 use Tregex patterns to detect information types automatically. In the case of ID60, they use semantic rules to extract relationships from information types. Finally, the authors of paper ID175 have created a method to “capture both positive and negative statements of data collection and sharing practices” based on Hearst patterns. This paper is relevant due to its ability to detect negative statements in comparison with other more limited approaches like those used by Zimmeck et al. in ID885 and by Yu et al. in ID983.

As Table 6 shows, contributions based on Statistical NLP use supervised (60%), unsupervised (7%), ANN-based techniques (26%), or a combination of them (7%). While supervised techniques are mostly used, ANN have begun to gain strength since their appearance around 2016.

Supervised algorithms are primarily used for classification tasks, such as detecting which personal data is collected, while unsupervised algorithms are used for clustering tasks such as topic modeling. As for the supervised algorithms, geometric algorithms such as Support Vector Machine (SVM) (ID59, ID62, ID64) and Logistic Regression (LR) (ID886, ID978, ID993) are mainly applied. Decision tree-based models are also used, like Decision Tree (DT) (ID81), the ensembles Random Forest (RF) (ID770, ID783), and AdaBoost (ID783), which tends to outperform the results of DT.

Unsupervised learning techniques apply Hidden Markov Models (HMM) to group segments of policies based on the privacy topic they address (ID1018) and a Latent Dirichlet Allocation (LDA) algorithm to determine the underlying topics in privacy policies (ID1045). Although one of the most interesting attributes of these approaches is the absence of a labeled corpus, it is important to highlight that in both cases, the authors had to create a labeled dataset to evaluate the accuracy of their models.

ANN-based techniques have been applied to tasks such as text classification (ID10), answerability prediction (ID810), or vagueness identification (ID815). We have found different approaches and papers using different kinds of neural networks. Convolutional Neural Networks (CNNs) (ID28, ID805), Recurrent Neural Networks (RNN) (ID996) and Google’s algorithm BERT (Bidirectional Encoder Presentations from Transformers) (ID796, ID810) are mostly used. Certain works comparing ANN and supervised learning techniques (ID10, ID810) report better performance in the case of ANN-based predictions.

Another important aspect in ANN is the technique used to represent every word so that it can be used by the neural network. The analyzed papers have used different techniques to create this word representation: fastTest (ID28), Word2Vec (ID810), Glove (ID815, ID996) and ELMo (ID805). These tools can create a word representation from a pre-trained model or from a specific model trained for the occasion. Authors seem to agree that training the model with a related corpus improves the results.

Learning algorithms require annotated datasets for training or validation. The creation of this dataset is a time-consuming task. On the other hand, these datasets are of paramount importance since the results and performance of the final model depend on the quality and the completeness of this data. Table 7 collates the information on all the public datasets of annotated privacy policies found in our research.