Named Entity Recognition (NER) is a challenging learning task of identifying and classifying entity mentions in texts into predefined categories. In recent years, deep learning (DL) methods empowered by distributed representations, such as word- and character-level embeddings, have been employed in NER systems. However, for information extraction in Police narrative reports, the performance of a DL-based NER approach is limited due to the presence of fine-grained ambiguous entities. For example, given the narrative report “Anna stole Ada’s car”, imagine that we intend to identify the VICTIM and the ROBBER, two sub-labels of PERSON. Traditional NER systems have limited performance in categorizing entity labels arranged in a hierarchical structure. Furthermore, it is unfeasible to obtain information from knowledge bases to give a disambiguated meaning between the entity mentions and the actual labels. This information must be extracted directly from the context dependencies. In this paper, we deal with the Hierarchical Entity-Label Disambiguation problem in Police reports without the use of knowledge bases. To tackle such a problem, we present HELD, an ensemble model that combines two components for NER: a BLSTM-CRF architecture and a NER tool. Experiments conducted on a real Police reports dataset show that HELD significantly outperforms baseline approaches.

It is increasingly common to use chatbots as an interface to use services. Making this experience more humanized requires the chatbot to understand natural language and express itself using natural language. One crucial step to achieve this is to label the data with intentions and entities. After labeling, one can use the labeled data to train a Natural Language Understanding (NLU) component. The NLU component interprets the text extracting the intentions and entities present in that text. Manually label the data is an onerous and impracticable process due to the high volume of data. Thus, an unsupervised machine learning technique, such as data clustering, is usually used to find patterns in the data and thereby label them. For this task, it is essential to have an effective vector embedding representation of texts that depicts the semantic information and helps the machine understand the context, intent, and other nuances of the entire text. In this paper, we perform an extensive evaluation of different text embedding models for clustering, labeling, and training an NLU model using the text of attendances from the Coronavirus Platform Service of Ceará, Brazil. We also show how different text embeddings result in different clustering, thus capturing different intentions of patients.

Several crimes occur every day, and the first step in investigating these crimes begins with a police report. Victims report the criminal facts, which in turn must be detailed and contain accurate information about the incident or crime (e.g., factual, accurate, clear, concise, complete, and timely). In addition, the bulletin helps to safeguard the police operation itself, showing where the series of investigative operations that police agencies have been carrying out began. In cities with high crime rates, it is unfeasible to require the police to read and analyze all reported crime narratives. However, it would be helpful if employees could identify reports with similar modus operandi. Priority legal document retrieval is an information retrieval task used to retrieve past case documents related to specific cases and guide the police on how to act. Given a police report, the main objective of this work is to determine the most similar or duplicate police report. Another method is to encode the narrative as an embedded vector. In this article, we experimented with different pre-trained representations at the sentence level. We found the one that most effectively captures the semantic attributes of police report vocabulary and recognizes repeated reports. We are also investigating whether the summarized sentences identify duplicate police reports. Finally, we compare the effectiveness of the duplicated police report with the available sentence incorporation model trained in a large corpus. Our goal is to evaluate the performance of these embedding models (the one trained with our corpus and the pre-trained) to capture duplicate narratives.

In this article, we consider the External Sensor Trajectory Prediction problem for stolen vehicle trajectories. This analysis brings new challenges to the problem, as crime patterns are dynamic and drivers of stolen vehicles tend to move away from the sensors, which increases data dispersion. We analyze the effectiveness of different machine learning models and propose semantic enrichment with criminal data and points of interest to solve our problem. We also investigate the best attributes to improve EST prediction models, and how different spatial level representations can leverage prediction accuracy.

The movement of criminals is an important factor used in detecting crimes. Individuals sentenced to house arrest who wears an ankle monitor have their trajectories collected periodically. Each offender using an ankle monitor must adhere to a set of rules, for instance, be at his/her home during the night. Unfortunately, some of them break such rules, also some end up committing crimes again. In this demonstration1, we present a prototype system called Crime Monitor to monitor offenders in a semi-open regime. Crime Monitor reports the illegal activities to the police department in real-time based on trajectory features. Thus, the police can effectively prevent crimes from happening and handle them efficiently when they occur. We tackled the trajectory classification problem and used a deep learning model combining embedding with a recurrent neural network to classify illegal activities and learn the pattern regardless of who the criminal user is. We conduct experiments on a real dataset, and we show that DeepeST outperforms other approaches from state- of-the-art.

Huge volumes of data are produced every day on the Web. These are a big amount of videos, images, and texts that store unstructured information. Text summarization systems were created to facilitate the presentations of large amounts of textual data as well as to aid information retrieval over this type of data. The sentence compression has been developed due to the need for better summaries generated by these systems. However, when trained over domains with restricted amounts of labeled data for sentence compression, neural netword-based models tend to not be able to extract important features. Thus, to improve the performance of these models in this scenario, some pieces of information must be extracted and adapted before being used for training. Thus, we propose a sentence compression model capable of achieving competitive results, even when trained with smaller amounts of data, compared with other neural networkbased models, by using a set of linguistic features extracted from words alongside a rare words reduction strategy over the sentences.

There are relevant rates of violence in Brazil that have increased in recent years. Intelligence and efficiency are required to combat this issue, in order to reduce public money spending and time from public officials. Consequently, this operation may improve the safety of the population. There are several solutions that use intelligent systems to predict where and when a crime will occur, which allows police routes to be sent to areas with a higher risk of danger. In this paper, four machine learning methods are used to predict the location of where a crime will occur in a city of Fortaleza, Brazil. The final result shows that simple algorithms can be efficient in the task of crime prediction. In this paper, the Decision Tree and Bagging Regressor methods obtained the best predictions results.

The task of detecting anomalous data patterns is as important in practical applications as challenging. In the context of spatial data, recognition of unexpected trajectories brings additional difficulties, such as high dimensionality and varying pattern lengths. We aim to tackle such a problem from a probability density estimation point of view, since it provides an unsupervised procedure to identify out of distribution samples. More specifically, we pursue an approach based on normalizing flows, a recent framework that enables complex density estimation from data with neural networks. Our proposal computes exact model likelihood values, an important feature of normalizing flows, for each segment of the trajectory. Then, we aggregate the segments' likelihoods into a single coherent trajectory anomaly score. Such a strategy enables handling possibly large sequences with different lengths. We evaluate our methodology, named aggregated anomaly detection with normalizing flows (GRADINGS), using real world trajectory data and compare it with more traditional anomaly detection techniques. The promising results obtained in the performed computational experiments indicate the feasibility of the GRADINGS, specially the variant that considers autoregressive normalizing flows.

The ubiquity of GPS-enabled smartphones and automotive navigation systems allows to monitor and collect massive streams of trajectory data in real-time. This enables real-time analyses on mobility data in urban settings, which in turn have the potential to substantially improve traffic conditions, analyze congested areas, detect events in (quasi) real-time, and so on. While many existing approaches characterize past movements of moving objects from historical trajectory data, or address the problem of finding out clusters of moving objects from data streams, such approaches fail to capture how movement behaviors unravel over time - for instance, they fail to capture typically trafficked routes or traffic jams. In this work we propose NET-CUTiS, a novel approach that addresses the problem of discovering and monitor the evolution of clusters of trajectories over road networks from trajectory data streams. We conduct several experiments that demonstrate the validity of our proposal in terms of clustering quality and run-time performance.

Recent approaches for online clustering of moving objects location are restricted to instantaneous positions. Subse-quently, they fail to capture the behavior of moving objects over time. By continuously tracking sub-trajectories of moving object at each time window, it becomes possible to gain insight on the current behavior and potentially detect mobility patterns in real time. In our previous work [1], we proposed CUTiS, an incremental algorithm for discovering and maintaining the density-based clusters in trajectory data streams, while tracking the evolution of the clusters. This paper extends [1] to CUTiS* by proposing an indexing structure for sub-trajectory data based on a space-filling curve. The proposed index improves the performance of our approach without losing quality in the clusters results as we show in our experiments conducted on a real dataset.

Movement tracking becomes ubiquitous in many applications, which raises great interests in trajectory data analysis and mining. Most existing approaches cluster the whole trajectories offline. This allows characterizing the past movements of the objects but not current patterns. Recent approaches for online clustering of moving objects location are restricted to instantaneous positions. Subsequently, they fail to capture moving objects' behavior over time. By continuously tracking moving objects' sub-trajectories at each time window, rather than just the last position, it becomes possible to gain insight on the current behavior, and potentially detect mobility patterns in real time. In this work, we tackle the problem of discovering and maintaining the density based clusters in trajectory data streams, despite the fact that most moving objects change their position over time. We propose CUTiS, an incremental algorithm to solve this problem, while tracking the evolution of the clusters as well as the membership of the moving objects to the clusters. Our experiments were conducted on real data sets, and it shows the efficiency and the effectiveness of our method.

Trajectory pattern mining allows characterizing movement behavior, which leverages new applications and services. Most existing approaches analyse the whole object trajectory rather that the current movement. Besides existing approaches for online pattern discovery are restricted to instantaneous positions. Subsequently, they fail to capture the movement behaviour along time. By continuously tracking moving objects sub-trajectories at each time window, rather than just the last position, it becomes feasible to gain insight on the current behaviour, and potentially detect mobility patterns in real time. This demonstration presents a novel framework for online mobility pattern discovery in sub-trajectory data streams. Key innovations include: (i) Online discovery of mobility patterns and pattern evolution by tracking the sub-trajectories of moving objects, (ii) A novel structure, called micro-group, to represent the relationship among moving objects, and (iii) An incremental algorithm to maintain micro-groups and to capture their evolution on highly dynamic sub-trajectory data. We present various demonstration scenarios using a real data set.