A semantic-aided abnormal broadcast content analysis method Authors: Qi Zhang, Fengming Zhou, Yuzhi Zhang, Huan Wang, Jianxin Zhang Published: 2026 ABSTRACT Abnormal broadcast activities are becoming more and more rampant, since it is easy and cheap to achieve a radio broadcasting station. These abnormal broadcasts disseminate illegal information, which can have serious negative impacts on society. However, detecting and tracing these abnormal broadcast contents can be challenging as they are often scattered and difficult to identify. In this article, we propose a semantic-aided analysis method for broadcast content (SAAM), which is based on Latent Dirichlet Allocation (LDA) to construct a model for detecting and analyzing abnormal broadcasts. The LDA model intelligently captures the latent semantic topics implied in the broadcasting content, enabling the establishment of criteria for identifying anomaly and facilitating to investigate. The analysis method further exploits the Apriori algorithm to explore the associations between the content of abnormal broadcast and relevant organizations based on semantic topics from LDA. This allows for detailed analysis of abnormal broadcasts as well as tracing relevant organizations involved in such activities. In the experiment, the abnormal broadcast data is collected to verify the accuracy of SAAM in detecting and tracing abnormal broadcast. The experiment results show that, the accuracy rate of SAAM in detecting abnormal broadcasts is more than 0.91 and the tracing-accuracy is more than 0.8. Introduction With the advancement of national information construction, broadcast remains one of the most popular methods for multimedia communication, providing convenience to people’s lives while also providing a means for the spread of abnormal information1. The abnormal broadcast refers to an unauthorized broadcast signal that does not comply with national regulations. It is not authorized by the State Administration of Radio, Film and Television (SARFT), radio management agencies, or other management departments. It often broadcasts abnormal content and operates on the same frequency as authorized broadcast by suppressing the authorized broadcast2. At present, the radio transmitting equipment is gradually small, lightweight and its components are easy to assemble. Thus, radio is widely used by illegal organizations to spread fraud, vulgarity, pornography, cults or other abnormal broadcast content, and even installed in the border areas to preach sermons, reactionary, violent content, and so on. The content spread by abnormal broadcasts infringes on the rights and interests of the public, endangers national security, and harms the social environment3. Furthermore, abnormal broadcast transmission equipment has high-power and electromagnetic radiation, which seriously disrupts the order of electromagnetic waves, interferes with normal broadcast signals, and poses a threat to public safety. Abnormal broadcasts often display characteristics of concealment and trick. Illegitimate broadcast organizations often occupy legally authorized frequency bands to avoid detection by regulatory authorities. Moreover, in order to confuse the public, these organizations frequently update the text of abnormal content and mix it with regular programming, such as interactive communication, which results in the scattered and diversity of abnormal broadcast content. This makes it challenging to accurately define the features of abnormal content, further complicating efforts to identify abnormal broadcasts. Especially, the abnormal broadcast content released by illegal organizations lacks structured data directly related to the publishers, making it difficult to associate the corresponding organizations. The initial approach for identifying abnormal broadcasts primarily relies on manual monitoring and public complaints, which necessitates a significant amount of manpower and equipment. However, the effectiveness of these methods in identification is unsatisfactory. Currently, the common method involves analyzing and evaluating the frequency or audio characteristics of collected abnormal broadcast bands5, as described in references3,6–12. Since abnormal broadcasts fundamentally deviate from national regulations, it is more effective to identify abnormalities based on their content itself. Nowadays, keyword matching methods are employed in abnormal content analysis by creating rules with a word list to recognize abnormal information. To determine whether the tested text contains any abnormal content, they are accomplished by the manually screening keywords and matching with the tested text, such as the solutions in references13–17. Although this method is simple and easy to implement, it requires extensive human involvement in screening abnormal keywords. Additionally, rules obtained in this way are subjective and rely heavily on expert experience. And the rules are inadequate to solve the problem of word polysemy and context. That may lead to more false positives. Furthermore, machine learning techniques are frequently utilized for identifying abnormal content18–27. Through algorithmic model training, the machine learning approach effectively mines rules for identifying and classifying content. The machine learning domain is not reliant on manual selection of keywords and can efficiently accomplish anomaly detection of content. However, some analysis processes of the machine learning resemble black boxes resulting in limited interpretability of anomalies. Therefore, it is difficult to the further fine-grained analysis of the detected abnormal broadcast. For instance, we want to discover the association between the abnormal broadcasts and the relevant organization to trace their accountability organization. In fact, the abnormal broadcasts consist of diverse keywords and possess a central idea, despite their complex and scattered content. Therefore, by extracting the text’s topic, we can capture the main clues running throughout the entire broadcast to determine its abnormality. On the other hand, there are few characteristics indicating the accountability organization in abnormal broadcast content. However, these broadcasts released by specific organizations generally revolve around particular topics. For instance, when radio is utilized to promote an informal medicine to the public, constant insertions of small talk regarding the medicine’s efficacy and target audience occur. Therefore, through analyzing the correlation between semantic topics and accountability organizations, we can establish attribution criteria for abnormal broadcasts and related entities. Motivated by it, this article aims to develop a system that aggregates existing case of abnormal broadcasts for detecting abnormal broadcast as well as tracing the relevant organizations. We introduce a semantic aided analysis method for abnormal broadcasts (SAAM). The basic idea behind SAAM is that we employ sematic analysis to mine the hidden broadcast topic and the keywords related to the topic. LDA is an unsupervised machine learning algorithm that is able to extract text topics and obtain keywords related to the topic. At present, LDA has achieved success in the identification of malicious emails and social media, system logs, network traffic, and other anomalies detection28. For example, Jian Yang29 uses LDA algorithm to capture potential attack intent and reconstruct APT scenarios to achieve detection of multi-stage threats over a long period of time. Similarly, LDA is able to carry out topic mining for SAAM. The obtained keywords with high probability related to the topic can serve as the objective rules for anomaly detection. After the anomaly detection, an association analysis is required to further trace abnormal broadcasts. Apriori algorithm is widely applicable and easy to generate interpretable association rules30. It has shown excellent performance in fraud detection31, network traffic analysis32, and other fields by effectively mining the association relationships between items33. Therefore, we further invoke Apriori algorithm to extract the associations between the abnormal broadcasts and the accountability organization. Finally, the abnormal broadcast is discovered and the hidden organization is exposed. The contributions of this paper are mainly three-fold: Semantic-based anomaly broadcast detection mechanism: We design and implement SAAM abnormal broadcasts detection. The LDA algorithm is incorporated in SAAM to perform semantic analysis on broadcast text, which capture the hidden topics of broadcast. And SAAM automatically generates detection rules with the latent topic label of word extracted from LDA.Semantics-assisted traceability model of accountability for abnormal broadcasts: Based on the broadcast topics obtained from the above semantic analysis of abnormal broadcast content, the Apriori algorithm is used to mine the correlation between different broadcast topics and the corresponding accountability organizations. Thus, the attribution criteria of the related organizations of abnormal broadcasts are constructed.Dataset is constructed by capturing real broadcast: We gather broadcast data from various regions and transform them into textual data, to help in-depth analysis of abnormal broadcasts and conduct experiments. We organize and label the collected data, which results in a broadcast text database and historical records of abnormal content. We then construct test dataset with both abnormal and normal broadcast content. These datasets aim to facilitate validation of detection and further analysis of abnormal broadcasts. Semantic-based anomaly broadcast detection mechanism: We design and implement SAAM abnormal broadcasts detection. The LDA algorithm is incorporated in SAAM to perform semantic analysis on broadcast text, which capture the hidden topics of broadcast. And SAAM automatically generates detection rules with the latent topic label of word extracted from LDA. Semantic-based anomaly broadcast detection mechanism: We design and implement SAAM abnormal broadcasts detection. The LDA algorithm is incorporated in SAAM to perform semantic analysis on broadcast text, which capture the hidden topics of broadcast. And SAAM automatically generates detection rules with the latent topic label of word extracted from LDA. Semantics-assisted traceability model of accountability for abnormal broadcasts: Based on the broadcast topics obtained from the above semantic analysis of abnormal broadcast content, the Apriori algorithm is used to mine the correlation between different broadcast topics and the corresponding accountability organizations. Thus, the attribution criteria of the related organizations of abnormal broadcasts are constructed. Semantics-assisted traceability model of accountability for abnormal broadcasts: Based on the broadcast topics obtained from the above semantic analysis of abnormal broadcast content, the Apriori algorithm is used to mine the correlation between different broadcast topics and the corresponding accountability organizations. Thus, the attribution criteria of the related organizations of abnormal broadcasts are constructed. Dataset is constructed by capturing real broadcast: We gather broadcast data from various regions and transform them into textual data, to help in-depth analysis of abnormal broadcasts and conduct experiments. We organize and label the collected data, which results in a broadcast text database and historical records of abnormal content. We then construct test dataset with both abnormal and normal broadcast content. These datasets aim to facilitate validation of detection and further analysis of abnormal broadcasts. Dataset is constructed by capturing real broadcast: We gather broadcast data from various regions and transform them into textual data, to help in-depth analysis of abnormal broadcasts and conduct experiments. We organize and label the collected data, which results in a broadcast text database and historical records of abnormal content. We then construct test dataset with both abnormal and normal broadcast content. These datasets aim to facilitate validation of detection and further analysis of abnormal broadcasts. The rest of the paper is organized as follow. Related work is comprehensively reviewed in Sect. 2. Section 3 describes the anomaly broadcast analysis method based on fused semantic analysis. Section 4 presents and analyzes the experimental results. Finally, the conclusion is drawn in Sect. 5. Related work The traditional methods for identifying abnormal broadcasts primarily involve manual mobile patrol listening or relying on public complaints, which necessitate significant human and material resources, exhibit low efficiency, and are challenging to implement34. Currently, the commonly used method is to detect abnormal broadcasts based on broadcast audio features or frequency band features. One of the techniques utilizing audio feature recognition technology typically involves retrieving segments containing broadcast audio content and comparing the acoustic features of keywords to identify abnormal broadcasts. For instance, this may entail extracting speech acoustic matrix information and employing neural networks for recognition35. However, abnormal broadcasts often come with a substantial amount of background noise that significantly affects the comparison of acoustic features. Furthermore, the method for detecting anomaly broadcasts has evolved into identifying them through their frequency characteristics, such as using machine learning methods to study the spectrum information. Lin Qiuhai7 and colleagues introduced the Bayesian discriminative method, which extracts six signal features that effectively identify legal music broadcasts, legal news broadcasts, and illegal Frequency Modulation (FM) broadcasts. Chaowei Wang8 used standard Euclidean distance to extract the feature signals and then used Support Vector Machine (SVM) classifier based on clustered samples to classify the broadcast signals. Jitong Ma9 and colleagues used Long Short Term Memory (LSTM) identification network to identify abnormal broadcasts, which can be more effective in identifying abnormal broadcasts in terms of recognition accuracy, and through the special gate structure inside the LSTM framework, the identification features such as the channel state information and the Radio Frequency (RF) identification features are effectively extracted. Zhu Ye11 developed a method to detect abnormal broadcasts by reporting the characteristics of the zero-point signal spectrum at consecutive integer points. Park Dae-kyeong et al.12 proposed an unsupervised learning-based threat detection system using radio frequency signal characteristic data. Huan Wang et al.3 extract time–frequency features of radio signals and comprise an autoencoder and a Generative Adversarial Network to detect anomalous radio. The mentioned methods leverage broadcast signals for extracting and classifying features and has a good effect on identifying abnormal broadcasts signals36,37. The aforementioned approach aims to assess the signal characteristics of broadcasts for anomaly detection, but it is more informative to distinguish anomalies based on the content of the broadcasts. Existing methods for analyzing abnormal content can be broadly categorized into two types. One kind of them employs a keyword matching technique, where abnormal keywords are manually identified in advance and compiled into a specialized lexicon. This lexicon is then utilized to compare and determine whether the tested text contains any abnormal information. For example, Dimopoulos V13 converts character comparison into state transfer to match abnormal content strings through a finite state machine; Xian et al.14 propose an improved exhaustive automaton for sensitive word recognition in order to improve the problem of low matching efficiency; Deng Yigui et al.15 propose a decision tree-based information filtering algorithm which takes into account factors other than sensitive words; Manber et al.16 proposed a WM algorithm that constructs three tables of shift, hash, and prefix by introducing a block of characters for a large jump of bad characters. Liao Qinting et al.17 proposed an improved and efficient BMHS pattern matching algorithm for fast localization of sensitive information in text content. The involvement of manual intervention is inevitable in such methods. Usually, if the recognition rules are too subjective, the accuracy of abnormal content recognition will be affected. Another kind is to use machine learning for abnormal content detection. Lu Xiangyan18 proposed a sensitive information detection method using SVM algorithm training, which is useful for identifying unwanted information on the Internet. Tong Ying19 proposed an early warning algorithm based on deep neural network for highly sensitive information on the network. To eliminate the need for human expert intervention in feature engineering, Trivikram Muralidharan20 proposes a detection framework to detect malicious content in all email segments using deep integrated learning analysis. Zhao21 used pre-trained language model to analyzed text content to identify fake news. Che22and Birunda23 respectively prompt learning and random forest models to detect fake news. Zouak25 leveraged BERT (Bidirectional Encoder Representations from Transformers) model reveals the sematic content of spam email. Das24, Choi26, and Olivo27 respectively used LSTM and deep convolutional neural networks, SVM and autoencoder, ensemble classifier to detect spam. Machine learning overcomes the drawbacks of traditional keyword matching methods and enables the analysis of abnormal content without relying on manual keyword selection. BERT and sBERT have the capabilities for nuanced language analysis, and they are pre-trained model. Decision Tree, Support Vector Machine (SVM), neural network and Latent Dirichlet Allocation (LDA) are also commonly used for classification of natural language processing. Decision Tree, SVM and neural network are supervised learning methods, and they require manual labeled data for training29. LDA is an unsupervised learning topic modeling technique and does not require pre-training. Some of these machine learning algorithms often suffer from the “black box” dilemma, a lack of transparency that hinders their applicability in security contexts38. And it is difficult for them to achieve further trace. To address the issues, this paper proposes a novel approach that integrates LDA semantic fusion of the broadcast text for both anomaly detection and attribution. Methods This paper aims to extract the implied theme of the broadcast’s content, in order to detect its anomaly as well as mining the accountability organization. The overview of semantic-aided analysis method for abnormal broadcasts as shown in Fig. 1. The proposed method includes the following stages: preprocessing of broadcast data, identification of abnormal broadcast content, organization associating and integration of alarm information alerts. Abnormal content is disseminated in various forms by abnormal broadcast propagators. Such contents are centered around specific topics and there are certain correlations exist between the abnormal broadcasting topics and their organizations. Based on this fact, this paper firstly automatically generates rules from historical abnormal broadcast recordings by sematic and association analysis. Then, the new broadcast is analyzed semantically and to be matched with the generated rules to achieve anomaly analysis. According to Fig. 1 the rule generation and new broadcast anomaly analysis basic operation process are explained as follow. Fig. 1The architecture of SAAM. The architecture of SAAM. Firstly, the multi-source historical anomaly broadcast content is integrated and converted into text. Then LDA algorithm is utilized to extract the topics of these broadcast text to automatically generate the detection rules from these historical broadcasts. After that, based on the topic classification from LDA the association algorithm Apriori is used to deduce the association criteria between abnormal broadcasting topics and their accountability organizations. The pending broadcasts are also transformed into text. LDA analysis to get the topic of each broadcast text and the keywords corresponding to the topic. The topic keywords are matched with the pre-determined detection rules to detection conclusion. Finally, the detected abnormal broadcasts’ topic is matched with the organization association criteria in order to achieve the accountability institution tracing. Data preprocessing Firstly, the raw broadcast audio will be processed into text documents by speech recognition. The text documents will be further processed, including the removal of punctuation and special characters. Finally, the text is divided into word units to facilitate more accurate expression of the meaning of the text. Through the processing of speech recognition, data cleaning, and Chinese word segmentation in the data preprocessing module, the broadcast audio is transformed into easier-to-process word sequences for semantic analysis. Semantic analysis model for broadcasts After preprocessing, this paper uses semantic analysis to mine the hidden topics in the broadcast so as to realize the identification and trace of the abnormal broadcast. The LDA algorithm is currently a popular model for mining potential topics of text in the field of natural language processing and has been widely used in the fields of text mining and information retrieval34. In our method, the identification of abnormal broadcasts is based on the main idea of its content which is regarded as the feature of various types of abnormal broadcasts. Through LDA semantic analysis, the potential idea of each broadcast can be given, while the essence of the topic can be given by the probability distribution of high-probability words in the broadcast. The LDA semantic analysis can assist us in understanding the potential idea of each broadcast so that the abnormal broadcasts can be captured more accurately. Fig. 2LDA model for broadcast document. LDA model for broadcast document. LDA model is a multilayer Bayesian topic model based on the bag-of-words approach, which extracts topic information using unsupervised learning techniques39. There are three levels to the LDA representation: document, topic and word. Meanwhile the LDA considers a document is produced by multiple topics with varying probabilities, and each word is generated by multiple topics with different probabilities. Both the document-topic distribution and the topic-word distribution follow the Dirichlet distribution40. Figure 2 shows a graphical model of our proposed topic model for broadcasting. In this semantic model, each broadcast text is a document which is further characterized by a distribution of K topics. D is the total number of all the broadcast texts, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{d}$$\end{document} is the total number of words in a broadcast text \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{d}(\mathrm{w}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{e}\:d=\mathrm{1,2},\dots\:,D)$$\end{document} and V is the total number of words in the corpus. The LDA model assumes that a broadcast text\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{d}$$\end{document} can be represented by the topic distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\:\theta\:}}_{d}$$\end{document}, the broadcast topic is represented by the word distribution\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{\overrightarrow{\beta\:}}_{k}(\mathrm{w}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{e}\:k=\mathrm{1,2},\dots\:,K\:$$\end{document}),The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{d,n}$$\end{document} (where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n=\mathrm{1,2},\dots\:,{N}_{d}$$\end{document} ,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d=\mathrm{1,2},\dots\:,D$$\end{document}) is a word in broadcast text \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{d}$$\end{document}, and it is the base unit in LDA. The document level consists of the broadcast texts \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({f}_{1},{f}_{2},\dots\:,{f}_{D}\right)$$\end{document}, a total of K topics in all the broadcasts are denoted by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\overrightarrow{z}=\left[{z}_{1},{z}_{2},\dots\:{z}_{K}\right]$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{z}}_{d}=\left[{z}_{d,1},{z}_{d,2},\dots\:,{z}_{d,{N}_{d}}\right]$$\end{document} is the topic of words in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{\mathrm{d}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{z}_{d,n}\:$$\end{document}denotes the topic corresponding to the nth word in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{\mathrm{d}}$$\end{document}. In Fig. 2, broadcasts are assumed to be the corpus of LDA generative model. The steps for each document \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{d}$$\end{document} can be approximated by following process. First, the distribution of broadcast document-topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\theta\:}}_{d}=Dirichlet\left(\overrightarrow{\boldsymbol{\alpha\:}}\right)$$\end{document} and the distribution of topic-word \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\beta\:}}_{k}=Dirichlet\left(\overrightarrow{\eta\:}\:\right)$$\end{document} are taken the values according to the prior Dirichlet hyperparameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\overrightarrow{\boldsymbol{\alpha\:}}\mathrm{、}\overrightarrow{\boldsymbol{\eta\:}}$$\end{document}. Then, for each of the topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{d,n}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} is choose according to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:multinomial\left({\overrightarrow{\theta\:}}_{d}\right)$$\end{document}. Next, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{v}$$\end{document} is choose for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{d,n}$$\end{document} according to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:multinomial\left({\overrightarrow{\beta\:}}_{k}\right)$$\end{document}. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:multinomial(\bullet\:)$$\end{document} is a multinomial with only one trial. In this model there are D Dirichlet distributions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\theta\:}}_{d}=\left[{\theta\:}_{d1}\:,{\theta\:}_{d2}\:,\:{\theta\:}_{d3},...{\theta\:}_{dk}\right]$$\end{document} for D text topics, corresponding to D multinomial distributions with D topic numbers, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\overrightarrow{\alpha\:}\to\:{\overrightarrow{\theta\:}}_{d}\to\:{\overrightarrow{z}}_{d}$$\end{document} forms the Dirichlet-multi conjugate, and the topic conditional distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left({\overrightarrow{z}}_{d}|\overrightarrow{\alpha\:}\right)=\frac{{\Delta\:}({\overrightarrow{n}}_{d}+\overrightarrow{\alpha\:})}{{\Delta\:}\left(\overrightarrow{\alpha\:}\right)}$$\end{document} is computed for the dth document. The topic distribution of D broadcast text is given by Eq. (1):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left(\overrightarrow{z}|\overrightarrow{\alpha\:}\right)=\prod\:_{d=1}^{D}p\left({\overrightarrow{z}}_{d}|\overrightarrow{\alpha\:}\right)=\prod\:_{\mathrm{d}=1}^{\mathrm{D}}\frac{{\Delta\:}({\overrightarrow{\mathrm{n}}}_{\mathrm{d}}+\overrightarrow{{\upalpha\:}})}{{\Delta\:}\left(\overrightarrow{{\upalpha\:}}\right)}\:\:$$\end{document} Similarly, there are the Dirichlet distribution of K topics and words in the proposed model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\beta\:}}_{k}=\left[{\beta\:}_{k1}\:,{\:\beta\:}_{k2}\:,\:{\:\beta\:}_{k3},...{\beta\:}_{kv}\right]\:(\mathrm{w}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{e}\:v=\mathrm{1,2},\dots\:,V\:)\:$$\end{document}, and corresponds to a multinomial distribution with K topics. And the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\overrightarrow{\eta\:}\to\:{\overrightarrow{\beta\:}}_{k}\to\:{\overrightarrow{w}}_{k}$$\end{document} forms the Dirichlet-multi conjugate, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\:w}}_{k}$$\end{document} is the word corresponding to the topic when the topic is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document}. The conditional distribution of the words corresponding to all the topics \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left(\overrightarrow{w}|\overrightarrow{z},\overrightarrow{\eta\:}\right)$$\end{document} is given by Eq. (2):2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{p}\left(\overrightarrow{\mathrm{w}}|\overrightarrow{\mathrm{z}},\overrightarrow{{\upeta\:}}\right)=\prod\:_{\mathrm{k}=1}^{\mathrm{K}}\mathrm{p}\left({\overrightarrow{\mathrm{w}}}_{\mathrm{k}}|\overrightarrow{\mathrm{z}},\overrightarrow{{\upeta\:}}\right)=\prod\:_{\mathrm{k}=1}^{\mathrm{K}}\frac{{\Delta\:}({\overrightarrow{\mathrm{n}}}_{\mathrm{k}}+\overrightarrow{{\upeta\:}})}{{\Delta\:}\left(\overrightarrow{{\upeta\:}}\right)}$$\end{document} The joint distribution of topics and words \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left(\overrightarrow{w},\overrightarrow{z}|\overrightarrow{\alpha\:},\overrightarrow{\eta\:}\right)$$\end{document} is given by Eq. (3):3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left(\overrightarrow{w},\overrightarrow{z}\right)\propto\:p\left(\overrightarrow{w},\overrightarrow{z}|\overrightarrow{\alpha\:},\overrightarrow{\eta\:}\right)=p\left(\overrightarrow{z}|\overrightarrow{\alpha\:}\right)p\left(\overrightarrow{w}|\overrightarrow{z},\overrightarrow{\eta\:}\right)=\prod\:_{d=1}^{D}\frac{{\Delta\:}({\overrightarrow{n}}_{d}+\overrightarrow{\alpha\:})}{{\Delta\:}\left(\overrightarrow{\alpha\:}\right)}\prod\:_{k=1}^{K}\frac{{\Delta\:}({\overrightarrow{n}}_{k}+\overrightarrow{\eta\:})}{{\Delta\:}\left(\overrightarrow{\eta\:}\right)}$$\end{document} For \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\theta\:}}_{d}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\beta\:}}_{k}$$\end{document} the corresponding posterior distribution is given by the following Eq. (4) and Eq. (5):4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:p\left({\overrightarrow{\theta\:}}_{d}|{\overrightarrow{w}}_{\neg\:i},{\overrightarrow{z}}_{\neg\:i}\right)=Dirichlet\left({\overrightarrow{\theta\:}}_{d}\right|{\overrightarrow{n}}_{d,\neg\:i}+\overrightarrow{\alpha\:})$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left({\overrightarrow{\beta\:}}_{k}|{\overrightarrow{w}}_{\neg\:i},{\overrightarrow{z}}_{\neg\:i}\right)=Dirichlet\left({\overrightarrow{\beta\:}}_{k}\right|{\overrightarrow{n}}_{k,\neg\:i}+\overrightarrow{\eta\:})$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\mathrm{n}}}_{\mathrm{d},\neg\:i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{n}}_{k,\neg\:i}$$\end{document} are the statistic numbers after removing the word \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{d,i}\:$$\end{document}. Gibbs sampling is a sampling technique that relies on Markov Chain Monte Carlo (MCMC). In the context of the LDA algorithm, Gibbs sampling can iteratively approximate the true topic and word distributions39. Gibbs sampling samples the topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{i}$$\end{document} for word \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{i}$$\end{document} according to the conditional probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left({z}_{i}|\overrightarrow{w},{\overrightarrow{z}}_{\neg\:i}\right)$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{z}}_{\neg\:i}$$\end{document} represents the broadcast topics of all the wors except \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{i}$$\end{document}. This conditional probability can be obtained as:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left({z}_{i}=k|\overrightarrow{w},{\overrightarrow{z}}_{\neg\:i}\right)=\frac{{n}_{d,\neg\:i}^{k}+{\alpha\:}_{k}}{\sum\:_{s=1}^{K}{n}_{d,\neg\:i}^{s}+{\alpha\:}_{s}}\frac{{n}_{k,\neg\:i}^{t}+{\eta\:}_{t}}{\sum\:_{f=1}^{V}{n}_{k,\neg\:i}^{f}+{\eta\:}_{f}}$$\end{document} The Gibbs sampling samples the topics of all words repeatedly, and when the iterations converge, we calculate the broadcast-topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\theta\:}}_{d}$$\end{document} and the topic-word \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\beta\:}}_{k}$$\end{document} respectively, as follows:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{dk}=\frac{{n}_{d,\neg\:i}^{k}+{\alpha\:}_{k}}{\sum\:_{s=1}^{K}{n}_{d,\neg\:i}^{s}+{\alpha\:}_{s}}$$\end{document}and8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\:\beta\:}_{kv}=\frac{{n}_{k,\neg\:i}^{t}+{\eta\:}_{t}}{\sum\:_{f=1}^{V}{n}_{k,\neg\:i}^{f}+{\eta\:}_{f}}$$\end{document} An accountability organization association model for abnormal broadcasts Algorithm 1Apriori Apriori In the abnormal broadcasts, there is no obvious metadata which can be used to trace the relevant publishing organizations. Moreover, illegal broadcasters often conceal its identity deliberately, which makes it more difficult to discover the accountability organizations for the abnormal broadcasts. However, each organization’s broadcast usually has its own specific topics. The prior abnormal broadcast records database has a verified anomaly accountability organization. The semantic topics of these prior broadcasts can be obtained according to the broadcast semantic analysis model in Sect. 3.2. Therefore, we employ the Apriori algorithm41 to construct a broadcast organization association model which enables us to obtain the association mechanism between publishing organizations and the abnormal broadcast events. Thus, the accountability organization for abnormal broadcasts is exposed. There are M broadcast speech records constituting a database of abnormal broadcast content. All of the records are divided into K topics \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}\:(where\:k=\mathrm{1,2},\dots\:,K)\:$$\end{document} and they are related to Q organizations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}\:(where\:q=\mathrm{1,2},\dots\:,Q)$$\end{document}. This section analyses the relationship between topics and the organizations of abnormal broadcasts in the database. So frequent item is the anomaly broadcast topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} with organization \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}$$\end{document}. Apriori is an algorithm30 which makes multiple passes over the database to find all frequent item sets in a given prior abnormal broadcast database. Thus, the rules of association relationship are revealed42. This algorithm works as follows. The first pass of the algorithm counts the number of single item occurrences to determine the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{1}$$\end{document} or the single member frequent item sets. Then each subsequent pass generates the candidate item set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{S}$$\end{document} from the frequent item set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{s-1}$$\end{document}, using the Apriori candidate generation algorithm. Next, the database is scanned and the support of the candidates in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{S}$$\end{document} is determined to ensure that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{S}\:$$\end{document}item sets are frequent item sets. The Apriori is summarized in Algorithm 1. The steps of the algorithm are as follow:Step 1: Firstly, it scans the broadcast database and record the frequency of each item in each transaction mark. After that it generates the candidate item set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{1}$$\end{document} and remove the single item set whose frequency is lower than the minimum support threshold. Then it forms the frequent 1-item set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{1}$$\end{document}.Step 2: On the basis of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{1}$$\end{document}, connect to generate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{2}$$\end{document}-candidate item set, calculate the support of each item, and then prune the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{2}$$\end{document}-candidate item set to get frequent 2-itemsets \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{2}$$\end{document}.Step 3: Iterate the above process until no more higher order frequent item set are generated to obtain the frequent s-itemset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{s}$$\end{document}, which is the set of all frequent item set mined.Step 4: Calculate the confidence level of each item and eliminate the rules below the minimum confidence threshold to filter out the strong association rules that abnormal broadcast topics and accountability organizations. Step 1: Firstly, it scans the broadcast database and record the frequency of each item in each transaction mark. After that it generates the candidate item set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{1}$$\end{document} and remove the single item set whose frequency is lower than the minimum support threshold. Then it forms the frequent 1-item set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{1}$$\end{document}. Step 1: Firstly, it scans the broadcast database and record the frequency of each item in each transaction mark. After that it generates the candidate item set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{1}$$\end{document} and remove the single item set whose frequency is lower than the minimum support threshold. Then it forms the frequent 1-item set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{1}$$\end{document}. Step 2: On the basis of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{1}$$\end{document}, connect to generate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{2}$$\end{document}-candidate item set, calculate the support of each item, and then prune the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{2}$$\end{document}-candidate item set to get frequent 2-itemsets \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{2}$$\end{document}. Step 2: On the basis of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{1}$$\end{document}, connect to generate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{2}$$\end{document}-candidate item set, calculate the support of each item, and then prune the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{2}$$\end{document}-candidate item set to get frequent 2-itemsets \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{2}$$\end{document}. Step 3: Iterate the above process until no more higher order frequent item set are generated to obtain the frequent s-itemset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{s}$$\end{document}, which is the set of all frequent item set mined. Step 3: Iterate the above process until no more higher order frequent item set are generated to obtain the frequent s-itemset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{s}$$\end{document}, which is the set of all frequent item set mined. Step 4: Calculate the confidence level of each item and eliminate the rules below the minimum confidence threshold to filter out the strong association rules that abnormal broadcast topics and accountability organizations. Step 4: Calculate the confidence level of each item and eliminate the rules below the minimum confidence threshold to filter out the strong association rules that abnormal broadcast topics and accountability organizations. Fig. 3Anomaly broadcast and organization association mining algorithm. Anomaly broadcast and organization association mining algorithm. As shown in Fig. 3, There are a portion of the M abnormal broadcast records are related to the organization \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}$$\end{document} in the database; there are U abnormal broadcast records are topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document}; and X abnormal broadcast records out of the U records whose topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} are related to the organization \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}$$\end{document}. The Apriori algorithm in this section analyses the relationship between topics and organizations of abnormal broadcasts in the database, and the rules for Apriori to obtain the final frequent item set is shown in Eq. (9):9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{anomaly}}\;{\mathrm{broadcast}}\;{\mathrm{topic}}\:z_{k} \to {\mathrm{organization}}\:g_{q} \left[ {{\mathrm{support}} = a\% ,\:{\mathrm{confidence}} = b\% } \right]$$\end{document}In the above equation, it denotes the anomaly broadcast topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} and organization \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}$$\end{document} association rule support as a and confidence as b, where a and b are two values in the range of 0 to 100; where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:support=a\mathrm{\%}$$\end{document} means that a percent of the all anomalies are topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} and they are related to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{o}\mathrm{r}\mathrm{g}\mathrm{a}\mathrm{n}\mathrm{i}\mathrm{z}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}$$\end{document}; where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:confidence=b\mathrm{\%}$$\end{document} means that b percent of anomalies with subject \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} are related to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{o}\mathrm{r}\mathrm{g}\mathrm{a}\mathrm{n}\mathrm{i}\mathrm{z}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}$$\end{document}. The support and confidence of the abnormal broadcast topics \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} association with the organizations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}$$\end{document} are calculated by the following Eq. (10) and Eq. (11):10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:support=\frac{X}{M}\times\:100\mathrm{\%}$$\end{document}11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:confidence=\frac{X}{T}\times\:100\mathrm{\%}$$\end{document} The minimum support threshold and the minimum confidence threshold are set according to experiments. If the support and confidence of the abnormal broadcast topics \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} association with the organizations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}$$\end{document} are greater than thresholds respectively, we determine that the topics \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} is associated with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}$$\end{document}. By collecting all the association rule of abnormal broadcast topics and the accountability organizations, we obtain the accountability organization association rule base G. Semantics-assisted abnormal broadcasts analysis algorithm According to the semantic topic from Sect. 3.2 and the accountability organization association rules obtained from Sect. 3.3, this section detects abnormal radios and traces their accountability organizations without relying on subjective rules. Algorithm 2SAAM SAAM The priori confirmed abnormal broadcast records is preprocessed as Sect. 3.1 to form a corpus. After that we make a topic analysis of this corpus according to Sect. 3.2. Then the distribution of abnormal broadcast topic-word \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\overrightarrow{\beta\:}}_{k}}^{{\prime\:}}$$\end{document} is got from Eq. (8), and the keywords with high probability are taken as the rules for identification the abnormal broadcast topic. The distribution of broadcast-topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\overrightarrow{\theta\:}}_{d}}^{{\prime\:}}$$\end{document} is got from Eq. (7), and the topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} of each abnormal broadcast can be inferred from the document-topic probability. Next, we leverage topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} and the priori organization \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}$$\end{document} to explore the association between topic and accountability organization according to Sect. 3.3. The accountability organization association rule library is constructed. The broadcasts to be detected audios are preprocessed. Firstly, it is converted into text by speech recognition. Then the preprocessed broadcast texts are assumed to be the corpus of LDA. By the LDA semantic model in Sect. 3.2, the topic-word distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\beta\:}}_{k}$$\end{document} and broadcast-topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\theta\:}}_{d}$$\end{document} are estimated according to Eq. (8) and Eq. (7) respectively. If the largest element in the vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\theta\:}}_{d}$$\end{document} is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{dl}$$\end{document}, the topic of the broadcast text d is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{l}$$\end{document} which represented by the corresponding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\beta\:}}_{l}$$\end{document}. The similarity\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\:{S}_{k\:}(\mathrm{w}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{e}\:k=\mathrm{1,2},\dots\:,K\:$$\end{document})of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\beta\:}}_{l}$$\end{document}with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\overrightarrow{\beta\:}}_{k}}^{{\prime\:}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(\mathrm{w}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{e}\:k=\mathrm{1,2},\dots\:,K\:$$\end{document} is express as follow:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{k}=\frac{{\sum\:}_{i=1}^{I}{A}_{ki}{B}_{li}}{\sqrt{{\sum\:}_{i=1}^{I}{{(A}_{ki})}^{2}}\sqrt{{\sum\:}_{i=1}^{I}{{(B}_{li})}^{2}}},$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{A}_{ki}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{B}_{li}$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i=\mathrm{1,2},\dots\:I$$\end{document}) are the top I high probability word according to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{{\overrightarrow{\beta\:}}_{k}}^{{\prime\:}}$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\beta\:}}_{l}$$\end{document} respectively. The similarity threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Q}_{S}$$\end{document} can be obtained by experimental test. The maximum is taken from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{k}(k=\mathrm{1,2},\dots\:,K)$$\end{document}. If the maximum \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{k}$$\end{document} is greater than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Q}_{S}$$\end{document}, this broadcast is considered as an abnormal broadcast and it is marked as topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} corresponding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{k}$$\end{document}. The accountability organization is inferred by matching the marked abnormal broadcast topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} with the organization association rule base G from Sect. 3.3. Further tracing of the abnormal broadcast is now achieved. Finally, detection system outputs alerts with the message of identified topic and associated accountability organizations. The semantic-aided abnormal broadcast content analysis algorithm is summarized in Algorithm 2. Experiment Our experiment consists of three phases. The first stage is the broadcast dataset collection from the field in each city. In the second stage, the detection rules and tracing rules are mined. Finally, in the third stage, the performance of proposed method for abnormal broadcast analysis is presented after comparative experiment. Datasets As far as we know, there is no dataset about abnormal broadcast content. The dataset in this experiment is collected from several cities (Chengdu, Guangyuan, Luoyang and Mianyang) in China and all the collected radio broadcast are in the 87 ~ 108 MHz band. The historical dataset and test dataset are generated from the above collected radio broadcast. The historical dataset is used to emulate a prior abnormal broadcast. The test dataset is used to validate the performance of the proposed SAAM in detecting and tracing abnormal broadcast. Their abnormal contents are fake drugs and hospitals advertisements about eight categories: treatment of stretch marks, medicines for hypertension, medicines for eye diseases, traditional Chinese herb medicines, hospital for andrology, Nanshan Andrology Research Institute, Qiyang medicinal herbs, and ion medicines. The above advertisements are published by 8 accountability organizations respectively. We collate the above data, and convert them into text. Then we annotate their accountability organizations. In total, the historical dataset has 327 broadcast documents, which contains eight categories of abnormal radio broadcast, as shown in Table 1. The content of abnormal broadcasts is highly repetitive day by day and the rush hour for abnormal broadcasts is from 7 p.m. to the late at night every day. There are a few normal broadcasts during this period. we emulate the broadcast data during a rush hour in a day to be the test dataset. The test dataset consists of 448 test broadcast texts, including 244 abnormal broadcast texts from the above 8 categories of topics and 204 normal texts. The test set is randomly divided into five groups of varying numbers. It is shown in Table 2 Table 1The historical dataset.TypesUnit (piece)Treatment of stretch marks13Hypertension Medicine31Medicine for eye disease30Traditional Chinese Herb Medicine14Hospital for Andrology102Nanshan Andrology Research Institute40Qiyang medicinal herbs45Ion Medicines51Total327 Table 1The historical dataset.TypesUnit (piece)Treatment of stretch marks13Hypertension Medicine31Medicine for eye disease30Traditional Chinese Herb Medicine14Hospital for Andrology102Nanshan Andrology Research Institute40Qiyang medicinal herbs45Ion Medicines51Total327 The historical dataset. Table 2Test dataset.Unit (piece)Group IGroup IIGroup IIIGroup IVGroup VAbnormal broadcast data3040505668Normal broadcast data2030405262Total507090108130 Table 2Test dataset.Unit (piece)Group IGroup IIGroup IIIGroup IVGroup VAbnormal broadcast data3040505668Normal broadcast data2030405262Total507090108130 Test dataset. Detection and tracing for anomaly broadcast content we leverage the historical dataset in Sect. 4.1 to mining anomaly detection rules and the responsible organization association rules. The broadcast records in the dataset are all in Chinese. The data set in this experiment, each record is processed according to the Chinese words by slicing and Stop-Words. Detection We discovered anomaly detection rules from the historical dataset according to LDA model described in Sect. 3.2. Table 3Rules library for anomaly broadcast detection.Topic00.031*"Renshen” + 0.024*"Wen” + 0.015*"Pifu” + 0.010*"Qianwei” + 0.010*"Wenyou " + 0.009*"Chang” + 0.008*"Wenlu” + 0.008*"Duzi” + 0.008*"Tanli” + 0.007*"Duanlie”Topic10.016*"Chi” + 0.012*"Zhongyi” + 0.009*"Zhuanjia” + 0.007*"Jiang” + 0.006*"Zixun” + 0.006*"Yao” + 0.006*"Zhuren” + 0.006*"Yunyue” + 0.005*"Xiyao " + 0.005*” Bai Cao Tang “Topic20.017*"Yanjing” + 0.010*"Yanbing” + 0.008*"Shenyi " + 0.007*"Baoming " + 0.007*"Lilaoshi” + 0.006*"Jintian” + 0.005*"Zhenhao” + 0.005*"Shili” + 0.005*"Xiaoguo” + 0.005*"Liaofa”Topic30.036*"Gaoxieya” + 0.028*"Jiangya” + 0.018*"Xieya” + 0.017*"Liaocheng” + 0.015*"Chanpin” + 0.013*"Gaoxieyabing” + 0.013*"Shiyong” + 0.013*"Shuli” + 0.011*"Yaowu” + 0.010*"Zhong”Topic40.006*"Qianliexian” + 0.005*"Nuren” + 0.005*"Genben” + 0.003*"Yingdu” +0.003*"Nanren” + 0.003*"Yiyuan” + 0.003*"Yici” + 0.002*"Yihou” + 0.002*” " + 0.002*"Niao”Topic50.017* “Liaocheng” + 0.009* “Qianliexian” + 0.008* “Nan” + 0.008* “Guoyao” + 0.007* “Niao” + 0.006* “Huodong” + 0.005* “Mai” + 0.005* “Qichenzhan” + 0.005* “Nenggou " + 0.005* “Chiyao”Topic60.013*"Zhuanjia " + 0.010*"Yuyue” + 0.009*"Lu” + 0.007*"Youhui” + 0.007*"Miaoniao” + 0.007*” Zaoxie” + 0.006*"Yinjing " + 0.006*"Shengzhi " + 0.006*"Yiyuan " + 0.006*"zixun”Topic70.008*"Yuyue” + 0.008*"Manren” + 0.008*"Lu” + 0.007*"Zhuanjia” + 0.006*"Shengzhi” + 0.006*"Jishu” + 0.005*"Mianyangshi” + 0.005*"Yanjiuyuan " + 0.005*"Youhui” + 0.005*"Zhuanye” Table 3Rules library for anomaly broadcast detection.Topic00.031*"Renshen” + 0.024*"Wen” + 0.015*"Pifu” + 0.010*"Qianwei” + 0.010*"Wenyou " + 0.009*"Chang” + 0.008*"Wenlu” + 0.008*"Duzi” + 0.008*"Tanli” + 0.007*"Duanlie”Topic10.016*"Chi” + 0.012*"Zhongyi” + 0.009*"Zhuanjia” + 0.007*"Jiang” + 0.006*"Zixun” + 0.006*"Yao” + 0.006*"Zhuren” + 0.006*"Yunyue” + 0.005*"Xiyao " + 0.005*” Bai Cao Tang “Topic20.017*"Yanjing” + 0.010*"Yanbing” + 0.008*"Shenyi " + 0.007*"Baoming " + 0.007*"Lilaoshi” + 0.006*"Jintian” + 0.005*"Zhenhao” + 0.005*"Shili” + 0.005*"Xiaoguo” + 0.005*"Liaofa”Topic30.036*"Gaoxieya” + 0.028*"Jiangya” + 0.018*"Xieya” + 0.017*"Liaocheng” + 0.015*"Chanpin” + 0.013*"Gaoxieyabing” + 0.013*"Shiyong” + 0.013*"Shuli” + 0.011*"Yaowu” + 0.010*"Zhong”Topic40.006*"Qianliexian” + 0.005*"Nuren” + 0.005*"Genben” + 0.003*"Yingdu” +0.003*"Nanren” + 0.003*"Yiyuan” + 0.003*"Yici” + 0.002*"Yihou” + 0.002*” " + 0.002*"Niao”Topic50.017* “Liaocheng” + 0.009* “Qianliexian” + 0.008* “Nan” + 0.008* “Guoyao” + 0.007* “Niao” + 0.006* “Huodong” + 0.005* “Mai” + 0.005* “Qichenzhan” + 0.005* “Nenggou " + 0.005* “Chiyao”Topic60.013*"Zhuanjia " + 0.010*"Yuyue” + 0.009*"Lu” + 0.007*"Youhui” + 0.007*"Miaoniao” + 0.007*” Zaoxie” + 0.006*"Yinjing " + 0.006*"Shengzhi " + 0.006*"Yiyuan " + 0.006*"zixun”Topic70.008*"Yuyue” + 0.008*"Manren” + 0.008*"Lu” + 0.007*"Zhuanjia” + 0.006*"Shengzhi” + 0.006*"Jishu” + 0.005*"Mianyangshi” + 0.005*"Yanjiuyuan " + 0.005*"Youhui” + 0.005*"Zhuanye” Rules library for anomaly broadcast detection. The parameters in LDA estimation are configured with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=6.25,\beta\:=0.01,k=8$$\end{document}. After LDA analysis, the abnormal broadcast topic-word distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\overrightarrow{\beta\:}}_{k}}^{{\prime\:}}$$\end{document}and broadcast-topic distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\overrightarrow{\theta\:}}_{d}}^{{\prime\:}}$$\end{document}are obtained. According to distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\overrightarrow{\beta\:}}_{k}}^{{\prime\:}}$$\end{document}, we sort the words by the descending order of the corresponding probability and the top ten words in each topic are taken as the anomaly rules \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{A}_{ki}$$\end{document}. Then we get the detection rule library as shown in Table 3, where Topic0 ~ Topic7 represent eight broadcast topics, the word is in Chinese spell, and the decimal represents the probability of the word appearing in each topic. The broadcasts in the test dataset are analyzed according to Algorithm 2 in Sect. 3.4. The topic-word distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\beta\:}}_{k}$$\end{document} and broadcast-topic distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\theta\:}}_{d}$$\end{document} are inferred. Corresponding to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\theta\:}}_{d}$$\end{document}, we take the topic with the greatest probability in each broadcast record to be its topic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{l}$$\end{document}. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\overrightarrow{\beta\:}}_{l}$$\end{document} corresponding to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{l}$$\end{document} is the topic description of the broadcast and the similarity\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\:{S}_{k\:}$$\end{document}is calculated according to Eq. (12). The threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Q}_{S}$$\end{document} is set by the test experiments. The detection performance with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Q}_{S}$$\end{document} = 0.3, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Q}_{S}$$\end{document} = 0.4 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Q}_{S}$$\end{document} = 0.5 are shown in Table 4. F score is a combine indicator of detection Precision and Recall. When \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Q}_{S}$$\end{document} = 0.4, the F score of each test group is about 0.93. Therefore, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Q}_{S}$$\end{document} is set to be 0.4 in this experiment. If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{k\:}$$\end{document} is greater than 0.4, this broadcast is discriminate as anomaly, and the anomaly topic is marked as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document}. Table 4Experimental results for different matching thresholds S. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Q}_{S}=0.3/{Q}_{S}=0.4/{Q}_{S}=0.5$$\end{document} AccuracyPrecisionRecallF110.880/0.98/0.920.875/0.952/10.933/0.967/0.8670.903/0.983/0.86720.885/0.928/0.8710.900/0.857/10.9/0.875/0.7750.900/0.933/0.77530.911/0.911/0.8780.889/0.977/10.96/0.86/0.7800.923/0.914/0.78040.889/0.911/0.9070.879/0.977/10.911/0.86/0.8210.895/0.914/0.82150.907/0.938/0.8850.902/0.954/10.928/0.913/0.7790.915/0.933/0.779 Table 4Experimental results for different matching thresholds S. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Q}_{S}=0.3/{Q}_{S}=0.4/{Q}_{S}=0.5$$\end{document} AccuracyPrecisionRecallF110.880/0.98/0.920.875/0.952/10.933/0.967/0.8670.903/0.983/0.86720.885/0.928/0.8710.900/0.857/10.9/0.875/0.7750.900/0.933/0.77530.911/0.911/0.8780.889/0.977/10.96/0.86/0.7800.923/0.914/0.78040.889/0.911/0.9070.879/0.977/10.911/0.86/0.8210.895/0.914/0.82150.907/0.938/0.8850.902/0.954/10.928/0.913/0.7790.915/0.933/0.779 Experimental results for different matching thresholds S. Accountability organization tracing After the abnormal broadcast is detected and its topic determined, the association analysis is performed according to Sect. 3.3. The parameters in Apriori algorithm are configured with the minimum support threshold and the minimum confidence threshold are 0.03 and 0.2, respectively. As a result, a total of 8 abnormal broadcast topic-organization association rules are mined as shown in Table 5. They are stored as the organization association rule base. From the abnormal broadcast topic-organization association rules we can trace the abnormal broadcast with the organization relationships. QS=0.4, the anomaly broadcasts are detected and marked with topics in Sect. 4.2.1. Each broadcast’s topic is marching with the topics in Tables 5 to infer the accountability organization. The organization tracing results of the five test data groups are show in Table 6. Table 5Abnormal broadcast organization association rule library. 1{‘topic0’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Treatment of Stretch Marks’ }\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=3.96\:\mathrm{\%},\:confidence=68.42\:\mathrm{\%}\right]$$\end{document}2{‘topic1’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Traditional Chinese Herb Medicine ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=4.5732\:\mathrm{\%},\:confidence=93.33\:\mathrm{\%}\right]$$\end{document}3{‘topic2’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Medicine for Eye Disease ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=6.70\:\mathrm{\%},\:confidence=88.00\:\mathrm{\%}\right]$$\end{document}4{‘topic3’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Hypertension Medicine ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=4.57\:\mathrm{\%},\:confidence=93.54\mathrm{\%}\right]$$\end{document}5{‘topic4’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Nanshan Andrology Research Institute ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=11.89\:\mathrm{\%},\:confidence=97.05\:\mathrm{\%}\right]$$\end{document}6{‘topic5’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Qiyang Medicinal Herbs ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=13.41\:\mathrm{\%},\:confidence=73.33\:\mathrm{\%}\right]$$\end{document}7{‘topic6’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Ion Medicines ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=14.93\:\mathrm{\%},\:confidence=96.07\:\mathrm{\%}\right]$$\end{document}8{‘topic7’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Hospital for Andrology ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=28.35\:\mathrm{\%},\:confidence=98.80\:\mathrm{\%}\right]$$\end{document} Table 5Abnormal broadcast organization association rule library. 1{‘topic0’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Treatment of Stretch Marks’ }\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=3.96\:\mathrm{\%},\:confidence=68.42\:\mathrm{\%}\right]$$\end{document}2{‘topic1’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Traditional Chinese Herb Medicine ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=4.5732\:\mathrm{\%},\:confidence=93.33\:\mathrm{\%}\right]$$\end{document}3{‘topic2’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Medicine for Eye Disease ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=6.70\:\mathrm{\%},\:confidence=88.00\:\mathrm{\%}\right]$$\end{document}4{‘topic3’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Hypertension Medicine ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=4.57\:\mathrm{\%},\:confidence=93.54\mathrm{\%}\right]$$\end{document}5{‘topic4’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Nanshan Andrology Research Institute ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=11.89\:\mathrm{\%},\:confidence=97.05\:\mathrm{\%}\right]$$\end{document}6{‘topic5’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Qiyang Medicinal Herbs ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=13.41\:\mathrm{\%},\:confidence=73.33\:\mathrm{\%}\right]$$\end{document}7{‘topic6’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Ion Medicines ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=14.93\:\mathrm{\%},\:confidence=96.07\:\mathrm{\%}\right]$$\end{document}8{‘topic7’}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document}{‘ Hospital for Andrology ‘}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[support=28.35\:\mathrm{\%},\:confidence=98.80\:\mathrm{\%}\right]$$\end{document} Abnormal broadcast organization association rule library. Table 6Tracing-performance of accountability organizations for abnormal broadcasts.Group 1Group 2Group 3Group 4Group 5Tracing-Accuracy0.8670.8500.8200.8400.823 Table 6Tracing-performance of accountability organizations for abnormal broadcasts.Group 1Group 2Group 3Group 4Group 5Tracing-Accuracy0.8670.8500.8200.8400.823 Tracing-performance of accountability organizations for abnormal broadcasts. Evaluation in different size datasets In this section, four datasets are emulated to further evaluate SAAM across varied broadcast scenarios. The four datasets are generated to emulate broadcast data about one week at some certain location. During the on-site collection of the dataset, we found that the content of abnormal broadcasts was highly repetitive day by day. So, we randomly selected abnormal broadcasts from 327 historical data and injected them into normal broadcasts to generate four larger-scale datasets with 3000, 5000, 8000 texts. And the experiment sets different concentrations of anomalies with 25%, 10%, 5%, 1%. Table 7Performance of SAAM in different datasets.Dataset sizeConcentration of anomalyDetection-accuracyDetection-precisionDetection-recallTracing-accuracy300025%0.9400.8700.9100.810300010%0.9540.7010.9460.80250005%0.9580.5350.9440.82680001%0.9520.1670.9370.853 Table 7Performance of SAAM in different datasets.Dataset sizeConcentration of anomalyDetection-accuracyDetection-precisionDetection-recallTracing-accuracy300025%0.9400.8700.9100.810300010%0.9540.7010.9460.80250005%0.9580.5350.9440.82680001%0.9520.1670.9370.853 Performance of SAAM in different datasets. The detection results and tracing abilities are shown in Table 7. The detection accuracy rate and recall rate are all about 0.95. The ability to trace the organization can be above 0.8. However, the detection precision decreases as the concentration of anomaly decreases. When the concentration is 1%, SAAM still recalls 94.4% abnormal broadcast, and the detection precision greatly decreases to 0.16. The experimental results show that, for all the four broadcast scenarios, SAAM can detected anomalies and trace organizations. However, as the proportion of abnormal broadcasts decreases, the false positives increase. Runtime overhead We evaluated the overhead of LDA and Apriori algorithm in SAAM. The usages and runtime are listed in Tables 8 and 9. With the dataset having 327, 3270, 32,700 and 327,000 broadcast texts, we set up the experiments on a host with an AMD Ryzen 75,800 H having 8 cores and 3.2 GHz of speed each and the memory of 16GB as well as running on Windows 10. The Gibbs sampling method is adopted to iterate the parameters of the LDA model. The computation complexity of the LDA algorithm is related to the total number of words in the dataset, the number of topics K and the number of iterations. In this work, the number of iterations is 1000 and K=2. Therefore, in this experiment, the LDA time complexity depends on the size of the dataset. The larger the dataset, the greater the total number of words. In Table 8, the runtime increases as the size of the dataset grows. The computation complexity of the Apriori is related to the number of dataset records, the number of items in each record, the kinds of items in the dataset. In the accountability organization association model, there are 8 topics and 8 organizations, so the number of items in each record is 2 and 16 kinds of items in the dataset. Therefore, in this experiment, the Apriori time complexity depends on the size of the dataset. As the dataset gets larger, the runtime becomes longer. In the broadcast association with only 2 items of each record and just 16 kinds of items in the dataset, so the runtime of mining association rules is a few seconds, as shown in Table 9. The max memory usage is 2215.91 MB and the max CPU utilization is 103.39% about one of the 8 cores. The results show that, LDA and Apriori can run with affordable overhead on common hardware equipment.Table 8Usage and runtime for LDA. Dataset sizeDuration (s)Avg. memory MBMax memory (MB)CPU utilization (%)3271093.46303.11305.8103.393,2702569.05478.12478.99100.9632,70011882.371327.182215.91100.01 Table 8Usage and runtime for LDA. Dataset sizeDuration (s)Avg. memory MBMax memory (MB)CPU utilization (%)3271093.46303.11305.8103.393,2702569.05478.12478.99100.9632,70011882.371327.182215.91100.01 Usage and runtime for LDA. Table 9Usage and runtime for Apriori.Dataset sizeDuration (s)Avg. memory (MB)Max memory (MB)CPU utilization (%)3270.015153.82153.880.003,2700.0256154.66155.43.3883.84327,0004.199312.56385.18100.71 Table 9Usage and runtime for Apriori.Dataset sizeDuration (s)Avg. memory (MB)Max memory (MB)CPU utilization (%)3270.015153.82153.880.003,2700.0256154.66155.43.3883.84327,0004.199312.56385.18100.71 Usage and runtime for Apriori. Comparative experiment This section shows the performance improvement of our proposed detection method by comparing it with three benchmark approaches: Deterministic Finite Automaton (DFA) algorithm43, the Convolutional Neural Networks (CNN) algorithm44 and Sentence-BERT (SBERT)45. DFA is a commonly-used keyword matching method. It identifies anomalies by matching the segmented broadcast data with a preset sensitive lexicon, and determines whether the broadcast is abnormal based on the identification result. The CNN algorithm is commonly used in text categorization, by transforming text data into word embedding representation which is an input to the convolutional layer for feature extraction and classification. Jarrah MAM44 used CNN algorithm to detect hate speech in tweets, and the CNN algorithm is used to identify abnormal broadcasts in this section. SBERT is a model based on the Transformer for state-of-the art sentence embeddings in tasks including large-scale semantic similarity comparison, clustering and information retrieval via semantic search46. This section uses DFA, CNN, SBERT and SAAM to detect abnormal broadcasts in the test dataset Group V listed in Table 2. In the DFA experiments, the sensitive lexicon contains 2084 anomalies downloaded from the online public dataset47, which covers most of the anomalies currently in use. In the SAAM experiment, the detection rules are taken from Sect. 4.2.1. In the CNN experiment, after labelling the abnormal or normal broadcasts and preprocessing them, and the models are trained. And the CNN detect the abnormal broadcasts among them after the tested broadcast data are fed into the model. In the SBERT experiment, the historical dataset with 327 anomalies is fed to SBERT model after preprocessing. SBERT generates abnormal semantic vector as the abnormal rule. The tested data are embedded into vector by SBERT to calculate similarity with the rule. In this section, for the sake of analysis and comparison, we count the detection results of SAAM and the three benchmark approaches. The detection results statistics are shown in Table 10. Table 10The performance of the methods.MethodAccuracyPrecisionRecallF1-scoreSAAM0.9380.9540.9120.933DFA0.8150.8920.7350.806CNN0.9230.9670.8820.923SBERT0.8460.9140.7790.841 Table 10The performance of the methods.MethodAccuracyPrecisionRecallF1-scoreSAAM0.9380.9540.9120.933DFA0.8150.8920.7350.806CNN0.9230.9670.8820.923SBERT0.8460.9140.7790.841 The performance of the methods. We can see from Table 10, that all three methods are able to identify most abnormal broadcasts. However, all the four indicators of the DFA algorithm are lower than that of the three machine learning methods SAAM, CNN, and SBERT. It is due to the subjective keyword dose not adapt well to the changing data. In this experiment, SBERT extracts the abnormal features of mixed eight anomalies, and the identified rules are relatively broad. So that, detection results of SBERT are weaker than that of SAAM and CNN. CNN is training by the labeled data and learn the features of abnormal broadcast data. Due to limited training data the CNN model is slightly insufficient to detection new data, which results in a slightly lower accuracy rate and recall rate compared to the SAAM. But, in terms of precision rate, CNN is slightly higher than SAAM. It implies that SAAM can detect more abnormal broadcasts, but it detects more false positives than CNN. The F1-score can integrate the overall performance of both recall rate and precision rate. In the terms of F1-score, SAAM is better than CNN. Although CNN and SAAM have similar performance, the SAAM method is available for further tracing the accountability organization. The SAAM method can deeply analyze the abnormal broadcasts, while CNN and SBET can only achieve the black-box analysis process. The above results shows that the better performance of the SAAM for abnormal broadcast content recognition. Conclusions In this paper, we developed a new semantic-aided content analysis system SAAM for the detection problem of the abnormal broadcasts. SAAM considered the semantic context to carry out a semantic correlation analysis approach for broadcasts content to detect abnormal broadcasts and trace back to its organization. The semantic analysis relying on LDA to self-discover the topics and the key words related to anomaly, thus generating the abnormal identification rules from the historical broadcast text intelligently. Then, the topics from LDA was employed to assist Apriori algorithm to capture latent associations rules with accountability organizations for the abnormal broadcast text. When detecting the broadcast, SAAM performed LDA analysis on test broadcasts, then detected anomalies and traced them according to the above rules. An experiment has been conducted to demonstrate the achievable performance of our proposed method (SAAM). By comparing it with the subjective rule method (DFA) and machine learning methods (CNN, SBERT), we illustrated the improved performance of the proposed SAAM in both identifying the anomalies as well as tracing each accountability organization.