Sensitive word auditing method

Abstract

The invention discloses a sensitive word auditing method, which belongs to the field of auditing, and comprises the following steps of: firstly, preprocessing an input text, such as removing spaces, symbols, stop words and the like, so as to generate a word segmentation result; thirdly, constructing a dynamic synonym and a deformation word library, combining user behaviors and NLP technology updating, and expanding word meanings by applying a word vector model Word2Vec; then, based on BERT, a text context is analyzed, and metaphors and biguan languages are detected. Performing sensitive word primary screening and secondary screening by using a static word bank and an SVM (Support Vector Machine), and analyzing complex semantics by LSTM (Long Short Term Memory); the model is optimized based on a feedback self-learning mechanism, difficult contents are manually checked, an analysis report is generated, and finally a word bank and an algorithm are periodically updated to adapt to hotspots. The method has the beneficial effects that sensitive word auditing is improved through a self-learning mechanism, synonyms and deformed word banks can be dynamically updated, and private expressions are identified. BERT and LSTM model multi-level auditing are adopted, secondary screening is carried out in combination with SVM, and recognition of complex contexts and biguan languages is improved.

Claims

1. A method for reviewing sensitive words, characterized in that: The method for reviewing sensitive words comprises the following steps: S1. Preprocess the input text, including removing redundant spaces, special characters and emoticons; for non-text content, perform word segmentation to generate separate words and phrases; use word segmentation tools to segment the text and remove stop words and irrelevant characters; S2. Build a dynamic synonym and variant word library to deal with users' behavior of circumventing sensitive word detection through pinyin, homophones and variants; update the synonym and variant word library by collecting Internet data and user behavior analysis, combined with natural language processing technology; use the word vector model Word2Vec to expand word meanings and capture potential variants and synonymous expressions; S3. Use a deep learning method based on the BERT pre-trained language model to analyze the context of the text and determine whether there are sensitive metaphors and puns; S4. Perform preliminary screening through a static sensitive word database to quickly identify obvious sensitive words; apply traditional machine learning methods such as support vector machines to perform secondary screening of potential sensitive content; use deep neural network LSTM to further analyze the semantics of the text and identify complex sensitive content; S5. Adopt a feedback-based self-learning mechanism to optimize the audit model; retrain and optimize the model through manual audit results and user reporting data; use reinforcement learning technology to continuously correct model judgments; S6. For complex and ambiguous texts, the system automatically marks them and submits them to manual review. Combined with the manual review platform, it provides detailed text analysis reports to assist reviewers in making decisions. Through rule setting, difficult texts are processed first. S7. Monitor social hot spots in real time by regularly updating the vocabulary and algorithms.

2. According to claim 1, a method for reviewing sensitive words is characterized in that: The construction of the S2 dynamic synonym and variant word database includes the following steps: S2-1-1. Analyze user behavior, especially how users use inflected words and pinyin to bypass sensitive word detection when violating rules; S2-1-2. Use the word vector model Word2Vec to automatically mine synonyms and near-synonyms from a large amount of text; the Word2Vec is based on a large-scale corpus and captures the semantic similarity between words through the contextual relationship between words; For the recognition of pinyin, homophones and variant words, a rule-based vocabulary is constructed, which includes all pinyin variants, morphological changes and spelling errors. In combination with pinyin matching technology, Chinese characters are converted into pinyin and then compared to identify the situations where sensitive words are avoided by pinyin and near-phonetic characters. S2-1-3. Regularly analyze new sensitive word deformation and avoidance patterns in user-generated content, capture new synonyms and deformation words and incorporate them into the vocabulary; use cluster analysis to automatically identify and classify deformation words; mark deformation words that appear frequently and are potentially dangerous, and associate them with the original sensitive words.

3. According to claim 2, a method for reviewing sensitive words is characterized in that: The cluster analysis to automatically identify and classify inflected words includes the following steps: S2-1-3-1. Initialize K cluster centers, randomly selected and selected according to data distribution; S2-1-3-2. Calculate the distances from all words to the K cluster centers, and assign the words to the corresponding clusters according to the minimum distance; S2-1-3-3. Recalculate the center point of each cluster, that is, the mean vector of all words in the cluster; S2-1-3-4. Repeat S2-1-3-2 and S2-1-3-3 until the clustering results converge and the center point no longer changes; S2-1-3-5. Hierarchical clustering starts with each word, treating it as a separate cluster; S2-1-3-6. Calculate the similarity between clusters and select the two clusters with the smallest similarity to merge; S2-1-3-7. Repeat S2-1-3-6 until all clusters are merged into one large cluster and the preset number of clusters is reached.

4. According to claim 1, a method for reviewing sensitive words is characterized in that: Training the S2 using the word vector model Word2Vec to capture potential deformations and synonymous expressions includes the following steps: S2-2-1. Use a large-scale corpus, including social media data, news text, and forum posts, to train a Word2Vec model; the Word2Vec model captures the semantic similarity between words and identifies potential variant words in context; Using the word vectors generated by the Word2Vec model, similarities between words are calculated to identify words with similar meanings but different expressions, including through synonym substitution and the use of inflected words; S2-2-2. Automatically expand synonyms and near-synonyms by calculating the similarity between word vectors; The processing of inflected words uses the concept of "distance" of word vectors to capture inflected words with similar pinyin and similar glyphs; S2-2-3. Combine rule-based pinyin matching with the semantic expansion of the word vector model; include words with similar pinyin, which are filtered through pinyin rule preprocessing; and for more complex semantic variations, they are further captured through the word vector model.

5. According to claim 1, a method for reviewing sensitive words is characterized in that: The S2 dynamic synonym and variant word database also receives user-reported content including text, comments and messages; The system has a user reporting interface, which allows users to report potential sensitive words that the detection system has not found; the report content includes the original text, the reason for the report, and the specific word information reported; The reported text is segmented to generate a vocabulary, and the reported words are extracted together with the context, and the deformation and synonym relationship are identified.

6. A method for reviewing sensitive words according to claim 5, characterized in that: When the system finds that the content reported by the user contains new inflected words and sensitive words, the new inflected words and sensitive words are updated in real time to the vocabulary of synonyms and inflected words; Adopting an incremental update strategy, whenever the new inflected words and sensitive words are detected, the new inflected words and sensitive words are automatically added to the existing word library, while ensuring that the existing valid words are not overwritten; By means of an automated process, the identified new inflected words and sensitive words are associated with the original sensitive words to generate new synonyms and inflected words; Use machine learning algorithms, including supervised learning-based classification models and unsupervised learning-based clustering models, to automatically classify and associate newly reported inflected words; After the report content is updated, it will go through a manual review mechanism to ensure that all transformed words meet the review standards before being added to the vocabulary library; manual reviewers will check the automated detection results to confirm whether the newly added words are potential sensitive word variations.

7. A method for reviewing sensitive words according to claim 1, characterized in that: Training the S3BERT for metaphor and pun recognition involves the following steps: S3-1. The input text is processed by word segmentation to decompose the sentence into sub-word units; the BERT uses the WordPiece word segmentation algorithm to split the text into smaller sub-word units and process words that have not been seen in the vocabulary; each input sample also adds special tags [CLS] and [SEP]; S3-2. During fine-tuning, each input sample is trained based on its annotated "sensitive" and "non-sensitive" labels, so that the model predicts whether the text contains metaphors and puns; S3-3. Use the cross entropy loss function to adjust the model weights according to the difference between the predicted probability and the actual label; train the cross entropy loss function to minimize; Use the Adam optimizer to adaptively adjust the learning rate; S3-4. Input the token sequence of each sentence and paragraph; the output model outputs a category label, "sensitive" and "non-sensitive", to determine whether the text contains metaphors and puns.

8. A method for reviewing sensitive words according to claim 7, characterized in that: The cross entropy loss function calculation formula is as follows: For the binary classification problem, the calculation formula of the cross entropy loss function is as follows: Wherein, y is the true label, which takes values ​​of 0 and 1; is the probability of category 1 predicted by the model, ranging from For multi-classification problems, the calculation formula of the cross entropy loss function is as follows: Wherein, C is the number of categories; y i is the label value of the i-th class in the one-hot vector of the true label indicator; is the probability that the model predicts that it belongs to category i.

Images