C# 屏蔽词过滤

参考:https://www.cnblogs.com/kubidemanong/p/10834993.html

public class TreeNode
    {
        public char Char;
        public bool IsEnd;
        public int WordEndAt;
        private Dictionary<char, TreeNode> NodeDict;
        public TreeNode(char c)
        {
            Char = c;
            IsEnd = false;
            NodeDict = new Dictionary<char, TreeNode>();
        }

        public bool ContainChar(char ch)
        {
            return NodeDict.ContainsKey(ch);
        }

        public TreeNode GetChild(char c)
        {
            TreeNode Child = null;
            NodeDict.TryGetValue(c, out Child);
            return Child;
        }

        public TreeNode AddNode(char ch)
        {
            TreeNode insertNode = null;
            if (!NodeDict.TryGetValue(ch, out insertNode))
            {
                insertNode = new TreeNode(ch);
                NodeDict.Add(ch, insertNode);
            }
            return insertNode;
        }
    }
public class Trie
{

        public TreeNode Root { get; }
        private HashSet<char> SkipCharSet;      //忽略字符
        private HashSet<char> SeparateCharSet;  //常见分隔符
        private string skipCharList = " `-=[]\\‘,.·/~!@#$%^&*()_+{}|:\"<>?*\r\n";
        private TreeNode checkNode;
        private TreeNode backupNode;
        private char duplicateChar;
        public Trie()
        {
            Root = new TreeNode( );
            SkipCharSet = new HashSet<char>();
            SeparateCharSet = new HashSet<char>();
            foreach (char c in skipCharList)
            {
                SkipCharSet.Add(c);
                SeparateCharSet.Add(c);
            }
        }

        public void AddNode(TreeNode node, string word)
        {
            if (word.Length > 0)
            {
                char ch = word[0];
                TreeNode insertNode = node.AddNode(ch);
                if (word.Length == 1)
                {
                    insertNode.IsEnd = true;
                }
                word = word.Remove(0,1);
                AddNode(insertNode, word);
            }
        }

        private bool IsSkipChar(char c)
        {
            return SkipCharSet.Contains(c);
        }

        //是否是英文
        private bool IsEnglishChar(char ch)
        {
            string str = new string(ch, 1);
            Regex regEnglish = new Regex("^[a-zA-Z]");
            return regEnglish.IsMatch(str);
        }

        //是否是分割符
        private bool Separator(char ch)
        {
            return SeparateCharSet.Contains(ch);
        }

        //是否是单词的开头
        private bool IsWordBegin(string word,int pos)
        {
            if (pos == 0)
                return true;
            if (pos < word.Length)
            {
                char c1 = word[pos - 1];
                char c2 = word[pos];
                return (Separator(c1) && !Separator(c2)) || (!IsEnglishChar(c1) && IsEnglishChar(c2));
            }
            return false;
        }

        //是否是单词的结尾
        private bool IsWordEnd(string word,int pos)
        {
            if (pos == word.Length - 1)
                return true;
            if(pos < word.Length)
            {
                char c1 = word[pos];
                char c2 = word[pos + 1];
                return (!Separator(c1) && Separator(c2)) || (IsEnglishChar(c1) && !IsEnglishChar(c2));
            }
            return false;
        }

        private void CheckWord(string checkWord,int begin)
        {
            int index = begin;
            while(index + 1 < checkWord.Length)
            {
                ++index;
                char ch = checkWord[index];
                if (IsSkipChar(ch))
                {
                    if (checkNode.ContainChar(ch))
                        checkNode = checkNode.GetChild(ch);
                }
                else
                {
                    if(checkNode.ContainChar(ch))
                    {
                        checkNode = checkNode.GetChild(ch);
                        if(checkNode.IsEnd)
                        {
                            checkNode.WordEndAt = index;
                            backupNode = checkNode;
                            duplicateChar = ch;             
                            CheckWord(checkWord,index);     //继续匹配
                            break;
                        }
                    }
                    else
                    {
                        if (duplicateChar == ch)            //屏蔽fuccccccck例如这样的
                            backupNode.WordEndAt = index;
                        else
                            break;
                    }
                }
                duplicateChar = ch;
            }
        }

        public string Filter(string filterWord)
        {
            int begin = 0;
            checkNode = null;
            backupNode = null;
            duplicateChar =  ;
            string word = filterWord.ToLower();
            StringBuilder result = new StringBuilder(filterWord);
            while(begin < word.Length)
            {
                checkNode = Root;
                backupNode = Root;
                char ch = word[begin];
                duplicateChar = ch;
               //设置是否严格匹配, 即逐个字符检测是否有可能是敏感词,否则像英语一样只检测单词, 
//
严格匹配: [av]是敏感词,那么[avoid]被替换成[**oid]
                bool isStrict = !IsEnglishChar(ch);              
                bool isWordBegin = isStrict || IsWordBegin(word, begin);
                if(isWordBegin && checkNode.ContainChar(ch))
                {
                    checkNode = checkNode.GetChild(ch);
                    if(!IsSkipChar(ch))
                    {
                        CheckWord(word, begin);
                        if(backupNode.IsEnd && backupNode.WordEndAt > 0)
                        {
                            bool isWordEnd = isStrict || IsWordEnd(word, backupNode.WordEndAt); //到单词末尾才行 have 中有av 但是不是末尾 所以不是屏蔽词
                            if(isWordEnd)
                            {
                                for(int i = begin; i <= backupNode.WordEndAt;++i)
                                {
                                    result[i] = *;
                                }
                                begin = backupNode.WordEndAt;
                            }
                        }
                    }
                }
                ++begin;
            }

            return result.ToString();
        }
 }

测试用例:

 class Program
    {
        static void Main(string[] args)
        {
            Trie trie = new Trie();
            trie.AddNode(trie.Root, "fuc");
            trie.AddNode(trie.Root, "fuc bitch");
            trie.AddNode(trie.Root, "fuck");
            trie.AddNode(trie.Root, "bitch");
            trie.AddNode(trie.Root, "*");
            Console.WriteLine(trie.Filter("是哦fuckkkkkk山大的撒bi\tch"));
            Console.WriteLine(trie.Filter("have world fuc bitch"));
            Console.WriteLine(trie.Filter("1218fuck1"));

            Console.ReadKey();
        }
}

结果:

C# 屏蔽词过滤

 

C# 屏蔽词过滤

上一篇:Dubbo源码解析-Consumer发送请求全过程


下一篇:一张图揭示主流互联网架构体系,什么会导致Java应用程序的CPU使用率飙升