viterbi算法 python版

牛mm细心给我讲了一个小时，终于明白它的含义，然后花了一两节分布式数据库的课实现了。当时牛mm还说不可能这么快实现，结果不可能事还是发生了。发现python果真非常好用。不明白此算法可以看这篇blog http://blog.csdn.net/NirvanaFeng/archive/2009/05/12/4171799.aspx

初始化方法：

def InitDicForViterbi(nodes,posw,posdi,n): newWordList = [] # 解决未登录词 for i in nodes: if not posw.has_key(i): newWordList.append(i) maxPosList = NPos(posdi,n) print maxPosList GenerateDicPos(newWordList,maxPosList,posw,posdi) def InitViterbi(node,posw,posdi): viStatePath = [] for i in posw[node]: if i <> "@@@": print "i:",i viStatePath.append([posw[node][i]*posdi[i]["@@@"],[i]]) return viStatePath

viterbi算法函数

""" nodes 就是分好的词, posw 是词转换为词性的概率，posdi是词性之间的转换概率,n 是n个最大的词性将此用于未登录词中, weightNone是未出现的词性转移的概率 nodes format {word1,word2...} , posw format {word1:{pos1:fre,pos2:fre,..."@@@":totalnum},..."@@@":total} posdi format {pos1:{pos2:fre,pos3:fre...."@@@":total},...."@@@":total} """ def Viterbi(nodes,posw,posdi,n,weightNone): InitDicForViterbi(nodes,posw,posdi,n) viStatePath = InitViterbi(nodes[0],posw,posdi) length = len(nodes) currentNode = 1 while currentNode < length: currentPosList = posw[nodes[currentNode]] paths = [] # print "vstate:",viStatePath for k in currentPosList: if k <> "@@@": ajk = weightNone heap = [] for j in xrange(len(viStatePath)): # compute every state j to every state k in ti # temppath = viStatePath[j] # print "lastpos:",temppath lastpos = viStatePath[j][1][-1] lastweight = viStatePath[j][0] lastposList = posdi[lastpos] if lastposList.has_key(k): ajk =lastposList[k] currentweight = lastweight * ajk * currentPosList[k] # print "viStatePath:",viStatePath[j][1] pathNew = [data for data in viStatePath[j][1]] pathNew.append(k) # print "pathNew:",pathNew heappush(heap,[currentweight,pathNew]) # get the max possibility of state k in ti # print "path:",path paths.append(nlargest(1,heap)[0]) del viStatePath viStatePath = paths # print "paths:",paths currentNode = currentNode + 1 heap = [] # get the max possibility path for i in viStatePath: heappush(heap,i) return nlargest(1,heap)

结果打印输出函数：

"""nodes format {word1,word2,...} path is [weight,[pos1,pos2....]]""" def Result(nodes,path,edcode="utf-8"): realPath = path[0][1] ResultPrint(nodes,realPath,edcode) """nodes format {word1,word2,...} path is [pos1,pos2....]""" def ResultPrint(nodes,path,edcode="utf-8"): for i in xrange(len(nodes)): print nodes[i].decode(edcode),"/",path[i].decode(edcode)

下面是测试代码：

aa = ConvertGBKtoUTF("球球") bb = ConvertGBKtoUTF("娃娃") cc = ConvertGBKtoUTF("吃饭") dd = ConvertGBKtoUTF("好") ee = ConvertGBKtoUTF("dddwieoewkem") dictions = {aa:{bb:1,"@@@":4},bb:{cc:2,aa:3,"@@@":40},"@@@":400} posdi = {"n":{"s":3,"v":3,"@@@":40},"s":{"v":2,"e":3,"@@@":33},"v":{"@@@":1},"@@@":100} posw = {aa:{"n":1,"v":3,"@@@":29},bb:{"n":1,"s":1,"@@@":19},cc:{"n":1,"@@@":1},"@@@":10002} nodes = [aa,bb,cc,ee,dd,bb,aa,cc] path = Viterbi(nodes,posw,posdi,3,0.01) print path Result(nodes,path)

这里不要问为啥要encode,然后再decode 因为只有这样才能在屏幕上打印出中文。