Gensim - 创建LDA Mallet模型

本章将解释什么是潜在狄利克雷分配（LDA）Mallet模型以及如何在Gensim中创建它。

在上一节中，我们已经实现了LDA模型，并从20Newsgroup数据集的文档中获取了主题。那是Gensim内置的LDA算法版本。Gensim也存在Mallet版本，它提供了更好的主题质量。在这里，我们将对之前已经实现的示例应用Mallet的LDA。

什么是LDA Mallet模型？

Mallet是一个开源工具包，由Andrew McCullum编写。它基本上是一个基于Java的软件包，用于NLP、文档分类、聚类、主题建模以及许多其他机器学习文本应用。它为我们提供了Mallet主题建模工具包，其中包含LDA以及分层LDA的高效的基于采样的实现。

Mallet2.0是MALLET（Java主题建模工具包）的当前版本。在开始使用它与Gensim进行LDA之前，我们必须在系统上下载mallet-2.0.8.zip包并解压缩它。安装并解压缩后，将环境变量%MALLET_HOME%设置为指向MALLET目录，可以通过手动或我们将在实现LDA与Mallet时提供的代码来完成。

Gensim包装器

Python为潜在狄利克雷分配（LDA）提供了Gensim包装器。该包装器的语法为gensim.models.wrappers.LdaMallet。此模块（从MALLET折叠吉布斯采样）允许从训练语料库估计LDA模型，以及对新的、未见过的文档推断主题分布。

实现示例

我们将对之前构建的LDA模型使用LDA Mallet，并通过计算连贯性得分来检查性能差异。

提供Mallet文件路径

在将Mallet LDA模型应用于我们在前面示例中构建的语料库之前，我们必须更新环境变量并提供Mallet文件的路径。可以使用以下代码来完成：

import os
from gensim.models.wrappers import LdaMallet
os.environ.update({'MALLET_HOME':r'C:/mallet-2.0.8/'}) 
#You should update this path as per the path of Mallet directory on your system.
mallet_path = r'C:/mallet-2.0.8/bin/mallet' 
#You should update this path as per the path of Mallet directory on your system.

一旦我们提供了Mallet文件的路径，我们现在就可以在语料库上使用它。可以使用ldamallet.show_topics()函数完成，如下所示：

ldamallet = gensim.models.wrappers.LdaMallet(
   mallet_path, corpus=corpus, num_topics=20, id2word=id2word
)
pprint(ldamallet.show_topics(formatted=False))

输出

[
   (4,
   [('gun', 0.024546225966016102),
   ('law', 0.02181426826996709),
   ('state', 0.017633545129043606),
   ('people', 0.017612848479831116),
   ('case', 0.011341763768445888),
   ('crime', 0.010596684396796159),
   ('weapon', 0.00985160502514643),
   ('person', 0.008671896020034356),
   ('firearm', 0.00838214293105946),
   ('police', 0.008257963035784506)]),
   (9,
   [('make', 0.02147966482730431),
   ('people', 0.021377478029838543),
   ('work', 0.018557122419783363),
   ('money', 0.016676885346413244),
   ('year', 0.015982015123646026),
   ('job', 0.012221540976905783),
   ('pay', 0.010239117106069897),
   ('time', 0.008910688739014919),
   ('school', 0.0079092581238504),
   ('support', 0.007357449417535254)]),
   (14,
   [('power', 0.018428398507941996),
   ('line', 0.013784244460364121),
   ('high', 0.01183271164249895),
   ('work', 0.011560979224821522),
   ('ground', 0.010770484918850819),
   ('current', 0.010745781971789235),
   ('wire', 0.008399002000938712),
   ('low', 0.008053160742076529),
   ('water', 0.006966231071366814),
   ('run', 0.006892122230182061)]),
   (0,
   [('people', 0.025218349201353372),
   ('kill', 0.01500904870564167),
   ('child', 0.013612400660948935),
   ('armenian', 0.010307655991816822),
   ('woman', 0.010287984892595798),
   ('start', 0.01003226060272248),
   ('day', 0.00967818081674404),
   ('happen', 0.009383114328428673),
   ('leave', 0.009383114328428673),
   ('fire', 0.009009363443229208)]),
   (1,
   [('file', 0.030686386604212003),
   ('program', 0.02227713642901929),
   ('window', 0.01945561169918489),
   ('set', 0.015914874783314277),
   ('line', 0.013831003577619592),
   ('display', 0.013794120901412606),
   ('application', 0.012576992586582082),
   ('entry', 0.009275993066056873),
   ('change', 0.00872275292295209),
   ('color', 0.008612104894331132)]),
   (12,
   [('line', 0.07153810971508515),
   ('buy', 0.02975597944523662),
   ('organization', 0.026877236406682988),
   ('host', 0.025451316957679788),
   ('price', 0.025182275552207485),
   ('sell', 0.02461728860071565),
   ('mail', 0.02192687454599263),
   ('good', 0.018967419085797303),
   ('sale', 0.017998870026097017),
   ('send', 0.013694207538540181)]),
   (11,
   [('thing', 0.04901329901329901),
   ('good', 0.0376018876018876),
   ('make', 0.03393393393393394),
   ('time', 0.03326898326898327),
   ('bad', 0.02664092664092664),
   ('happen', 0.017696267696267698),
   ('hear', 0.015615615615615615),
   ('problem', 0.015465465465465466),
   ('back', 0.015143715143715144),
   ('lot', 0.01495066495066495)]),
   (18,
   [('space', 0.020626317374284855),
   ('launch', 0.00965716006366413),
   ('system', 0.008560244332602057),
   ('project', 0.008173097603991913),
   ('time', 0.008108573149223556),
   ('cost', 0.007764442723792318),
   ('year', 0.0076784101174345075),
   ('earth', 0.007484836753129436),
   ('base', 0.0067535595990880545),
   ('large', 0.006689035144319697)]),
   (5,
   [('government', 0.01918437232469453),
   ('people', 0.01461203206475212),
   ('state', 0.011207097828624796),
   ('country', 0.010214802708381975),
   ('israeli', 0.010039691804809714),
   ('war', 0.009436532025838587),
   ('force', 0.00858043427504086),
   ('attack', 0.008424780138532182),
   ('land', 0.0076659662230523775),
   ('world', 0.0075103120865437)]),
   (2,
   [('car', 0.041091194044470564),
   ('bike', 0.015598981291017729),
   ('ride', 0.011019688510138114),
   ('drive', 0.010627877363110981),
   ('engine', 0.009403467528651191),
   ('speed', 0.008081104907434616),
   ('turn', 0.007738270153785875),
   ('back', 0.007738270153785875),
   ('front', 0.007468899990204721),
   ('big', 0.007370947203447938)])
]

评估性能

现在我们还可以通过计算连贯性得分来评估其性能，如下所示：

ldamallet = gensim.models.wrappers.LdaMallet(
   mallet_path, corpus=corpus, num_topics=20, id2word=id2word
)
pprint(ldamallet.show_topics(formatted=False))

输出

Coherence Score: 0.5842762900901401