How to use TaggedDocument in gensim?

Question

I have two directories from which I want to read their text files and label them, but I don't know how to do this via TaggedDocument. I thought it would work as TaggedDocument([Strings],[Labels]) but this doesn't work apparently.

This is my code:

from gensim import models
from gensim.models.doc2vec import TaggedDocument
import utilities as util
import os
from sklearn import svm
from nltk.tokenize import sent_tokenize
CogPath = "./FixedCog/"
NotCogPath = "./FixedNotCog/"
SamplePath ="./Sample/"
docs = []
tags = []
CogList = [p for p in os.listdir(CogPath) if p.endswith('.txt')]
NotCogList = [p for p in os.listdir(NotCogPath) if p.endswith('.txt')]
SampleList = [p for p in os.listdir(SamplePath) if p.endswith('.txt')]
for doc in CogList:
     str = open(CogPath+doc,'r').read().decode("utf-8")
     docs.append(str)
     print docs
     tags.append(doc)
     print "###########"
     print tags
     print "!!!!!!!!!!!"
for doc in NotCogList:
     str = open(NotCogPath+doc,'r').read().decode("utf-8")
     docs.append(str)
     tags.append(doc)
for doc in SampleList:
     str = open(SamplePath + doc, 'r').read().decode("utf-8")
     docs.append(str)
     tags.append(doc)

T = TaggedDocument(docs,tags)

model = models.Doc2Vec(T,alpha=.025, min_alpha=.025, min_count=1,size=50)

and this is the error I get:

Traceback (most recent call last):
  File "/home/farhood/PycharmProjects/word2vec_prj/doc2vec.py", line 34, in <module>
    model = models.Doc2Vec(T,alpha=.025, min_alpha=.025, min_count=1,size=50)
  File "/home/farhood/anaconda2/lib/python2.7/site-packages/gensim/models/doc2vec.py", line 635, in __init__
    self.build_vocab(documents, trim_rule=trim_rule)
  File "/home/farhood/anaconda2/lib/python2.7/site-packages/gensim/models/word2vec.py", line 544, in build_vocab
    self.scan_vocab(sentences, progress_per=progress_per, trim_rule=trim_rule)  # initial survey
  File "/home/farhood/anaconda2/lib/python2.7/site-packages/gensim/models/doc2vec.py", line 674, in scan_vocab
    if isinstance(document.words, string_types):
AttributeError: 'list' object has no attribute 'words'

Solution 1

The input for a Doc2Vec model should be a list of TaggedDocument(['list','of','word'], [TAG_001]). A good practice is using the indexes of sentences as the tags. For example, to train a Doc2Vec model with two sentences (i.e. documents, paragraphs):

s1 = 'the quick fox brown fox jumps over the lazy dog'
s1_tag = '001'
s2 = 'i want to burn a zero-day'
s2_tag = '002'

docs = []
docs.append(TaggedDocument(words=s1.split(), tags=[s1_tag])
docs.append(TaggedDocument(words=s2.split(), tags=[s2_tag])

model = gensim.models.Doc2Vec(vector_size=300, window=5, min_count=5, workers=4, epochs=20)
model.build_vocab(docs)

print 'Start training process...'
model.train(docs, total_examples=model.corpus_count, epochs=model.iter)

#save model
model.save(model_path)

Solution 2

So I just experimented a bit and found this on github:

class TaggedDocument(namedtuple('TaggedDocument', 'words tags')):
    """
    A single document, made up of `words` (a list of unicode string tokens)
    and `tags` (a list of tokens). Tags may be one or more unicode string
    tokens, but typical practice (which will also be most memory-efficient) is
    for the tags list to include a unique integer id as the only tag.

    Replaces "sentence as a list of words" from Word2Vec.

so I decided to change how I use the TaggedDocument function by generating a TaggedDocument class for each document, the important thing is that you have to pass the tags as a list.

for doc in CogList:
     str = open(CogPath+doc,'r').read().decode("utf-8")
     str_list = str.split()
     T = TaggedDocument(str_list,[doc])
     docs.append(T)

Solution 3

You can use gensim's common_texts as an example:

from gensim.test.utils import common_texts
from gensim.models.doc2vec import Doc2Vec, TaggedDocument

documents = [TaggedDocument(doc, [i]) for i, doc in enumerate(common_texts)]
model = Doc2Vec(documents, vector_size=5, window=2, min_count=1, workers=4)

This will use common_texts and TaggedDocument to create the document representation expected by the Doc2Vec algorithm.

Comments

[gojomo]

Separate from your main question: having the ending min_alpha be the same value as the starting alpha means your training isn't doing proper stochastic gradient descent. Also, it's rare for min_count=1 to be helpful in Word2Vec/Doc2Vec training – keeping such rare words just tends to make training take longer and interfere with the quality of the remaining word-vecs/doc-vecs.

[Farhood]

about min_alpha , I've copied it from a sample code followed by this code: for epoch in range(10): model.train(docs) model.alpha -= 0.002 # decrease the learning rate model.min_alpha = model.alpha # fix the learning rate, no decay and about the min_count: my data set is very limited and some words are not that much frequent but weigh a lot in the meaning, also I have filtered most stop words and frequent daily words.

[gojomo]

That's a bad sample to follow. If you're passing your corpus in when creating the Doc2Vec instance, it will automatically do all its training passes, and automatically manage the learning-rate from alpha to min_alpha, and you shouldn't call train() yourself. (And if you do, like you've shown without any other specifics, the latest gensim versions will throw an error because it's such a common mistake.) It is a rare, expert thing to call train() yourself or much with the default alpha/min_alpha.

[gojomo]

Yes: Doc2Vec is expecting the corpus to be an iterable collection, where each individual item (document) is shaped like a TaggedDocument. (That is, it has a words list and tags list.)