Doc2Vec Get most similar documents

Question

I am trying to build a document retrieval model that returns most documents ordered by their relevancy with respect to a query or a search string. For this I trained a doc2vec model using the Doc2Vec model in gensim. My dataset is in the form of a pandas dataset which has each document stored as a string on each line. This is the code I have so far

import gensim, re
import pandas as pd

# TOKENIZER
def tokenizer(input_string):
    return re.findall(r"[\w']+", input_string)

# IMPORT DATA
data = pd.read_csv('mp_1002_prepd.txt')
data.columns = ['merged']
data.loc[:, 'tokens'] = data.merged.apply(tokenizer)
sentences= []
for item_no, line in enumerate(data['tokens'].values.tolist()):
    sentences.append(LabeledSentence(line,[item_no]))

# MODEL PARAMETERS
dm = 1 # 1 for distributed memory(default); 0 for dbow 
cores = multiprocessing.cpu_count()
size = 300
context_window = 50
seed = 42
min_count = 1
alpha = 0.5
max_iter = 200

# BUILD MODEL
model = gensim.models.doc2vec.Doc2Vec(documents = sentences,
dm = dm,
alpha = alpha, # initial learning rate
seed = seed,
min_count = min_count, # ignore words with freq less than min_count
max_vocab_size = None, # 
window = context_window, # the number of words before and after to be used as context
size = size, # is the dimensionality of the feature vector
sample = 1e-4, # ?
negative = 5, # ?
workers = cores, # number of cores
iter = max_iter # number of iterations (epochs) over the corpus)

# QUERY BASED DOC RANKING ??

The part where I am struggling is in finding documents that are most similar/relevant to the query. I used the infer_vector but then realised that it considers the query as a document, updates the model and returns the results. I tried using the most_similar and most_similar_cosmul methods but I get words along with a similarity score(I guess) in return. What I want to do is when I enter a search string(a query), I should get the documents (ids) that are most relevant along with a similarity score(cosine etc). How do I get this part done?

Solution

You need to use infer_vector to get a document vector of the new text - which does not alter the underlying model.

Here is how you do it:

tokens = "a new sentence to match".split()

new_vector = model.infer_vector(tokens)
sims = model.docvecs.most_similar([new_vector]) #gives you top 10 document tags and their cosine similarity

Edit:

Here is an example of how the underlying model does not change after infer_vec is called.

import numpy as np

words = "king queen man".split()

len_before =  len(model.docvecs) #number of docs

#word vectors for king, queen, man
w_vec0 = model[words[0]]
w_vec1 = model[words[1]]
w_vec2 = model[words[2]]

new_vec = model.infer_vector(words)

len_after =  len(model.docvecs)

print np.array_equal(model[words[0]], w_vec0) # True
print np.array_equal(model[words[1]], w_vec1) # True
print np.array_equal(model[words[2]], w_vec2) # True

print len_before == len_after #True

Comments

[Clock Slave]

are you sure that it doesn't update the model. The infer_vector method takes parameters like alpha and min_alpha. I'm assuming they are learning rates. There's not much given in the documentation so I am not really sure if they are learning rates or some other parameters. Also, I came to think that it was updating the model because every time I passed the same sentence to infer_vector and then to most_similar, I got different results each time

[Erock]

infer_vector like the training is has non-deterministic elements. You will get different vectors on each call. There are a number of discussions out there on Gensim's mailing list and their issue log on github. Here is a good one good example: github.com/RaRe-Technologies/gensim/issues/447. Also, you can test if the model changes. See my edit.

[Antoine]

it's clearly stated in doc2vec paper that at inference time, all the parameters of the model are fixed. So the model definitely doesn't get updated.

[user2849678]

@ClockSlave Yes, infer_vector is changing the model. I am reloading the model, after infer_vector & the output is deterministic. Very useful post!