2
$\begingroup$

I have a list of 50k sentences such as : 'bone is making noise', 'nose is leaking' ,'eyelid is down' etc..

I'm trying to use Doc2Vec to find the most similar sentence from the 50k given a new sentence. 

tagged_data = [TaggedDocument(words=word_tokenize(_d.lower()), tags=[str(i)]) for i, _d in enumerate(data)]

max_epochs = 100
vec_size = 20
alpha = 0.025

model = Doc2Vec(size=vec_size,
                alpha=alpha, 
                min_alpha=0.025,
                min_count=1,
                dm =0)

model.build_vocab(tagged_data)

for epoch in range(max_epochs):
    print('iteration {0}'.format(epoch))
    model.train(tagged_data,
                total_examples=model.corpus_count,
                epochs=model.iter)
    # decrease the learning rate
    model.alpha -= 0.0002
    # fix the learning rate, no decay
    model.min_alpha = model.alpha

test_data = word_tokenize("The nose is leaking blood after head injury".lower())
v1 = model.infer_vector(test_data)
#print("V1_infer", v1)

similar_doc = model.docvecs.most_similar(positive=[model.infer_vector(test_data)],topn=3)

for i in range(0,len(similar_doc)):
    print(tagged_data[int(similar_doc[i][0])],similar_doc[i][1])

Such that for the sentence "The nose is leaking blood after head injury" i would like to get the sentence with the highest similarity score ( i guess that it will bring sentences with words like leak or even synonyms like dripping?) . But the sentence i get back are unrelated and change each iteration of model.infer_vector(test_data)

Any idea about what is wrong?

Improve this question
$\endgroup$
4
  • $\begingroup$ Are you using pre-trained work vectors ? If not, 50k sentences are not enough for model to learn patterns. $\endgroup$
    – Shamit Verma
    Apr 2, 2019 at 14:55
  • $\begingroup$ @ShamitVerma , Each sentence in the training is unique (and not labled , its just a list of sentences.. ) . How do i use pre-tained word vector in that setup ? $\endgroup$
    – Latent
    Apr 2, 2019 at 15:00
  • $\begingroup$ If you replace words with word-vectors (E.g. : Fasttext or Glove) model will have prior knowledge that leak and drip are similar. $\endgroup$
    – Shamit Verma
    Apr 2, 2019 at 15:11
  • $\begingroup$ @ShamitVerma , can you please elaborate some more or link me to relevant source? i'm not sure i understand how both frameworks should work together. isnt Doc2vec already project the sentence i give him into vector space similar to the one in Glove? $\endgroup$
    – Latent
    Apr 3, 2019 at 14:04

1 Answer 1

Reset to default
2
$\begingroup$

Doc2Vec (and words vectors) need significant amount of data to learn useful vector representation. 50k sentences is not sufficient for this. To overcome this, you can feed word vectors as initial weights in Embedding Layer of network.

For example, code from following question :

How to implement LSTM using Doc2Vec vectors?

model_doc2vec = Sequential()
model_doc2vec.add(Embedding(voacabulary_dim, 100, input_length=longest_document, weights=[training_weights], trainable=False))
model_doc2vec.add(LSTM(units=10, dropout=0.25, recurrent_dropout=0.25, return_sequences=True))
model_doc2vec.add(Flatten())
model_doc2vec.add(Dense(3, activation='softmax'))
model_doc2vec.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

Output of "Flatten" layer will be vector representation of a sentence / document.

Article with example code.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Not sure how is the code you added is related. There is no supervised model to train here. How do i use the flatten vectors to find similar sentences . Given i've projected it to vector space and i have the 1D vector representation for each word (remmber that i've looked for sentences ..) , what's next? $\endgroup$
    – Latent
    Apr 7, 2019 at 6:57
  • $\begingroup$ You have to input a sentence into the model and then use output of "Flatten()" layer. This will be the projection of sentence into 1D space. $\endgroup$
    – Shamit Verma
    Apr 8, 2019 at 3:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.