Python – How to compare sentence similarities using embeddings from BERT

cosine-similarityhuggingface-transformersnlppythonvector

I am using the HuggingFace Transformers package to access pretrained models. As my use case needs functionality for both English and Arabic, I am using the bert-base-multilingual-cased pretrained model. I need to be able to compare the similarity of sentences using something such as cosine similarity. To use this, I first need to get an embedding vector for each sentence, and can then compute the cosine similarity.

Firstly, what is the best way to extratc the semantic embedding from the BERT model? Would taking the last hidden state of the model after being fed the sentence suffice?

import torch
from transformers import BertModel, BertTokenizer

model_class = BertModel
tokenizer_class = BertTokenizer
pretrained_weights = 'bert-base-multilingual-cased'

tokenizer = tokenizer_class.from_pretrained(pretrained_weights)
model = model_class.from_pretrained(pretrained_weights)

sentence = 'this is a test sentence'

input_ids = torch.tensor([tokenizer.encode(sentence, add_special_tokens=True)])
with torch.no_grad():
    output_tuple = model(input_ids)
    last_hidden_states = output_tuple[0]

print(last_hidden_states.size(), last_hidden_states)

Secondly, if this is a sufficient way to get embeddings from my sentence, I now have another problem where the embedding vectors have different lengths depending on the length of the original sentence. The shapes output are [1, n, vocab_size], where n can have any value.

In order to compute two vectors' cosine similarity, they need to be the same length. How can I do this here? Could something as naive as first summing across axis=1 still work? What other options do I have?

Best Answer

In addition to an already great accepted answer, I want to point you to sentence-BERT, which discusses the similarity aspect and implications of specific metrics (like cosine similarity) in greater detail. They also have a very convenient implementation online. The main advantage here is that they seemingly gain a lot of processing speed compared to a "naive" sentence embedding comparison, but I am not familiar enough with the implementation itself.

Importantly, there is also generally a more fine-grained distinction in what kind of similarity you want to look at. Specifically for that, there is also a great discussion in one of the task papers from SemEval 2014 (SICK dataset), which goes into more detail about this. From your task description, I am assuming that you are already using data from one of the later SemEval tasks, which also extended this to multilingual similarity.

Related Solutions

Python – How to execute a program or call a system command

Use the subprocess module in the standard library:

import subprocess
subprocess.run(["ls", "-l"])

The advantage of subprocess.run over os.system is that it is more flexible (you can get the stdout, stderr, the "real" status code, better error handling, etc...).

Even the documentation for os.system recommends using subprocess instead:

The subprocess module provides more powerful facilities for spawning new processes and retrieving their results; using that module is preferable to using this function. See the Replacing Older Functions with the subprocess Module section in the subprocess documentation for some helpful recipes.

On Python 3.4 and earlier, use subprocess.call instead of .run:

subprocess.call(["ls", "-l"])

Python – How to select rows from a DataFrame based on column values

To select rows whose column value equals a scalar, some_value, use ==:

df.loc[df['column_name'] == some_value]

To select rows whose column value is in an iterable, some_values, use isin:

df.loc[df['column_name'].isin(some_values)]

Combine multiple conditions with &:

df.loc[(df['column_name'] >= A) & (df['column_name'] <= B)]

Note the parentheses. Due to Python's operator precedence rules, & binds more tightly than <= and >=. Thus, the parentheses in the last example are necessary. Without the parentheses

df['column_name'] >= A & df['column_name'] <= B

is parsed as

df['column_name'] >= (A & df['column_name']) <= B

which results in a Truth value of a Series is ambiguous error.

To select rows whose column value does not equal some_value, use !=:

df.loc[df['column_name'] != some_value]

isin returns a boolean Series, so to select rows whose value is not in some_values, negate the boolean Series using ~:

df.loc[~df['column_name'].isin(some_values)]

For example,

import pandas as pd
import numpy as np
df = pd.DataFrame({'A': 'foo bar foo bar foo bar foo foo'.split(),
                   'B': 'one one two three two two one three'.split(),
                   'C': np.arange(8), 'D': np.arange(8) * 2})
print(df)
#      A      B  C   D
# 0  foo    one  0   0
# 1  bar    one  1   2
# 2  foo    two  2   4
# 3  bar  three  3   6
# 4  foo    two  4   8
# 5  bar    two  5  10
# 6  foo    one  6  12
# 7  foo  three  7  14

print(df.loc[df['A'] == 'foo'])

yields

     A      B  C   D
0  foo    one  0   0
2  foo    two  2   4
4  foo    two  4   8
6  foo    one  6  12
7  foo  three  7  14

If you have multiple values you want to include, put them in a list (or more generally, any iterable) and use isin:

print(df.loc[df['B'].isin(['one','three'])])

yields

     A      B  C   D
0  foo    one  0   0
1  bar    one  1   2
3  bar  three  3   6
6  foo    one  6  12
7  foo  three  7  14

Note, however, that if you wish to do this many times, it is more efficient to make an index first, and then use df.loc:

df = df.set_index(['B'])
print(df.loc['one'])

yields

       A  C   D
B              
one  foo  0   0
one  bar  1   2
one  foo  6  12

or, to include multiple values from the index use df.index.isin:

df.loc[df.index.isin(['one','two'])]

yields

       A  C   D
B              
one  foo  0   0
one  bar  1   2
two  foo  2   4
two  foo  4   8
two  bar  5  10
one  foo  6  12

Best Answer

Related Solutions

Python – How to execute a program or call a system command

Python – How to select rows from a DataFrame based on column values

Related Topic