How to get BioBERT embeddings

Question

I have field within a pandas dataframe with a text field for which I want to generate BioBERT embeddings. Is there a simple way with which I can generate the vector embeddings? I want to use them within another model.

here is a hypothetical sample of the data frame

Visit Code	Problem Assessment
1234	ge reflux working diagnosis well
4567	medication refill order working diagnosis note called in brand benicar 5mg qd 30 prn refill

I have tried this package, but receive an error upon installation https://pypi.org/project/biobert-embedding

Error:

Collecting biobert-embedding
  Using cached biobert-embedding-0.1.2.tar.gz (4.8 kB)
ERROR: Could not find a version that satisfies the requirement torch==1.2.0 (from biobert-embedding) (from versions: 0.1.2, 0.1.2.post1, 0.1.2.post2, 1.7.1)
ERROR: No matching distribution found for torch==1.2.0 (from biobert-embedding)

Any help is GREATLY appreciated!

Answer 1

Try to install it as follows:

pip install biobert-embedding==0.1.2 torch==1.2.0 -f https://download.pytorch.org/whl/torch_stable.html

I extended your sample dataframe to illustrate how you can now calculate the sentence vectors for your problem assessments and use these to calculate for example the cosine similarity between similar visit codes.

>>> from biobert_embedding.embedding import BiobertEmbedding
>>> from scipy.spatial import distance
>>> import pandas as pd

>>> data = {'Visit Code': [1234, 1235, 4567, 4568], 
        'Problem Assessment': ['ge reflux working diagnosis well', 
                               'other reflux diagnosis poor', 
                               'medication refill order working diagnosis note called in brand benicar 5mg qd 30 prn refill',
                               'medication must be refilled diagnosis note called in brand Olmesartan 10mg qd 40 prn refill']}

>>> df = pd.DataFrame(data)
>>> df

	Visit Code	Problem Assessment
0	1234	ge reflux working diagnosis well
1	1234	other reflux diagnosis poor
2	4567	medication refill order working diagnosis note called in brand benicar 5mg qd 30 prn refill
3	4567	medication must be refilled diagnosis note called in brand Olmesartan 10mg qd 40 prn refill

>>> biobert = BiobertEmbedding()
>>> df['sentence embedding'] = df['Problem Assessment'].apply(lambda sentence: biobert.sentence_vector(sentence))
>>> df

	Visit Code	Problem Assessment	sentence embedding
0	1234	ge reflux working diagnosis well	tensor([ 2.7189e-01, -1.6195e-01, 5.8270e-02, -3.2730e-01, 7.5583e-02, ...
1	1234	other reflux diagnosis poor	tensor([ 1.6971e-01, -2.1405e-01, 3.4427e-02, -2.3090e-01, 1.6007e-02, ...
2	4567	medication refill order working diagnosis note called in brand benicar 5mg qd 30 prn refill	tensor([ 1.5370e-01, -3.9875e-01, 2.0089e-01, 4.1506e-02, 6.9854e-02, ...
3	4567	medication must be refilled diagnosis note called in brand Olmesartan 10mg qd 40 prn refill	tensor([ 2.2128e-01, -2.0283e-01, 2.2194e-01, 9.1156e-02, 1.1620e-01, ...

>>> df.groupby('Visit Code')['sentence embedding'].apply(lambda sentences: 1 - distance.cosine(sentences.values) )


Visit Code
1234    0.950492
4567    0.969715
Name: sentence embedding, dtype: float64

We can see that, as expected, the similar sentences lie very close together