Get the aligned BERT embedding for sequence labeling tasks
In the origin script extract_features.py in BERT, tokens may be splited into pieces as follows:
orig_tokens = ["John", "Johanson", "'s", "house"]
bert_tokens = ["[CLS]", "john", "johan", "##son", "'", "s", "house", "[SEP]"]
orig_to_tok_map = [1, 2, 4, 6]
We investigate 3 align strategies (first
, mean
and max
) to maintain an original-to-tokenized alignment. Take the "Johanson
-> johan
, ##son
" as example:
first
: take the representation ofjohan
as the whole wordJohanson
mean
: take the reduce_mean value of representations ofjohan
and##son
as the whole wordJohanson
max
: take the reduce_max value of representations ofjohan
and##son
as the whole wordJohanson
sh run.sh input_file outout_file BERT_BASE_DIR
# For example:
sh run.sh you_data you_data.bert path/to/bert/uncased_L-12_H-768_A-12
You can modify layers
and align_strategies
in the run.sh
.
After the above procedure, you are expected to get a output file of contextual embeddings (e.g., your_data_6_mean). Then you can load this file like conventional word embeddings. For example in a python script:
with open("your_data_6_mean", "r", encoding="utf-8") as bert_f"
for line in bert_f:
bert_vec = [[float(value) for value in token.split()] for token in line.strip().split("|||")]