tokenize with Spacy

[jbesomi]

The actual tokenizer is very fast as it uses a simple regex pattern but at the same time very imprecise.

A better alternative might be to make use of spaCy.

Naively tokenize a Pandas Series with spaCy is very simple:

def tokenize_with_spacy(s: pd.Series) -> pd.Series:
   
    nlp = spacy.load("en_core_web_sm")# disable=["ner", "tagger", "parser"])
    
    tokenized = []
    for doc in nlp.pipe(s):
        tokenized.append(list(map(str, doc)))
    
    return pd.Series(tokenized, index=s.index)

This should be somehow fast as in theory spaCy makes use of multi-threading.

The reason we haven't implemented this yet is that we want to make sure this solution is enough faster. We want to provide a simple tool to analyze a quite large amount of text data; say 100k Pandas Row should take no longer than 15-30 seconds to tokenize ... ?

For now, the task consists in:

1. Compare the "spacy solution" w.r.t the actual version and benchmark the function on large datasets (150k rows or so)
   1. This should be done in a single, clean notebook that will be shared there
2. Evaluate if we can do better by parallelizing even more the process (we can probably parallelize both by row and by sentence?)

[jbesomi]

Dask vs. spaCy

It's faster to use pipe from spaCy or to directly use Dask (Dask Dataframe)?

Dask might be the solution we were looking for ...

[mk2510]

As described in #162 , dask is not feasible from a UX perspective. Here are our results from experimenting with tokenize. See THE ATTACHED PDF for a notebook of the results.

Speed Comparison

We now compare:

1. current implementation without parallelization
2. current implementation with parallelization (see Speed-Up Preprocessing + NLP #162)
3. tokenize_with_spacy with spacy built-in parallelization through n_process
4. tokenize_with_spacy with our custom parallelization

Results below.

We can see that

• our current implementation is much faster than spaCy (22 vs 51 seconds with both parallelized)
• as shown in Speed-Up Preprocessing + NLP #162, our parallelization works better than spaCy's.

Thus, our options:

1. keep everything as proposed in Speed-Up Preprocessing + NLP #162 (-> multiprocessing applied to current solution)
2. option 1, but we give users a parameter use_spacy that works like our tokenize_with_spacy_own_parallelization above, and explain to them that this might give them better results but takes about 3x as long.

We don't really have a preference.