Splitting text DF into individual sentences DF: how to create a longer Pandas Data Frame with lambda and apply?

Question

This question might look long but I promise it is really not complicted.

I have a DF with textblocks and some ID columns. I want to create a new DF that contains each sentence as its own row.

original_df = pd.DataFrame(data={"year":[2018,2019], "text_nr":[1,2], "text":["This is one sentence. This is another!","Please help me. I am lost. "]})
original_df
>>>
       year  text_nr  text
    0  2018  1        "This is one sentence. This is another!"
    1  2019  2        "Please help me. I am lost."

I would like to split each textblock into individual sentences using spacy and create a new DF that looks like this:

sentences_df
>>>
   year  text_nr  sent_nr sentence
0  2018      1       1   "This is one sentence". 
1  2018      1       2   "This is another!"
2  2019      2       1   "Please help me."
3  2019      2       2   "I am lost."

I have found a way to do it like this:

import spacy
nlp = spacy.load("en_core_news_sm")
sentences_list = []

for i, row in original_df.iterrows():
    doc = nlp(row["text"])
    sentences = [(row["year"],row["text_nr"],str(i+1),sent.string.replace('\n','').replace('\t','').strip()) for i, sent in enumerate(doc.sents)]
    sentences_list = sentences_list+sentences

sentences_df = pd.DataFrame(sentences_list, columns = ["year",text_nr","sent_nr","sentence"])

But it is not very elegant and I read that df.apply(lambda: ...) method is much faster. However, when I try it, I never manage to get the correct result.I tried these two ways:

1. First Attempt:

nlp = spacy.load("en_core_news_sm")
def sentencizer (x, nlp_model):
    sentences = {}
    doc = nlp_model(x["text"])
    for i, sent in enumerate(doc.sents):
        sentences["year"]=x["year"]
        sentences["text_nr"]=x["text_nr"]
        sentences["sent_nr"] = str(i+1)
        sentences["sentence"] = sent.string.replace('\n','').replace('\t','').strip()
    return sentences
sentences_df = original_df.head().apply(lambda x: pd.Series(sentencizer(x,nlp)),axis=1)

This only ever gets the last sentence

sentences_df
>>>
   year  text_nr sent_nr  sentence
0  2018        1       2  "This is another!"
1  2019        2       2  "I am lost!"

2. Second Attempt

nlp = spacy.load("en_core_news_sm")
def sentencizer (x, nlp_model):
    sentences = {"year":[],"text_nr":[],"sent_nr":[],"sentence":[]}
    doc = nlp_model(x["text"])
    for i, sent in enumerate(doc.sents):
        sentences["year"].append(x["year"])
        sentences["text_nr"].append(x["text_nr"])
        sentences["sent_nr"].append(str(i+1))
        sentences["sentence"].append(sent.string.replace('\n','').replace('\t','').strip())
    return sentences
sentences_df = original_df.apply(lambda x: pd.Series(sentencizer(x,nlp)),axis=1)

This yields me a DF with lists as entrys:

sentences_df
>>>
   year          text_nr sent_nr    sentence
0  [2018, 2018]  [1, 1]  [1, 2]  ["This is one sentence.", "This is another!"]
1  [2019, 2019]  [2, 2]  [1, 2]  ["Please help me.", "I am lost."]

I could probably try to expand this last df, but I am sure there is a way to do this correctly in one go. I would like to use spacyfor splitting the text as it has more advanced sentence boundry detection than just using regex/ string splitting. You do not need to download spacy to help me (->string.split() is fine for the dummy data here). I just need to find a logic that works along the same lines as the following so I can rewrite it to use it with spacy.

nlp = spacy.load("en_core_news_sm")
doc = nlp("This is a sentence.\n This is another! ")
sentences = [sent.string.strip() for sent in doc.sents] #doc.sents is a generator
sentences
>>>
["This is a sentence", "This is another!"]

So something along the lines of this would be great:

text = "This is a sentence.\n This is another! "
sentences = [sent.replace("\n","").strip() for sent in text.split(".")]
sentences
>>>
["This is a sentence", "This is another!"]

Thanks a lot for any help. I am quite new to programming so please have mercy :)

Answer 1

Found a solution that works:

nlp = spacy.load("en_core_news_sm")
def splitter(x,nlp):
    doc = nlp(x["text"])
    a = [str(sent) for sent in doc.sents]
    b = len(a)
    dictionary = {"text_nr": np.repeat(x["text_nr"],b), "sentence_nr": list(range(1, b+1)), "sentence": a}
    dictionaries = [{key : value[i] for key, value in dictionary.items()} for i in range(b)]
    for dictionary in dictionaries:
        rows_list.append(dictionary)

original_df.apply(lambda x: splitter(x,nlp), axis = 1)
new_df = pd.DataFrame(rows_list, columns=['text_nr', 'sentence_nr','sentence'])

Answer 2

Something along this line might work:

# update punctuations list if needed
punctuations = '\.\!\?'
(original_df.drop('text',axis=1)
    .merge(original_df.text
               .str.extractall(f'(?P<sentence>[^{punctuations}]+[{punctuations}])\s?')
               .reset_index('match'),
           left_index=True, right_index=True, how='left')
)

Output:

   year  text_nr  match               sentence
0  2018        1      0  This is one sentence.
0  2018        1      1       This is another!
1  2019        2      0        Please help me.
1  2019        2      1             I am lost.