CODEWITHSUNDEEP
pythonpandasstringdataframepandas-groupby
Mi.
Mi.

Reputation: 510

Combining Rows in a DataFrame

I have a DF that has the results of a NER classifier such as the following:

df =

s        token        pred       tokenID
17     hakawati       B-Loc         3
17     theatre        L-Loc         3
17     jerusalem      U-Loc         7
56     university     B-Org         5
56     of             I-Org         5
56     texas          I-Org         5
56     here           L-Org         6
...
5402   dwight         B-Peop        1    
5402   d.             I-Peop        1
5402   eisenhower     L-Peop        1

There are many other columns in this DataFrame that are not relevant. Now I want to group the tokens depending on their sentenceID (=s) and their predicted tags to combine them into a single entity:

df2 =


s        token                        pred               
17     hakawati  theatre           Location
17     jerusalem                   Location
56     university of texas here    Organisation
...
5402   dwight d. eisenhower        People

Normally I would do so by simply using a line like data_map = df.groupby(["s"],as_index=False, sort=False).agg(" ".join) and using a rename function. However since the data contains different kind of Strings (B,I,L - Loc/Org ..) I don't know how to exactly do it.

Any ideas are appreciated.

Any ideas?

Upvotes: 1

Views: 67

Answers (2)

jpp
jpp

Reputation: 164623

One solution via a helper column.

df['pred_cat'] = df['pred'].str.split('-').str[-1]

res = df.groupby(['s', 'pred_cat'])['token']\
        .apply(' '.join).reset_index()

print(res)

      s pred_cat                       token
0    17      Loc  hakawati theatre jerusalem
1    56      Org    university of texas here
2  5402     Peop        dwight d. eisenhower

Note this doesn't match exactly your desired output; there seems to be some data-specific treatment involved.

Upvotes: 1

Piotr
Piotr

Reputation: 2117

You could group by both s and tokenID and aggregate like so:

def aggregate(df):
    token = " ".join(df.token)
    pred = df.iloc[0].pred.split("-", 1)[1]
    return pd.Series({"token": token, "pred": pred})

df.groupby(["s", "tokenID"]).apply(aggregate)

# Output
                             token  pred
s    tokenID                            
17   3            hakawati theatre   Loc
     7                   jerusalem   Loc
56   5         university of texas   Org
     6                        here   Org
5402 1        dwight d. eisenhower  Peop

Upvotes: 0

Related Questions