Effects of data imbalance on NER model

1

I have a dataset that is very imbalanced, and I have to train a ner model on it. I was wondering how will imbalanced data affect model performance ? Below is the distribution of the data for example -

Tops                    2981
Coat & Jacket           2327
Pants                   2277
Dress                   2180
Shorts                  1969
Jeans                   1765
Sandals                 1645
Boots                   1637
Sweaters                1633
Skirt                   1600
Sneakers                1593
Bra                     1549
Flats                   1508
Jumpsuits               1252
Panties                 1140
Pumps                   1114
Mules                   1060
Earrings                 958
Bracelets                897
Clutches                 870
Necklaces                848
Totes                    821
Cover ups                797
Rings                    793
Shoulder Bag             743
Hosiery                  738
Watches                  688
One Piece                676
Sunglasses               662
Belts                    662
                        ... 
Bodysuits                292
Slippers                 281
Hair Accessories         278
Luggage and Travel       231
Robe                     227
Slingback                169
Gown                     150
Hobo Bag                 134
Pajamas                  122
Belt bag                 110
Saddle Bag               103
Chemise                  102
Cold Weather              86
Diaper Bag                77
Anklets                   73
Camera                    70
Accessories & Cases       67
Hoodies & SweatShirt      50
Pins                      39
Eyewear                   30
Briefcase                 22
Brooches                  22
Messenger Bag             21
Bustier                   17
Rompers                   15
Sock-Bootie               14
Two Piece                  8
Accessories                8
Underwear                  1
Convertible Bag            1
Name: Style, Length: 70, dtype: int64

Is their any sampling technique that I can use to remove the imbalance ?

2

Honestly these don’t sound like very good categories for direct NER classification. I don’t think it will be efficient to have the model learn exactly which item-names are of each type based on labelled examples. Have you tried annotating into much coarser-grained entity types, and then using a lookup table to resolve the fine-grained type?

For instance, if you want to find names of actors in text, it’s often better to recognize people, and then resolve the PERSON entities to pages in an ontology (e.g. by resolving them to a Wikipedia page). Once you have the Wikipedia page, deciding who is and is not an actor is easy. I expect the same sort of strategy would work well for your problem.

First, identify whether some piece of text belongs to one of a small number of categories. You can experiment with which categories work best, but try to pay attention to whether the local context can decide the issue, or whether too much “world knowledge” is needed. Once you have the items, then try to match them to a catalogue entry or some other knowledge-base. You can then attach very fine-grained labels, by reading them off the catalogue.