vkovenko/cross_domain_uk_reviews · Datasets at Hugging Face

index int64 0 663k	split stringclasses 3 values
0	train
1	test
2	train
3	train
4	train
5	test
6	train
7	train
8	train
9	train
10	test
11	test
12	train
13	train
14	train
15	train
16	val
17	train
18	train
19	train
20	train
21	train
22	val
23	val
24	train
25	train
26	test
27	train
28	train
29	train
30	train
31	test
32	train
33	train
34	train
35	val
36	train
37	train
38	train
39	test
40	train
41	train
42	train
43	train
44	train
45	train
46	train
47	test
48	val
49	train
50	train
51	train
52	train
53	train
54	train
55	test
56	train
57	test
58	train
59	train
60	train
61	train
62	train
63	train
64	train
65	val
66	val
67	val
68	train
69	train
70	val
71	train
72	train
73	train
74	train
75	test
76	train
77	train
78	train
79	train
80	train
81	train
82	train
83	test
84	train
85	train
86	train
87	train
88	val
89	train
90	train
91	train
92	train
93	test
94	train
95	val
96	train
97	train
98	train
99	test

index

int64

663k

split

stringclasses

3 values

train

test

train

test

train

test

train

val

train

val

train

test

train

test

train

val

train

test

train

test

val

train

test

train

test

train

val

train

val

train

test

train

test

train

val

train

test

train

val

train

test

The dataset is relevant to Ukrainian reviews in three different domains:

Hotels.
Reustarants.
Products.

The dataset is comrpised of several .csv files, which one can found useful:

processed_data.csv - the processed dataset itself.
train_val_test_indices.csv - csv file with train/val/test indices. The split was stratified w.r.t dataset name (hotels, reustarants, products) and rating.
bad_ids.csv - csv file with ids of bad samples marked using model filtering approach, only ids of those samples for which difference between actual and predicted rating is bigger than 2 points are maintained in this file.

The data is scrapped from Tripadvisor (https://www.tripadvisor.com/) and Rozetka (https://rozetka.com.ua/).

The dataset was initially used for extraction of key-phrases relevant to one of rating categories, based on trained machine learning model (future article link will be here).

Dataset is processed to include two additional columns: one with lemmatized tokens and another one with POS tags. Both lemmatization and POS tagging are done using pymorphy2 (https://pymorphy2.readthedocs.io/en/stable/) library.

The words are tokenized using a specific regex tokenizer to account for usage of apostroph.

Those reviews which weren't in Ukrainian were translated to it using Microsoft translator and re-checked manually afterwards.

Datasets:

vkovenko
/

cross_domain_uk_reviews

Models trained or fine-tuned on vkovenko/cross_domain_uk_reviews

vkovenko/deep_lstm_attention_ukr_reviews_rating_estimation