metadata

language:
  - he
size_categories:
  - 1K<n<10K
task_categories:
  - token-classification
pretty_name: HebrewMetaphors
dataset_info:
  features:
    - name: text
      dtype: string
    - name: source
      dtype: string
    - name: label
      dtype: int64
    - name: verb
      dtype: string
    - name: tweet_id
      dtype: string
  splits:
    - name: train
      num_bytes: 669160
      num_examples: 4944
    - name: val
      num_bytes: 168781
      num_examples: 1271
    - name: test
      num_bytes: 206836
      num_examples: 1593
  download_size: 449410
  dataset_size: 1044777
model:
  - tdklab/hebert-finetuned-hebrew-metaphor
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: val
        path: data/val-*
      - split: test
        path: data/test-*

Dataset Card for "HebrewMetaphors"

Dataset Summary

A common dataset for text classification task is IMDb. Large Movie Review Dataset. This is a dataset for binary sentiment classification. The first step in our project was to create a Hebrew dataset with an IMDB-like structure but different in that, in addition to the sentences we have, there will also be verb names, and a classification of whether the verb name is literal or metaphorical in the given sentence. Using an API, sentences that contained the verbs we selected were pulled from Twitter and Wikipedia and Prodigy was used to classify them. The accuracy of the classification was confirmed by having two different persons classify each line twice.

Supported Tasks and Leaderboards

Classification

Languages

Hebrew

Dataset Structure

The IMDb dataset is a json file with objects containing the text of the review, and a number 0 or 1 which is a negative or positive review

For example:

{
    "label": 1, //(pos)
    "text": "A wonderful little production ... "
}, {
    "label": 0, //(neg)
    "text": ""It actually pains me to say it, but this movie was horrible on every level ..."
}

After retrieving data from Wikipedia and Twitter, classifying and processing, we received the following data:

{'text': 'מתחיל לאמץ לשלוח את הבגדים למכבסה אחרי טיול לחול כי למי יש כוח לכבס כל כך הרבה בגדים אחרי חופש',
 'source': 'twitter',
 'label': 0,
 'verb': 'לכבס',
 'tweet_id': '1546803262065606658'},
 
{'text': ' בינג בונג מגיח מן המערב ומתחיל לזרוע הרס בממלכת עוץ',
 'source': 'wikipedia',
 'label': 1, //(Metaphor)
 'verb': 'לזרוע',
 'tweet_id': None},
 
 {'text': 'הם ישתמשו במערבל כדי להפנות את המטוס לשדה התעופה שהם יבחרו וכך לשדוד אותו',
 'source': 'wikipedia',
 'label': 0, //(Literal)
 'verb': 'לשדוד',
 'tweet_id': None}

As you can see, we have the following fields:

'text': A sentence in Hebrew that was taken from Wikipedia or Twitter that included the verb name.
'source': There are two possible sources for this text: 'twitter' or 'wikipedia'.
'label': Classification of whether the verb name is literal ('label': 0) or metaphorical ('label': 1) in the given sentence.
'verb': A verb name in Hebrew on which the text is labeled.
'tweet_id': If the text was taken from Twitter, the tweet's id, or None if it was retrieved from Wikipedia.

In the next section we will explain how we created the dataset.

Data Statistics

Verb	After Agreement Between Taggers	Disagreements because of different answers	Disagreements because of different ignores	Total Literals After Agreement	Total Metaphors After Agreement	Split For Train Set	Split For Validation Set	Split For Test Set
lakhalom_twitter	131	42	42	26	105	83	21	27
lakhalom_wiki	93	53	69	34	59	58	16	19
lakhatokh_twitter	220	41	40	100	120	140	36	44
lakhatokh_wiki	288	54	6	286	2	182	47	59
lauf_twitter	166	8	29	14	152	104	28	34
lauf_wiki	172	11	67	161	11	108	28	36
lefareq_twitter	70	109	71	21	49	43	12	15
lefareq_wiki	96	48	213	89	7	60	16	20
lehadliq_twitter	270	14	19	194	76	172	43	55
lehadliq_wiki	375	13	13	370	5	239	61	75
lekhabes_twitter	203	2	26	155	48	129	33	41
lekhabes_wiki	55	1	0	51	4	34	9	12
lekhofef_twitter	289	7	11	18	271	183	47	59
lekhofef_wiki	107	3	6	73	34	67	18	22
lerasek_twitter	141	6	4	18	123	89	23	29
lerasek_wiki	154	3	12	110	44	98	25	31
levashel_twitter	280	3	19	264	16	177	46	57
levashel_wiki	228	3	19	221	7	144	37	47
limkhoq_twitter	222	16	22	123	99	141	36	45
limkhoq_wiki	212	11	29	109	103	134	35	43
liqpots_twitter	178	18	25	77	101	112	29	37
liqpots_wiki	168	6	28	163	5	107	27	34
liqroa_twitter	261	26	16	36	225	166	42	53
liqroa_wiki	226	18	19	147	79	143	37	46
liqtsor_twitter	256	18	28	43	213	163	41	52
liqtsor_wiki	167	11	2	99	68	106	27	34
lirqod_twitter	187	14	31	138	49	119	30	38
lirqod_wiki	214	18	34	211	3	135	35	44
lishbor_twitter	185	8	30	38	147	117	30	38
lishbor_wiki	244	6	50	69	175	156	39	49
lishdod_twitter	163	21	16	91	72	102	27	34
lishdod_wiki	209	2	4	209	0	133	34	42
lishtot_twitter	275	6	20	219	56	175	44	56
lishtot_wiki	192	7	2	192	0	122	31	39
litkhon_twitter	280	13	20	44	236	178	45	57
litkhon_wiki	87	0	0	85	2	54	15	18
litpor_twitter	193	4	13	26	167	122	31	40
litpor_wiki	178	2	14	153	25	113	29	36
lizroa_twitter	246	3	51	11	235	156	40	50
lizroa_wiki	127	2	171	67	60	80	21	26
total	7808	651	1291	4555	3253	4944	1271	1593

Contributions

Created by Doron Ben-chorin, Matan Ben-chorin, Tomer Tzipori, Guided by Dr. Oren Mishali. This is our project as part of computer engineering studies in the Faculty of Electrical Engineering combined with Computer Science at Technion, Israel Institute of Technology. For more cooperation, please contact email:

Doron Ben-chorin: [email protected]

Matan Ben-chorin: [email protected]

Tomer Tzipori: [email protected]