Word Embeddings of Sake

About

• This tool enables "sake operation"
• The terms of sake are represented by a vector in an Euclidean space
• The sake embedding supports 14 categories.
• You can change categories of the sake such as type of rice, yeast, and rice polishing rate

玉乃光_純米大吟醸_京の琴_祝１００％ - 祝 + 雄町 = 玉乃光_純米大吟醸_備前雄町１００％
金鵄正宗_純米大吟醸_祝 - 精米歩合45% + 精米歩合60% = 金鵄正宗_特別純米

Setup

• Download sake dataset from here

$ cd data
$ git clone https://github.com/yoichi1484/sake_dataset
$ cp sake_dataset/json/sake_dataset_v1.json ./

Usage

Training sake embeddings with gensim

python train.py

Analogy of sake vectors

import utils
model = utils.load_sake_embedding('model.txt')
result = model.most_similar(positive=['brand:英勲', 'rice:山田錦'], 
                            negative=['rice:祝'], topn=1)
result[0] # ('brand+name:美丈夫_大吟醸_薫', 0.465287983417511)

Search sake data with queries

import pprint
api = utils.SearchAPI()
query1 = "brand+name:美丈夫_大吟醸_薫"
query2 = "dgree_of_sweetness/dryness:-0.08"
query3 = "rice_polishing_rate:40"
results = api.and_search(query1, query2, query3)
pprint.pprint(results[0], width=40)

Output

{'alcohol_rate': {'max': '',
                  'mean': '',
                  'min': ''},
 'amino_acid_content': {'max': '',
                        'mean': '',
                        'min': ''},
 'brand': '美丈夫',
 'brand+name': '美丈夫 大吟醸 薫',
 'brewer': '濵川商店',
 'city': '安芸郡田野町',
 'dgree_of_sweetness/dryness': '-0.08',
 ...
 }

Examples of sake-operation

• Input sake: "金鵄正宗_純米大吟醸_祝"

# About "金鵄正宗_純米大吟醸_祝"

{'alcohol_rate': {'max': '16.00',
                  'mean': '16.50',
                  'min': '17.00'},
 'amino_acid_content': {'max': '1.10',
                        'mean': '1.10',
                        'min': '1.10'},
 'brand': '金鵄正宗',
 'brand+name': '金鵄正宗 純米大吟醸 祝',
 'brewer': 'キンシ正宗',
 'city': '京都市伏見区',
 'dgree_of_sweetness/dryness': '0.11',
 'method_for_making_sake': [],
 'name': '純米大吟醸 祝',
 'prefecture': '京都府',
 'rice': ['祝'],
 'rice_polishing_rate': '45',
 'sake_class': '純米大吟醸',
 'sake_meter_value': {'max': '1.00',
                      'mean': '1.00',
                      'min': '1.00'},
 'titratable_acidity': {'max': '1.20',
                        'mean': '1.20',
                        'min': '1.20'},
 'yeast': []}

Operation of sake vectors: Change rice

• [brand+name] - [rice A] + [rice B]
• "金鵄正宗_純米大吟醸_祝" - "祝 (rice)" + "山田錦 (rice)" = "松屋久兵衛"
• "祝" -> "山田錦"
• "松屋久兵衛" is a sake brewed by "キンシ正宗", which is the same brewery as "金鵄正宗_純米大吟醸_祝", but the rice is changed as "山田錦"

import utils
import pprint
model = utils.load_sake_embedding('model.txt')
api = utils.SearchAPI()

# Analogy: change the type of rice
result = model.most_similar(positive=['brand+name:金鵄正宗_純米大吟醸_祝', 
                                      'rice:山田錦'], negative=['rice:祝'], topn=3)

[('brand:松屋久兵衛', 0.45839840173721313), # most similar word
 ('brand:切子', 0.4513291120529175),
 ('brand+name:北洋_袋取り雫酒', 0.44058412313461304)]

Search about 'brand:松屋久兵衛'

results = api.and_search('brand:松屋久兵衛')
pprint.pprint(results[0], width=40)

{'alcohol_rate': {'max': '16.00',
                 'mean': '16.50',
                 'min': '17.00'},
'amino_acid_content': {'max': '1.10',
                       'mean': '1.10',
                       'min': '1.10'},
'brand': '松屋久兵衛',
'brand+name': '松屋久兵衛 ',
'brewer': 'キンシ正宗',       # Same brewer as 'brand+name:金鵄正宗_純米大吟醸_祝'
'city': '京都市伏見区',
'dgree_of_sweetness/dryness': '0.11',
'method_for_making_sake': [],
'name': '',
'prefecture': '京都府',
'rice': ['山田錦'],          # Type of the rice is changed by the analogy
'rice_polishing_rate': '35',
'sake_class': '純米大吟醸',
'sake_meter_value': {'max': '1.00',
                     'mean': '1.00',
                     'min': '1.00'},
'titratable_acidity': {'max': '1.20',
                       'mean': '1.20',
                       'min': '1.20'},
'yeast': []}

Operation of sake vectors: Change rice polishing rate

• [brand+name] - [rice_polishing_rate A] + [rice_polishing_rate B]
• 45% -> 60%

import utils
import pprint
model = utils.load_sake_embedding('model.txt')
api = utils.SearchAPI()

# Analogy: change the rice polishing rate
result = model.most_similar(positive=['brand+name:金鵄正宗_純米大吟醸_祝', 'rice_polishing_rate:60'], 
                            negative=['rice_polishing_rate:45'], topn=3)

[('brand+name:金鵄正宗_特別純米', 0.6118292808532715), # most similar word
 ('brand+name:金閣_荒武者', 0.571118950843811),
 ('brand+name:松屋久兵衛', 0.566016674041748)]

Search about 'brand+name:金鵄正宗_特別純米'

results = api.and_search('brand+name:金鵄正宗_特別純米')
pprint.pprint(results[0], width=40)

{'alcohol_rate': {'max': '15.00',
                  'mean': '15.50',
                  'min': '16.00'},
 'amino_acid_content': {'max': '1.30',
                        'mean': '1.30',
                        'min': '1.30'},
 'brand': '金鵄正宗',
 'brand+name': '金鵄正宗 特別純米',
 'brewer': 'キンシ正宗',       # Same brewer as 'brand+name:金鵄正宗_純米大吟醸_祝'
 'city': '京都市伏見区',
 'dgree_of_sweetness/dryness': '-0.17',
 'method_for_making_sake': [],
 'name': '特別純米',
 'prefecture': '京都府',
 'rice': ['五百万石'],
 'rice_polishing_rate': '60', # The rice polishing rate is changed
 'sake_class': '特別純米',
 'sake_meter_value': {'max': '1.50',
                      'mean': '1.50',
                      'min': '1.50'},
 'titratable_acidity': {'max': '1.40',
                        'mean': '1.40',
                        'min': '1.40'},
 'yeast': []}