sensai/examples/tangram.ipynb

In [1]:

import os
from os.path import join
from glob import glob
import pandas as pd
from sentence_transformers import SentenceTransformer
import tangram

DATASET_DIR = os.environ.get("DATASET_DIR", "../input/sensai")
print("DATASET_DIR", DATASET_DIR)

DATASET_DIR /home/uetchy/repos/src/github.com/holodata/sensai-huggingface

In [2]:

# load tokenizer
st = SentenceTransformer('all-MiniLM-L6-v2')

In [3]:

print('Loading data...')
df = pd.concat(
    [pd.read_parquet(x) for x in glob(join(DATASET_DIR, '*.parquet'))],
    ignore_index=True)

Loading data...

In [8]:

print('Sampling data...')

# Chats with body size larger than 1
# df = df[df['body'].str.len() > 1]

# Merge labels
df['label'] = df['label'].apply(lambda x: {
    'deleted': 'toxic',
    'hidden': 'toxic',
    'nonflagged': 'safe'
}[x])

# Drop duplicates
df = df.drop_duplicates()
df.info()

# Balance item count for each category
g = df.groupby('label')
sample_size = min(g.size().min(), 4_000_000)
print('sample_size', sample_size)
df = g.apply(lambda x: x.sample(sample_size)).reset_index(drop=True)

df.info()
print(df.groupby('label').size())
df.sample(5)

Out [8]:

Sampling data...
<class 'pandas.core.frame.DataFrame'>
Int64Index: 9121058 entries, 0 to 12942624
Data columns (total 3 columns):
 #   Column      Dtype 
---  ------      ----- 
 0   authorName  object
 1   body        object
 2   label       object
dtypes: object(3)
memory usage: 278.4+ MB
sample_size 3647717
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 7295434 entries, 0 to 7295433
Data columns (total 3 columns):
 #   Column      Dtype 
---  ------      ----- 
 0   authorName  object
 1   body        object
 2   label       object
dtypes: object(3)
memory usage: 167.0+ MB
label
safe     3647717
toxic    3647717
dtype: int64

	authorName	body	label
2646589	それもえ	ナイスー	safe
6722778	LassieOW	K U S A K E N ? ? ? K U S A K E N ? ? ? K U S...	toxic
4490817	ヤルダバオトここに	猫左下	toxic
522319	IkeShamus	5	safe
6359809	susi Susanti	keigo tsundere	toxic

In [9]:

print('Formatting data...')

# Add counting features
# df['blen'] = df['body'].str.len()

# Concat author name and message with a separator
df['body'] = df.apply(lambda x: (x['authorName'] or '<EMPTY>') + '<SEP>' + x['body'], axis=1)
df.drop(columns=['authorName'], inplace=True)

df.info()
df.sample(5)

Out [9]:

Formatting data...
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 7295434 entries, 0 to 7295433
Data columns (total 2 columns):
 #   Column  Dtype 
---  ------  ----- 
 0   body    object
 1   label   object
dtypes: object(2)
memory usage: 111.3+ MB

	body	label
6998597	allormist<SEP>demon slayer copied this idea	toxic
2536692	Rei<SEP>ああああああああ	safe
3713136	Alex<SEP>Look at the first part of the Quest. ...	toxic
5614434	Nov4Zyz<SEP>かわいい	toxic
517705	Aida<SEP>lmaoooo	safe

In [13]:

print('Calculating embedding vectors...')

emb = pd.DataFrame(st.encode(df['body'].to_list(), convert_to_numpy=True))
df = pd.concat([df, emb], axis=1).drop(columns=['body'])

df.head()

Out [13]:

Calculating embedding vectors...

	label	0	1	2	3	4	5	6	7	8	...	374	375	376	377	378	379	380	381	382	383
0	safe	0.009612	0.052424	0.083845	-0.031595	0.028893	0.059027	0.045335	-0.014343	0.037504	...	-0.088224	0.037281	0.020476	0.067415	0.000324	-0.049588	0.006397	0.139169	0.049218	-0.025690
1	safe	0.003328	0.114213	0.032118	-0.055552	-0.008358	0.024322	0.080483	0.007664	0.062874	...	-0.000377	0.046901	0.018417	-0.003170	-0.027469	0.005639	0.133500	0.032221	0.030415	-0.019664
2	safe	0.031809	-0.024030	0.026037	-0.039061	-0.002786	0.052565	0.126942	-0.014587	0.047931	...	-0.014579	0.009263	0.010371	0.045266	-0.076779	-0.023678	0.021364	0.036397	-0.001076	-0.060540
3	safe	-0.061874	0.046060	0.021520	0.021778	0.009127	0.021292	0.184221	0.004618	-0.069447	...	0.000147	0.002003	0.089495	0.040357	0.007321	0.016096	0.012340	0.062434	0.005810	0.003983
4	safe	-0.053528	-0.045918	0.076154	-0.003261	0.005603	-0.058807	0.071698	0.005038	0.010522	...	0.077442	-0.080605	-0.038363	0.013772	-0.049518	-0.004730	0.142463	-0.017870	0.013901	-0.037920

5 rows × 385 columns

In [14]:

df.info()
df.to_csv('./data/input_7m_balanced_deduped_amc_MiniLM.csv', index=False)

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 7295434 entries, 0 to 7295433
Columns: 385 entries, label to 383
dtypes: float32(384), object(1)
memory usage: 10.5+ GB

In [ ]:

!tangram train -f ./data/input_7m_balanced_deduped_amc_MiniLM.csv -o ./models/sensai_7m_balanced_deduped_amc_MiniLM.tangram -t label

In [4]:

model = tangram.Model.from_path('./models/sensai_7m_balanced_deduped_amc_MiniLM.tangram')

def remap_label(pred) -> str:
    label = '👍'
    if pred.class_name == 'toxic':
        if pred.probability <= 0.6:
            label = '⚖️'
        else:
            label = '🚽'
    return label


def predict_batch(texts):
    preds = model.predict([{str(k): v for k,v in enumerate(x)} for x in st.encode(texts)])
    return [(remap_label(pred), pred.probability) for pred in preds]

def predict(text: str):
    return predict_batch([text])[0]

In [ ]:

chats = [
    "lol",
    "草",
    "かわいい！！！", # cute !!!
    "INAAAAAAAAAA",
    "[ENG] Heh ? don't I have way less recovery items ? don't they come back on each continue ????",
    "What's with this sassy lost child?",
    "Your village looks so cute and pretty",
    "imagine the dead team member watching him dance for no reason",
    "A型だけど、気にならないどころかむしろ安心感ある",
    "Tenchou pls even if you don't think your singing is as good, you are still one of the most electric and infectious souls I've ever seen. Ganbatte bossu!!!",
    "Nice butt",
    "My dog died let's goooooooooooo",
    "when will you graduate?",
    "lol you guys are pathetic",
    "死ね死ね死ね死ね死ね死ね死ね死ね死ね死ね死ね死ね死ね死ね死ね", # die die die die die die
    "卒業卒業卒業卒業", # graduate graduate graduate graduate
]

for text, pred in zip(chats, predict_batch(chats)):
    print(pred[0], ':', text)

👍 : lol
🚽 : 草
👍 : かわいい！！！
👍 : INAAAAAAAAAA
🚽 : [ENG] Heh ? don't I have way less recovery items ? don't they come back on each continue ????
👍 : What's with this sassy lost child?
👍 : Your village looks so cute and pretty
⚖️ : imagine the dead team member watching him dance for no reason
👍 : A型だけど、気にならないどころかむしろ安心感ある
🚽 : Tenchou pls even if you don't think your singing is as good, you are still one of the most electric and infectious souls I've ever seen. Ganbatte bossu!!!
🚽 : Nice butt
🚽 : My dog died let's goooooooooooo
👍 : when will you graduate?
⚖️ : lol you guys are pathetic
🚽 : 死ね死ね死ね死ね死ね死ね死ね死ね死ね死ね死ね死ね死ね死ね死ね
🚽 : 卒業卒業卒業卒業

In [5]:

deleted_by_mods = df[df['label'] != 'nonflagged']
safe = df[df['label'] == 'nonflagged']
print('# of deleted chats', len(deleted_by_mods))
print('# of safe chats', len(safe))

# of deleted chats 6471313
# of safe chats 6471313

In [6]:

N = 50000

def adjust_label(pred):
    if pred.class_name == 'safe':
        if pred.probability <= 0.5:
            return 'toxic'
        else:
            return 'safe'
    return 'toxic'

def predict_labels(chats):
    return [(x.class_name, x.probability) for x in model.predict([{str(k): v for k,v in enumerate(x)} for x in st.encode(chats.apply(lambda x: (x['authorName'] or '<EMPTY>') + '<SEP>' + x['body'], axis=1).to_list())])]

deleted_labels = predict_labels(deleted_by_mods.sample(N))
safe_labels = predict_labels(safe.sample(N))

In [7]:

def f_beta(prec, rec, beta):
    return (1+beta**2) * ( (prec * rec) / ( beta**2 * prec + rec ) )

fn = len([x for x in deleted_labels if x[0] == 'safe'])
fp = len([x for x in safe_labels if x[0] != 'safe'])
tp = N - fn
tn = N - fp
pp = tp + fp
fnr = fn / N
fpr = fp / N
sensitivity = tp / N
precision = tp / pp
accuracy = (tp + tn) / (N*2)
fb = f_beta(precision, sensitivity, 2)
print('N', len(deleted_labels)+len(safe_labels))
print('FPR', fpr)
print('FNR', fnr)
print('Sensitivity', sensitivity)
print('Precision', precision)
print('Accuracy', accuracy)
print('F(2) score', fb)

N 100000
FPR 0.14654
FNR 0.181
Sensitivity 0.819
Precision 0.8482300060070013
Accuracy 0.83623
F(2) score 0.8246837201996158

In [23]:

predict_labels(pd.DataFrame.from_dict([{'authorName': 'Skamor', 'body': '味ってわかる？"あじ"って読むんやで'}]))

Out [23]:

[('safe', 0.5598429441452026)]

In [26]:

pd.DataFrame.from_dict([{'authorName': 'Hi', 'body': 'Hi'}])

Out [26]:

	authorName	body
0	Hi	Hi

In [111]:

import json

def convertRawMessageToString(rawMessage):
    def handler(run):
        msgType = list(run.keys())[0]
        payload = run[msgType]
        if msgType == 'text':
            return payload
        elif msgType == 'emoji':
            if 'isCustomEmoji' in payload:
                return "\uFFFD"
            else:
                return payload['emojiId']
        else:
            raise 'Invalid type: ' + msgType + ', ' + payload
    return "".join([handler(run) for run in rawMessage])

with open('addChatItemAction.jsonl') as f:
    lines = list(map(json.loads, f.readlines()))
chats = pd.DataFrame.from_dict(lines)
chats = chats[['authorName', 'message']]
chats['message'] = chats['message'].apply(convertRawMessageToString)
chats = chats.rename(columns={'message': 'body'})
chats['label'] = pd.Series(map(lambda x: 'hold' if x[1] < 0.85 else {'safe': 'safe', 'toxic': 'annoying'}[x[0]], predict_labels(chats)))

In [106]:

chats[(chats['label'] == 'safe') & (chats['body'].str.len() > 3)].sample(20)

Out [106]:

	authorName	body	label
3690	Miguel Islas	lmao	safe
451	mamin	ㅋㅋㅋㅋㅋㅋ	safe
7977	Pierah	🧡💛💙💜❤	safe
1993	Kuga Raian	LMAOOOO😳😳😳	safe
5752	BOY tenten	こんにちはー	safe
9304	linkis20	💙💙💙💙💙	safe
10540	MegaSlayer92	💙💙💙💙💙💙	safe
5562	Brandermau	You're welcome	safe
8890	零羅	かわいい	safe
8155	Kronii's Apostle [8th of the Twelve]	ayy pog moment	safe
7414	linkis20	💙❤🧡💜💛	safe
7976	LongtimeNoc-KFP-Nephamily	Is happy to make one of your dreams finally co...	safe
10969	Bishop Johnson	💙💙💙💙	safe
7504	Arthur	❤🧡💛💙💜	safe
9343	Nova Star	You did great	safe
544	BreakingPhobia103	FEET�	safe
11250	Oliver Chen	LOOOOOOOOOL	safe
2791	Kronii's Apostle [8th of the Twelve]	is mumei really that beeg?	safe
9991	linkis20	💙💙💙💙💙	safe
7843	»pingpong	💙💛💜❤🧡	safe