Thanks to visit codestin.com
Credit goes to www.scribd.com

Scribd
Upload
102 views29 pages

DMT Video Reivew Process

The document summarizes the video review process at Dolphin. It involves triggering a review, retrieving data, calculating risk scores, applying review rules, and outputting the results. The process uses various services like Kafka, nsq, and RPC to distribute tasks. It loads configurations, retrieves data from sources like callback APIs and databases, runs the content through risk prediction models, and sends it to reviewers or filters based on the scores and rules. Pipelines are configured to orchestrate the various steps.

Uploaded by

Wei Dai
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
102 views29 pages

DMT Video Reivew Process

The document summarizes the video review process at Dolphin. It involves triggering a review, retrieving data, calculating risk scores, applying review rules, and outputting the results. The process uses various services like Kafka, nsq, and RPC to distribute tasks. It loads configurations, retrieves data from sources like callback APIs and databases, runs the content through risk prediction models, and sends it to reviewers or filters based on the scores and rules. Pipelines are configured to orchestrate the various steps.

Uploaded by

Wei Dai
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 29

DMT Video Review Process

赵新晨

2020.04.17
Overview Dolphin

aweme.review.content Risk Predict Gorilla


Trigger MGetObjectData

Kafka / nsq

data
Event Handler
Load Conf Retrieve Data Calculate Risk Apply Rule Output TCS SetTaskResult

RPC
Pipeline
EnqueueID
SendToReview

First Y Y First
SetToReviewRisk Review && Video Quality Filter Review &&
Pass Pass

N N

Review Judge
Trigger & Event Handler Dolphin

aweme.review.content Risk Predict Gorilla


Trigger MGetObjectData

Kafka / nsq

data
Event Handler
Load Conf Retrieve Data Calculate Risk Apply Rule Output TCS SetTaskResult

RPC
Pipeline
EnqueueID
SendToReview

First Y Y First
SetToReviewRisk Review && Video Quality Filter Review &&
Pass Pass

N N

Review Judge
Trigger & Event Handler
Recall Review Third Review
First Review Report Review Circulation Review
Second Review Hot Review

ugc_item_change counter service feedback msg counter service

nsq Kafka nsq Kafka

filter_first_review counter_review_to_pipeline mt_review_feedback mt_circulation_review

nsq RPC: SendToReview RPC: SendToReview RPC: SendToReview

send_first_review

RPC: SendToReview

aweme.review.platform
Send To Review Dolphin

aweme.review.content Risk Predict Gorilla


Trigger MGetObjectData

Kafka / nsq

data
Event Handler
Load Conf Retrieve Data Calculate Risk Apply Rule Output TCS SetTaskResult

RPC
Pipeline
EnqueueID
SendToReview

First Y Y First
SetToReviewRisk Review && Video Quality Filter Review &&
Pass Pass

N N

Review Judge
Send To Review

Trigger

RPC: SendToReview
Kafka / nsq

Event Handler aweme.review.platform

RPC nsq
retry
SendToReview aweme.script.review_platform

EnqueueID RPC: EnqueueID

content.review.review_queue

Pipeline
Pipeline Conf Dolphin

aweme.review.content Risk Predict Gorilla


Trigger MGetObjectData

Kafka / nsq

data
Event Handler
Load Conf Retrieve Data Calculate Risk Apply Rule Output TCS SetTaskResult

RPC
Pipeline
EnqueueID
SendToReview

First Y Y First
SetToReviewRisk Review && Video Quality Filter Review &&
Pass Pass

N N

Review Judge
Pipeline Conf

• retrievers
• review_callback

• etcd

• frame_extract

• risk_predict

• kms_encoder

• …

• output_adapters
• tcs

• risk_predict

• fatal

• record_handler
• diting

• qps_quota
Retrieve Data Dolphin

aweme.review.content Risk Predict Gorilla


Trigger MGetObjectData

Kafka / nsq

data
Event Handler
Load Conf Retrieve Data Calculate Risk Apply Rule Output TCS SetTaskResult

RPC
Pipeline
EnqueueID
SendToReview

First Y Y First
SetToReviewRisk Review && Video Quality Filter Review &&
Pass Pass

N N

Review Judge
Pipeline Conf

• retrievers
• review_callback

• etcd

• frame_extract

• kms_encoder

Calculate Risk

Risk Model Set Risk Model Set Risk Model Set

mt_video_create_m
lab.drug.musically lab.abuse_animals.tiktok
lab.belly_dance.musically_v2 lab.sensitive_flag.mtv … …
lab.sensitive_text.english lab.bloody.musically
…

aweme.review.content Risk Predict Gorilla Dolphin

Load Conf Retrieve Data Calculate Risk Apply Rule Output

Pipeline
Pipeline Conf

• retrievers
• risk_predict

Apply Rule Dolphin

aweme.review.content Risk Predict Gorilla


Trigger MGetObjectData

Kafka / nsq

data
Event Handler
Load Conf Retrieve Data Calculate Risk Apply Rule Output TCS SetTaskResult

RPC
Pipeline
EnqueueID
SendToReview

First Y Y First
SetToReviewRisk Review && Video Quality Filter Review &&
Pass Pass

N N

Review Judge
Pipeline Rule (without Gorilla)

• Input
• risk predict results

• item details

• author details

• …

• Output
• queue

• to_review_risk

• rule_results

Gorilla

• Input
• risk predict results

• item details

• author details

• …

• Output
• queue

• to_review_risk

• rule_results

Try it out: https://ares.bytedance.net/gorilla/


Pipeline Output Dolphin

aweme.review.content Risk Predict Gorilla


Trigger MGetObjectData

Kafka / nsq

data
Event Handler
Load Conf Retrieve Data Calculate Risk Apply Rule Output TCS SetTaskResult

RPC
Pipeline
EnqueueID
SendToReview

First Y Y First
SetToReviewRisk Review && Video Quality Filter Review &&
Pass Pass

N N

Review Judge
Pipeline Conf

• output_adapters
• tcs

• targets

• templates

TCS
Pipeline Output Dolphin

aweme.review.content Risk Predict Gorilla


Trigger MGetObjectData

Kafka / nsq

data
Event Handler
Load Conf Retrieve Data Calculate Risk Apply Rule Output TCS SetTaskResult

RPC
Pipeline
EnqueueID
SendToReview

First Y Y First
SetToReviewRisk Review && Video Quality Filter Review &&
Pass Pass

N N

Review Judge
Pipeline Conf

• output_adapters
• callback

Pipeline Conf

• fatal
• tcs

• callback

• qps_quota
Pipeline Output Dolphin

aweme.review.content Risk Predict Gorilla


Trigger MGetObjectData

Kafka / nsq

data
Event Handler
Load Conf Retrieve Data Calculate Risk Apply Rule Output TCS SetTaskResult

RPC
Pipeline
EnqueueID
SendToReview

First Y Y First
SetToReviewRisk Review && Video Quality Filter Review &&
Approve Pass

N N

Review Judge
Video Quality Filter

Status/Rate Rules

duration_cheat

Pass Rules Post Actions


Filter Rules fake_seed_item
is_ad watermark_recheck
is_ad
request request request static_frame_detection request status / rate
quality_reviewed record_op_history Review
is_not_video Judge
watermark
encode_abnormal update_reason_code
…
hawk_review_hash_tag
… …
filtered tag_mohu

Rule Definition：https://bytedance.feishu.cn/wiki/wikcnkkaHGSFIrlGKtrAYSE5UKe
Review Judge Dolphin

aweme.review.content Risk Predict Gorilla


Trigger MGetObjectData

Kafka / nsq

data
Event Handler
Load Conf Retrieve Data Calculate Risk Apply Rule Output TCS SetTaskResult

RPC
Pipeline
EnqueueID
SendToReview

First Y Y First
SetToReviewRisk Review && Video Quality Filter Review &&
Approve Pass

N N

Review Judge
Video Status :
• 22：user deleted
Review Judge (D Status) • 141：violation & reviewed
• 144：visible to self
• A strictest • 140：private
• A1 latest • 143：visible to friend
• A1-1： pre-moderation punish strictest • 102：normal
• pre-moderation for video
• pre-moderation for background sound
• A1-2：report
• A1-3：vv recall
• video recall
• background sound recall
requested status final status
• A1-4：comment recall
• A1-5：duet/react unpass video recall
• A1-6：second moderation forwarding to porn/reaction
• A1-7：similar video dedup
• A1-8：pre-moderation
• A2：appeal, operation tools, double check for unpass video of big V
• B：users operation
• C：infringement
• D：music
• E：account
• X：self-see during moderation
Review Judge (D Rate) Video Rate:
• 10：not recommended
• 12：fresh
• P0
• operation tools

• rate 10 during moderation

• Lark assistant tools


P3: rate = 10 Review Judge final rate = 10
• appeal
P2: rate = 12
• P1
• similar video dedup

• manual moderation (second, third, hot)

• P2 P1: rate = 12 Review Judge final rate = 12

• video quality filter


P2: rate = 10
• P3
• initial rate

• associative punish
Documentation

• Content Safety Playbook - https://bytedance.feishu.cn/wiki/wikcnKrpHWeKbWePdBa5uYg4h0c

• Pipeline Docs - https://bytedance.feishu.cn/docs/cIESARRJg58S2QaQ2p7VCe

• TCS - https://press.bytedance.net/tcs/

• Jimu - https://doc.bytedance.net/docs/473/

• Video Quality Filter - https://bytedance.feishu.cn/wiki/wikcnkkaHGSFIrlGKtrAYSE5UKe

• Review Judge

• Douyin Status - https://bytedance.feishu.cn/docs/doccnhjyo0WG0GomFYa9y6zwr6c

• Douyin Rate - https://bytedance.feishu.cn/docs/doccnjfRteMAHz89KTvWUKm3xlg

• MT - https://bytedance.feishu.cn/docs/CoRZo7ehHKCvyjLsuLAmHd#YwOcKM

• Key Terms on Mod. Tool & Platforms - https://bytedance.feishu.cn/sheets/shtcncBfuCV4qKo2rtxXZ0YPCNb#624385


Q&A
Q&A Dolphin

aweme.review.content Risk Predict Gorilla


Trigger MGetObjectData

Kafka / nsq

data
Event Handler
Load Conf Retrieve Data Calculate Risk Apply Rule Output TCS SetTaskResult

RPC
Pipeline
EnqueueID
SendToReview

First Y Y First
SetToReviewRisk Review && Video Quality Filter Review &&
Pass Pass

N N

Review Judge

You might also like