Back office ops · Production

Google Docs auto-generated document summaries powered by a Pegasus-based ML model

The problem

Document readers struggle to keep up with the volume of incoming documents daily, while writers find composing summaries cognitively challenging and time-consuming.

First attempt

Early fine-tuning corpora had inconsistencies and high variation across document and summary types, causing the model to be easily confused and unable to learn document-summary relationships.

Workflow diagram · grounded in source

Document opened in Docs

trigger

“A blue summary icon appears in the top left corner when a document summary suggestion is available”

ML model comprehends and generates

ai_action

“a machine learning (ML) model that comprehends document text and, when confident, generates a 1-2 sentence natural language description of the document content”

Writer reviews suggestion

human_review

“the document writer maintains full control — accepting the suggestion as-is, making necessary edits to better capture the document summary or ignoring the suggestion altogether”

Summary aids document navigation

output

“Readers can also use this section, along with the outline, to understand and navigate the document at a high level”

Reported outcome

Google deployed a distilled Pegasus-based summarization model in Google Docs for Workspace business customers, generating 1-2 sentence summary suggestions with significant improvements in serving latency and memory footprint.

Reported metrics

Serving latency and memory footprintsignificant improvements in latency and memory footprint

Model quality after distillationquality was still on par with the original model

fine-tuning examples needed (Pegasus vs Transformer baseline)as few as 1,000

Reported stack

Google DocsPegasusTransformerBERTGPTT5TPUsNLUNLG

Source

https://ai.googleblog.com/2022/03/auto-generated-summaries-in-google-docs.html

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Google deployed a distilled Pegasus-based summarization model in Google Docs for Workspace business customers, generating 1-2 sentence summary suggestions with significant improvements in serving latency and memory fo…

What tools did this team use?

Google Docs, Pegasus, Transformer, BERT, GPT, T5, TPUs, NLU, NLG.

What results were reported?

Serving latency and memory footprint: significant improvements in latency and memory footprint; Model quality after distillation: quality was still on par with the original model; fine-tuning examples needed (Pegasus vs Transformer baseline): as few as 1,000 (source-reported, not independently verified).

What failed first in this deployment?