# US 2020/0401661 A1 · Session embeddings for summarizing activity

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2020/0401661 A1
- Document ID: US20200401661A1
- Status: Published application
- Kind: A1
- Application number: 16/446,349
- Filed: 2019-06-19
- Published: 2020-12-24
- Earliest claimed: 2019-06-19
- Assignees: Microsoft Technology Licensing, LLC
- Applicants: Microsoft Technology Licensing, LLC
- Inventors: Nagaraj Kota; Venkatesh Duppada
- Web page: https://patentgrep.com/patent/US20200401661A1

## CPC

- G06F 40/30: Semantic analysis
- G06F 16/24578: {using ranking}
- G06F 16/9535: Search customisation based on user profiles and personalisation
- G06F 40/284: Lexical analysis, e.g. tokenisation or collocates
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/0895: Weakly supervised learning, e.g. semi-supervised or self-supervised learning
- G06N 3/09: Supervised learning
- G06N 20/00: Machine learning
- G06Q 10/1053: Employment or hiring
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06Q: INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR

## Abstract

The disclosed embodiments provide a system for processing data. During operation, the system obtains a first sentence representing a first sequence of actions between a user and a set of jobs. Next, the system applies a language model to token embeddings of a first set of tokens in the first sentence and position embeddings of token positions in the first sentence to produce a first set of output embeddings. The system then combines the first set of output embeddings into a first session embedding that encodes the first sequence of actions. Finally, the system outputs the first session embedding for use in characterizing job-seeking activity of the user.

## Claims

### Claim 1

A method, comprising:
obtaining a first sentence representing a first session with a user and a second sentence representing a second session with the user, wherein each of the first and second sessions comprises a sequence of actions between the user and a set of jobs;
applying, by one or more computer systems, a language model to token embeddings of a first set of tokens in the first sentence and position embeddings of token positions in the first sentence to produce a first session embedding for the first session;
applying, by the one or more computer systems, the language model to the token embeddings of a second set of tokens in the second sentence and the position embeddings of the token positions in the second sentence to produce a second session embedding for the second session; and
outputting the first and second session embeddings for use in characterizing job-seeking activity of the user.

### Claim 2 (depends on 1)

The method of claim 1, further comprising:
combining the first and second session embeddings into a session history embedding representing the first and second sessions.

### Claim 3 (depends on 2)

The method of claim 2, wherein combining the first and second session embeddings into the session history embedding comprises:
generating the session history embedding as an element-wise maximum of the first and second session embeddings.

### Claim 4 (depends on 2)

The method of claim 2, wherein combining the first and second session embeddings into the session history embedding comprises:
inputting the first and second session embeddings into a sequence model; and
obtaining the session history embedding as output from the sequence model.

### Claim 5 (depends on 2)

The method of claim 2, wherein combining the first and second session embeddings into the session history embedding comprises:
identifying a subset of tokens in the first and second sentences that represent a subset of actions in the first and second sessions; and
aggregating output embeddings generated by the language model from the subset of tokens into the session history embedding.

### Claim 6 (depends on 1)

The method of claim 1, further comprising:
obtaining a third sentence representing a third session with the user and a fourth sentence representing a fourth session with the user; and
applying the language model to the token embeddings of a third set of tokens in the third sentence and a fourth set of tokens in the fourth sentence, a separator embedding of a separator token between the third and fourth sets of tokens, a first segment embedding for the third sentence, a second segment embedding for the fourth sentence, and the position embeddings of the token positions in the third and fourth sentences to produce a session history embedding.

### Claim 7 (depends on 1)

The method of claim 1, further comprising:
training the language model to predict at least one of:
masked tokens in a set of sample sentences; and
a next sentence in the set of sample sentences from a current sentence in the set of sample sentences.

### Claim 8 (depends on 1)

The method of claim 1, wherein outputting the first and second session embeddings for use in characterizing the job-seeking activity of the user comprises:
inputting the first and second session embeddings and additional features related to a job into a machine learning model;
receiving, as output from the machine learning model, a score representing a likelihood of a positive outcome between the user and the job; and
generating a recommendation related to the job based on the score.

### Claim 9 (depends on 8)

The method of claim 8, wherein the recommendation comprises at least one of:
a query suggestion for a job search by the user; and
a recommendation of the job to the user.

### Claim 10 (depends on 1)

The method of claim 1, wherein the first session embedding comprises an output embedding generated by the language model from a classification token inputted into the language model.

### Claim 11 (depends on 1)

The method of claim 1, wherein the language model comprises a bidirectional transformer encoder.

### Claim 12 (depends on 1)

The method of claim 1, wherein the first and second sentences comprise at least one of:
an action type;
a job identifier;
a title;
a company;
a location; and
a search term; and
a keyword.

### Claim 13

A system, comprising:
one or more processors; and
memory storing instructions that, when executed by the one or more processors, cause the system to:
obtain a first sentence representing a first sequence of actions between a user and a set of jobs;
apply a language model to token embeddings of a first set of tokens in the first sentence and position embeddings of token positions in the first sentence to produce a first set of output embeddings;
combine the first set of output embeddings into a first session embedding that encodes the first sequence of actions; and
output the first session embedding for use in characterizing job-seeking activity of the user.

### Claim 14 (depends on 13)

The system of claim 13, wherein the memory further stores instructions that, when executed by the one or more processors, cause the system to:
apply the language model to a second sentence representing a second sequence of actions between the user and an additional set of jobs to produce a second set of output embeddings;
combine the second set of output embeddings into a second session embedding that encodes the second sequence of actions; and
combine the first and second session embeddings into a session history embedding representing the first and second sequences of actions.

### Claim 15 (depends on 14)

The system of claim 14, wherein combining the first and second session embeddings into the session history embedding comprises:
inputting the first and second session embeddings into a sequence model; and
obtaining the session history embedding as output from the sequence model.

### Claim 16 (depends on 13)

The system of claim 13, wherein outputting the first session embedding for use in characterizing the job-seeking activity of the user comprises:
inputting the first session embedding and additional features related to a job into a machine learning model;
receiving, as output from the machine learning model, a score representing a likelihood of a positive outcome between the user and the job; and
generating a recommendation related to the job based on the score.

### Claim 17 (depends on 16)

The system of claim 16, wherein the recommendation comprises at least one of:
a query suggestion for a job search by the user; and
a recommendation of the job to the user.

### Claim 18 (depends on 13)

The system of claim 13, wherein the language model comprises a bidirectional transformer encoder.

### Claim 19 (depends on 13)

The system of claim 13, wherein the first session embedding comprises an output embedding generated by the language model from a classification token inputted into the language model.

### Claim 20

A non-transitory computer-readable storage medium storing instructions that when executed by a computer cause the computer to perform a method, the method comprising:
obtaining a first sentence representing a first session with a user and a second sentence representing a second session with the user, wherein each of the first and second sessions comprises a sequence of actions performed by the user with an online system;
applying a language model to token embeddings of a first set of tokens in the first sentence and position embeddings of token positions in the first sentence to produce a first session embedding for the first session, wherein the first session embedding comprises an output embedding generated by the language model from a classification token added to the first sentence;
applying the language model to the token embeddings of a second set of tokens in the second sentence and the position embeddings of the token positions in the second sentence to produce a second session embedding for the second session; and
outputting the first and second session embeddings for use in characterizing activity of the user.