# US 12,112,131 B2 · Systems and methods for factual extraction from language model

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

- Number: US 12,112,131 B2
- Document ID: US12112131B2
- Status: Granted
- Kind: B2
- Application number: 17/588,043
- Filed: 2022-01-28
- Published: 2023-03-16
- Granted: 2024-10-08
- Earliest claimed: 2021-09-10
- Assignees: Salesforce, Inc.
- Applicants: Salesforce, Inc.
- Inventors: Benjamin Newman; Nazneen Rajani; Prafulla Kumar Choubey
- Web page: https://patentgrep.com/patent/US12112131B2

## CPC

- G06F 40/279: Recognition of textual entities
- G06F 40/126: Character encoding
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/045: Combinations of networks
- G06N 3/096: Transfer learning
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Embodiments described herein provide a system and method for extracting factual information. The system transforms a query into a natural language prompt in a format of a query subject and a queried relation. The system encodes, via an embedding layer of a pre-trained language model, the natural language prompt into a first embedding. The system encodes, via the adapter model, the first embedding into a second embedding based on a probability that the second embedding returns the factual information when the second embedding is fed the first attention layer of the pre-trained language model. The system decodes, by the first attention layer of the pre-trained language mode, the second embedding into a response to the query. The system extracts the factual information from the decoded response to the query.

## Claims

### Claim 1

A method for extracting factual information from a language model, the method comprising:
receiving, via a communication interface, a query for factual information;
encoding, via an embedding layer of a pre-trained language model, the natural language prompt into a first embedding;
encoding, via an adapter model the first embedding into a second embedding that includes a continuous representation based on a probability that the second embedding will return the factual information when the second embedding is fed to a first attention layer of the pre-trained language model, wherein the adapter model is placed between the embedding layer of the pre-trained language model and a first attention layer of the pre-trained language model;
decoding, via the first attention layer of the pre-trained language model, the second embedding into a response to the query; and
extracting the factual information from the decoded response to the query.

### Claim 2 (depends on 1)

The method of claim 1, wherein encoding, via the adapter model, the first embedding into a second embedding comprises;
determining, via an LSTM (Long Short Term Memory) layer, an LSTM encoding in response to the first embedding; and
determining, via a Multi-Layer Perceptron (MLP), the second embedding in response to the LSTM encoding.

### Claim 3 (depends on 1)

The method of claim 1, wherein encoding, via the adapter model, the first embedding into a second embedding comprises;
determining, via an LSTM (Long Short Term Memory) layer, an LSTM encoding based on the first embedding;
down-sampling the LSTM output, via a pooling layer; and
determining, via a Multi-Layer Perceptron (MLP), a part of the second embedding based on the down-sampled LSTM encoding.

### Claim 4 (depends on 3)

The method of claim 3, further comprising;
transmitting the first embedding as part of the second embedding to the first attention layer.

### Claim 5 (depends on 1)

The method of claim 1, wherein encoding, via the adapter model, the first embedding into a second embedding comprises:
determining a first entity and a second entity in the first embedding;
determining, via a classifier, a relationship between the first entity and the second entity; and
determining, via a look-up table, the second embedding associated with the relationship.

### Claim 6 (depends on 1)

The method of claim 1, wherein encoding, via the adapter model, the first embedding into a second embedding comprises;
determining a first entity and a second entity in the first embedding;
determining, via an oracle model, a relationship between the first entity and the second entity, wherein the oracle model returns a relationship given the first entity and the second entity; and
determining, via a look-up table, the second embedding associated with the relationship.

### Claim 7 (depends on 1)

The method of claim 1, wherein encoding, via the adapter model, the first embedding into a second embedding comprises:
determining, via the adapter model, a plurality of alternate embeddings based on the first embedding;
determining, via the adapter model, a probability associated with each of the plurality of alternate embeddings returning the factual information when fed to the first attention layer of the pre-trained language model; and
determining, via the adapter model, the second embedding based on the probability associated with each of the plurality of alternate embeddings.

### Claim 8 (depends on 1)

The method of claim 1, wherein encoding, via the adapter model, the first embedding into a second embedding comprises:
determining a first entity and a mask in the first embedding;
determining, via the adapter model, a plurality of alternate embeddings based on the position of the first entity and the mask;
determining, via the adapter model, a probability associated with each of the plurality of alternate embeddings returning the factual information when fed to the first attention layer of the pre-trained language model; and
determining, via the adapter model, the second embedding based on the probability associated with each of the plurality of alternate embeddings.

### Claim 9 (depends on 1)

The method of claim 1, further comprising:
receiving, via a communication interface, a training data set, the training dataset including a plurality of sets of similar natural language queries for factual information;
determining a first entity and a second entity in each query in a set of similar natural language queries from the plurality of sets of similar natural language queries;
masking at least the first entity or the second entity in each query in the set of similar natural language queries;
determining a masked embedding for each masked query in the set of similar natural language queries;
determining a response to each masked embedding via the pre-trained language model, via the first attention layer of the pre-trained language model;
determining whether the response matches the masked entity in the natural language query;
determine a probability that a masked embedding will return the response that matches the masked entity in the natural language query; and
update an adapter model that chooses a query where the response from the pre-defined language model matches the masked entity based on the determined probability.

### Claim 10

A system for extracting factual information from a language model, the system comprising:
a communication interface for receiving a query for factual information;
a memory storing a plurality of machine readable instructions; and
a processor reading and executing the instructions from the memory to perform operations comprising:
receive, via the communication interface, a query for factual information;
transform the query into a natural language prompt in a format of a query subject and a queried relation;
encode, via an embedding layer of a pre-trained language model, the natural language prompt into a first embedding;
encode, via the adapter model, the first embedding into a second embedding based on a probability that the second embedding returns the factual information when the second embedding is fed the first attention layer of the pre-trained language model;
decode, by the first attention layer of the pre-trained language mode, the second embedding into a response to the query; and
extract the factual information from the decoded response to the query.

### Claim 11 (depends on 10)

The system of claim 10, wherein the operation of encoding, via the adapter model, the first embedding into a second embedding comprises:
determine, via an LSTM (Long Short Term Memory) layer, an LSTM encoding based on the first embedding; and
determine, via a Multi-Layer Perceptron (MLP), the second embedding based on the LSTM encoding.

### Claim 12 (depends on 10)

The system of claim 10, wherein the operation of encoding, via the adapter model, the first embedding into a second embedding comprises:
determine, via an LSTM layer, an LSTM encoding based on the first embedding;
down-sample the LSTM output, via a pooling layer; and
determine, via a Multi-Layer Perceptron (MLP), a part of the second embedding based on the down-sampled LSTM encoding.

### Claim 13 (depends on 12)

The system of claim 12, wherein the operation of encoding, via the adapter model, the first embedding into a second embedding comprises:
transmit the first embedding as part of the second embedding to the first attention layer.

### Claim 14 (depends on 10)

The system of claim 10, wherein the operation of encoding, via the adapter model, the first embedding into a second embedding comprises:
determine a first entity and a second entity in the first embedding;
determine, via a classifier, a relationship between the first entity and the second; and
determine, via a look-up table, the second embedding associated with the relationship.

### Claim 15 (depends on 10)

The system of claim 10, wherein the operation of encoding, via the adapter model, the first embedding into a second embedding comprises:
determine a first entity and a second entity in the first embedding;
determine, via an oracle model, a relationship between the first entity and the second entity, wherein the oracle model returns a relationship given the first entity and the second entity; and
determine, via a look-up table, the second embedding associated with the relationship.

### Claim 16 (depends on 10)

The system of claim 10, wherein the operation of encoding, via the adapter model, the first embedding into a second embedding comprises: determining, via the adapter model, a plurality of alternate embeddings based on the first embedding;
determine, via the adapter model, a probability associated with each of the plurality of alternate embeddings returning the factual information when fed to the first attention layer of the pre-trained language model; and
determine, via the adapter model, the second embedding based on the probability associated with each of the plurality of alternate embeddings.

### Claim 17 (depends on 10)

The system of claim 10, wherein the operation of encoding, via the adapter model, the first embedding into a second embedding comprises:
determine a first entity and a mask in the first embedding;
determine, via the adapter model, a plurality of alternate embeddings based on the position of the first entity and the mask;
determine, via the adapter model, a probability associated with each of the plurality of alternate embeddings returning the factual information when fed to the first attention layer of the pre-trained language model; and
determine, via the adapter model, the second embedding based on the probability associated with each of the plurality of alternate embeddings.

### Claim 18 (depends on 10)

The system of claim 10, wherein the processor performs operations comprising:
receive, via a communication interface, a training data set, the training dataset including a plurality of sets of similar natural language queries for factual information;
determine a first entity and a second entity in each query in a set of similar natural language queries from the plurality of sets of similar natural language queries;
generate a mask for at least the first entity or the second entity in each query in the set of similar natural language queries;
determine a masked embedding for each masked query in the set of similar natural language queries;
determine a response to each masked embedding via the pre-trained language model, via the first attention layer of the pre-trained language model;
determine whether the response matches the masked entity in the natural language query;
determine a probability that a masked embedding will return the response that matches the masked entity in the natural language query; and
update an adapter model that chooses a query where the response from the pre-defined language model matches the masked entity based on the determined probability.

### Claim 19

A processor-readable non-transitory storage medium storing a plurality of processor-executable instructions for extracting factual information from a language model, the instructions being executed by one or more processors to perform operations comprising:
receiving, via a communication interface, a query for factual information;
transforming the query into a natural language prompt in a format of a query subject and a queried relation;
encoding, via an embedding layer of a pre-trained language model, the natural language prompt into a first embedding;
encoding, via an adapter model, the first embedding into a second embedding based on a probability that the second embedding returns the factual information when the second embedding is fed the first attention layer of the pre-trained language model, wherein the adapter model is placed between the embedding layer of the pre-trained language model and the first embedding layer of the pre-trained language model;
decoding, by the first attention layer of the pre-trained language mode, the second embedding into a response to the query; and
extracting the factual information from the decoded response to the query.

### Claim 20 (depends on 19)

A processor-readable non-transitory storage medium storing a plurality of processor-executable instructions for extracting factual information from a language model, of claim 19, the instructions being executed by one or more processors to perform operations comprising:
determining, via an LSTM (Long Short Term Memory) layer, an LSTM encoding based on the first embedding; and
determining, via a Multi-Layer Perceptron (MLP), the second embedding based on the LSTM encoding.