# US 2021/0279622 A1 · Learning with limited supervision for question-answering with light-weight markov models

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

- Number: US 2021/0279622 A1
- Document ID: US20210279622A1
- Status: Published application
- Kind: A1
- Application number: 16/813,098
- Filed: 2020-03-09
- Published: 2021-09-09
- Earliest claimed: 2020-03-09
- Assignees: Adobe Inc.
- Applicants: Adobe Inc.
- Inventors: Trung Huu Bui; Tong Sun; Natwar Modani; Lidan Wang; Franck Dernoncourt
- Web page: https://patentgrep.com/patent/US20210279622A1

## CPC

- G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks
- G06F 40/205: Parsing
- G06F 40/279: Recognition of textual entities
- G06F 40/30: Semantic analysis
- G06N 3/09: Supervised learning
- G06N 20/00: Machine learning
- G06N 3/08: Learning methods
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

Methods for natural language semantic matching performed by training and using a Markov Network model are provided. The trained Markov Network model can be used to identify answers to questions. Training may be performed using question-answer pairs that include labels indicating a correct or incorrect answer to a question. The trained Markov Network model can be used to identify answers to questions from sources stored on a database. The Markov Network model provides superior performance over other semantic matching models, in particular, where the training data set includes a different information domain type relative to the input question or the output answer of the trained Markov Network model.

## Claims

### Claim 1

One or more computer storage media having computer-useable instructions that, when used by one or more computing devices, cause the one or more computing devices to train a Markov Network model for identifying answers to questions by performing operations comprising:
receiving training data comprising question-answer pairs;
identifying object texts and relation texts within the question-answer pairs;
generating a Markov Network comprising a node network of object nodes and relation nodes, wherein the object nodes are based on the identified object texts and the relation nodes are based on the identified relation texts;
training the Markov Network by optimizing a feature function associated with the Markov Network, the training resulting in a trained Markov Network model; and
storing the trained Markov Network model for use in identifying an answer to a question.

### Claim 2 (depends on 1)

The media of claim 1, wherein the question-answer pairs are associated with binary labels, and wherein the binary labels comprise a first label that indicates a correct answer to a question and a second label that indicates an incorrect answer to the question.

### Claim 3 (depends on 1)

The media of claim 1, wherein the training data comprises question-answer pairs associated with a non-specific domain type and the trained Markov Network model is for use in identifying an answer from a specific domain type.

### Claim 4 (depends on 1)

The media of claim 1, wherein generating the Markov Network further comprises:
forming a relation side structure of the Markov Network that includes the relation nodes, each relation node representing an identified relation text, the relation side structure comprising a plurality of relation binary cliques that each include a relation node pair having a relation first node of a first relation text identified from a question and a relation second node of a second relation text identified from an answer of a question-answer pair, wherein each relation first node for each relation node pair of the plurality of relation binary cliques is dependent upon each relation second node for each relation node pair of the plurality of relation node pairs; and
forming an object side structure of the Markov Network that includes the object nodes, each object node representing an identified object text, the object side structure comprising a plurality of object binary cliques that each include an object node pair having an object first node of a first object text identified from the question and an object second node of a second object text identified from the answer of the question-answer pair, wherein each object first node for each object node pair of the plurality of object binary cliques is dependent upon each object second node for each object node pair of the plurality of object node pairs.

### Claim 5 (depends on 4)

The media of claim 4, wherein:
a relative position of each of the relation first nodes is determined from a relative position of each of the identified relation texts within the question-answer pair, and wherein each of the relation first nodes is dependent upon adjacent relation first nodes; and
a relative position of each of the object first nodes is determined from a relative position of each of the identified object texts within the question-answer pair, and wherein each of the object first nodes is dependent upon adjacent object first nodes.

### Claim 6 (depends on 5)

The media of claim 5, wherein an exponential decay factor is applied to the Markov Network based on the relative position of each of the relation first nodes and the relative position of each of the object first nodes.

### Claim 7 (depends on 4)

The media of claim 4, wherein the feature function generates a probability score indicating a similarity between each of the plurality of relation binary cliques and each of the plurality of object binary cliques.

### Claim 8 (depends on 7)

The media of claim 7, wherein training the Markov Network further comprises jointly optimizing parameters associated with the feature function using a maximum likelihood as optimization criteria.

### Claim 9 (depends on 1)

The media of claim 1, wherein the feature function is selected from a deep learning model or an information retrieval model.

### Claim 10 (depends on 8)

The media of claim 8, wherein the feature function is Bidirectional Encoder Representations from Transformers (BERT).

### Claim 11

A computerized method identifying answers to questions using a trained Markov Network model for identifying answers to questions, the method comprising:
receiving a question from a user computing device;
identifying an answer to the question using a trained Markov Network model, the trained Markov Network model determined by training a Markov Network comprising a node network of object nodes and relation nodes, the object nodes based on object texts and the relation nodes based on relation texts identified from question-answer pairs of training data; and
providing the answer to the user computing device in response to receiving the question.

### Claim 12 (depends on 11)

The method of claim 11, wherein the question received from the computing device is associated with a specific domain type and the training data is associated with a non-specific domain type.

### Claim 13 (depends on 11)

The method of claim 11, wherein the training data comprises a first portion associated with a non-specific domain type and a second portion associated with a specific domain type, and the answer is identified from a dataset of the specific domain type.

### Claim 14 (depends on 11)

The method of claim 11, wherein the Markov Network comprises:
a relation side structure that includes the relation nodes, each relation node representing an identified relation text, the relation side structure comprising a plurality of relation binary cliques that each include a relation node pair having a relation first node of a first relation text identified from a question and a relation second node of a second relation text identified from an answer of a question-answer pair, wherein each relation first node for each relation node pair of the plurality of relation binary cliques is dependent upon each relation second node for each relation node pair of the plurality of relation node pairs; and
an object side structure that includes the object nodes, each object node representing an identified object text, the object side structure comprising a plurality of object binary cliques that each include an object node pair having an object first node of a first object text identified from the question and an object second node of a second object text identified from the answer of the question-answer pair, wherein each object first node for each object node pair of the plurality of object binary cliques is dependent upon each object second node for each object node pair of the plurality of object node pairs.

### Claim 15 (depends on 14)

The method of claim 14, wherein:
each of the relation first nodes is dependent upon adjacent relation first nodes; and
each of the object first nodes is dependent upon adjacent object first nodes.

### Claim 16 (depends on 14)

The method of claim 14, wherein the Markov Network further comprises an exponential decay factor applied to a relative position of each of the relation first nodes and a relative position of each of the object first nodes.

### Claim 17 (depends on 14)

The method of claim 14, wherein the Markov Network is associated with a feature function that generates a probability score indicating a similarity between each of the plurality of relation binary cliques and each of the plurality of object binary cliques.

### Claim 18 (depends on 17)

The method of claim 17, wherein the Markov Network is trained by jointly optimizing parameters associated with the feature function using a maximum likelihood as optimization criteria.

### Claim 19 (depends on 14)

The method of claim 14, wherein the Markov Network is associated with a feature function, the feature function selected from a deep learning model or an information retrieval model.

### Claim 20

A computer system comprising:
means for receiving a question from a user computing device;
means for identifying an answer to the question using a trained Markov Network model, the trained Markov Network model determined by training a Markov Network comprising a node network of object nodes and relation nodes, the object nodes based on object texts and the relation nodes based on relation texts identified from question-answer pairs of training data; and
means for providing the answer to the user computing device in response to receiving the question.
