# US 2022/0044134 A1 · Intelligent question answering on tabular content

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

- Number: US 2022/0044134 A1
- Document ID: US20220044134A1
- Status: Published application
- Kind: A1
- Application number: 16/989,495
- Filed: 2020-08-10
- Published: 2022-02-10
- Earliest claimed: 2020-08-10
- Assignees: Accenture Global Solutions Limited
- Applicants: Accenture Global Solutions Limited
- Inventors: Eldhose Joy; Sundar Prasad Jayaraman; Harsha Jawagal; Vishal Radhesham Supekar
- Web page: https://patentgrep.com/patent/US20220044134A1

## CPC

- G06N 5/04: Inference or reasoning models
- G06F 16/334: Query execution (filtering based on additional data G06F16/335)
- G06N 3/045: Combinations of networks
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06N 20/00: Machine learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

An intelligent table-based question answering system receives an input user query requesting information from an input table to generate a response to the input query. Two types of user queries including textual user queries and scalar user queries that require at least one mathematical operation to be executed can be processed for answer generation. The input table is processed to generate a paragraph of table text which includes row and column information extracted from the input table. The input query is provided along with the paragraph of table text to a deep Quans model which outputs a candidate answer that forms a textual portion of the response. Also, the candidate answer is reverse mapped to the input table and a portion of the input table including the candidate answer is provided as a tabular portion of the response.

## Claims

### Claim 1

An intelligent question answering system, comprising:
at least one processor;
a non-transitory processor-readable medium storing machine-readable instructions that cause the processor to:
train a TAQuAns (table QuAns) model to respond to user queries regarding information in tables;
access at least one input table based on data from an input user query that pertains to the at least one input table;
extract row information regarding one or more rows and including one or more column headers from the at least one input table;
extract column information regarding one or more columns and including one or more row headers from the at least one input table;
form a paragraph of table text from the row information and the column information extracted from the at least one input table;
provide the paragraph of table text and the input user query to the TAQuAns model;
obtain at least a portion of the paragraph of table text as a candidate answer to the input user query from the TAQuAns model based on the input user query; and
provide the candidate answer as a textual portion of a response to the input user query via an output interface.

### Claim 2 (depends on 1)

The intelligent question answering system of claim 1, wherein the processor is to further:
reverse map the candidate answer to a portion of the at least one input table that includes information used in the candidate answer.

### Claim 3 (depends on 1)

The intelligent question answering system of claim 1, wherein to provide the response the processor is to further:
provide the portion of the at least one input table that is mapped to the textual portion as a tabular portion of the response via the output interface.

### Claim 4 (depends on 1)

The intelligent question answering system of claim 1, wherein the processor is to further:
determine whether the input user query is a textual user query or a scalar user query, wherein the textual user query requires textual processing based on the paragraph of table text and the scalar user query requires execution of at least one predefined mathematical operation on contents of the input table.

### Claim 5 (depends on 4)

The intelligent question answering system of claim 4, wherein to determine whether the input user query is a textual user query or a scalar user query the processor is to further:
determine if the input user query includes one or more of numerical data types or mathematical keywords that signify that a mathematical operation is required.

### Claim 6 (depends on 5)

The intelligent question answering system of claim 5, wherein if it is determined that the input user query is a scalar user query the processor is to further:
identify a mathematical formula for the at least one predefined mathematical operation from a collection of predefined mathematical formulae and a dataset for executing the at least one predefined mathematical operation, wherein the dataset is identified from the paragraph of table text; and
execute the at least one predefined mathematical operation on the dataset.

### Claim 7 (depends on 5)

The intelligent question answering system of claim 5, wherein if it is determined that the input user query is a scalar user query the processor is to further:
further determine if the scalar user query is a complex scalar user query that requires multiple mathematical operations;
identify a mathematical operation to be executed from the multiple mathematical operations, wherein a mathematical formula for the identified mathematical operation is selected from a collection of predefined mathematical formulae corresponding to different mathematical operations and a dataset for executing the identified mathematical operation, wherein the dataset is identified from the paragraph of table text;
execute the identified mathematical operation on the dataset;
select a next mathematical operation from the multiple mathematical operations, a mathematical formula for the next mathematical operation is selected from the collection of predefined mathematical formulae and a result of a prior mathematical operation;
execute the next mathematical operation; and
repeat the steps of identifying the next mathematical operation, executing the next mathematical operation until all the multiple mathematical operations are executed.

### Claim 8 (depends on 1)

The intelligent question answering system of claim 1, wherein the TAQuAns model is based on Deep Quans which is a variation of Bidirectional Encoder Representations from Transformers (BERT).

### Claim 9 (depends on 1)

The intelligent question answering system of claim 1, wherein to provide the response to the input user query the processor is to:
provide the textual portion of the response via a chatbot interface.

### Claim 10 (depends on 1)

The intelligent question answering system of claim 1, wherein to form the paragraph of table text from the row information, the processor is to:
identify contents of cells in a topmost row as column headers;
identify contents of cells in a leftmost column of the at least one input table as pertaining to a primary entity;
for each row:
combine the contents from the cells in the leftmost column with a corresponding one of the column headers to form row-level sentences;
and
combine the row-level sentences to form a paragraph of row information for the at least one input table.

### Claim 11 (depends on 10)

The intelligent question answering system of claim 10, wherein to form the paragraph of table text from the column information, the processor is to:
identify columns and the column headers;
for each column:
combine each of the column headers with contents of cells in a corresponding column to form a column-level sentence;
combine the column-level sentences generated for the columns of the at least one input table to form a paragraph of column information; and
append the paragraph of row information to the paragraph of column information to form the paragraph of table text.

### Claim 12 (depends on 10)

The intelligent question answering system of claim 10, wherein the processor is to:
process the at least one input table to generate a data representation of the input table, wherein the data representation includes generated text that represents a natural language text derived from the at least one input table, row_position_ids that represent mappings derived from an aggregation of positions spanning the generated text and the rows of the at least one input table, column_position_ids that represent mappings derived from an aggregation of positions spanning the generated text and the columns of the at least one input table and data_features which include contents of corresponding target cells, rows or columns.

### Claim 13 (depends on 1)

The intelligent question answering system of claim 1, wherein to train the TAQuans model, the processor is to:
train the TAQuans model on training data that includes sample user queries and representations of sample tables associated with the sample user queries wherein each of the representations includes generated text, row_position_ids, column_position_ids, and data features extracted from the sample tables.

### Claim 14

A method of answering user queries for tabular data, comprising:
accessing at least one input table using an input user query that pertains to the at least one input table;
extracting row and column information from the input table;
storing data from a first row of the at least one input table as corresponding column headers of the at least one input table;
storing data from a first column as pertaining to a primary entity that forms relationships with the column headers of the at least one input table;
combining the primary entity with each of the column headers to form row-level sentences;
combining the column headers with contents of corresponding columns to form column-level sentences;
combining the row-level sentences and the column-level sentences to form a paragraph of text corresponding to the at least one input table;
generating representation for the at least one input table, wherein the representation includes generated text, row_ids, column_ids and data features extracted from the at least one table;
providing the input user query and the representation to a TAQuans model;
obtaining from the TAQuans model, a candidate answer to the input user query pertaining to the at least one input table; and
generating a response to the input user query wherein the response includes the candidate answer and a tabular portion with highlighted portion of the at least one input table that includes the candidate answer.

### Claim 15 (depends on 14)

The method of claim 14, further comprising:
determining that the input user query is a scalar user query that requires at least one mathematical operation.

### Claim 16 (depends on 15)

The method of claim 15, wherein the at least one mathematical operation includes one or more of a count, a sum and a mean.

### Claim 17 (depends on 14)

The method of claim 14, wherein at least one input table includes a plurality of input tables and extracting the row and column information further comprises:
selecting, using keywords extracted from the input user query, a relevant input table from the plurality of input tables.

### Claim 18 (depends on 14)

The method of claim 14, wherein a textual portion of the response includes one or more grammatically accurate sentences and the textual portion of the response is output via a voice interface.

### Claim 19

A non-transitory processor-readable storage medium comprising machine-readable instructions that cause a processor to:
train a TAQuAns (table QuAns) model to respond to user queries regarding information in tables;
access at least one input table using an input user query that pertains to the at least one input table;
extract row information regarding one or more rows and including one or more column headers from the at least one input table;
extract column information regarding one or more columns and including one or more row headers from the at least one input table;
form a paragraph of table text from the row information and the column information extracted from the at least one input table;
provide the paragraph of table text and the input user query to the TAQuAns model;
obtain at least a portion of the paragraph of table text as a candidate answer to the input user query from the TAQuAns model based on the input user query; and
provide the candidate answer as a textual portion of a response to the input user query via an output interface.

### Claim 20 (depends on 19)

The non-transitory processor-readable storage medium of claim 19, further comprising instructions that cause the processor to:
determine that the input user query is a scalar user query that requires a mathematical operation, wherein the mathematical operation includes one or more of a count, a sum, and a mean.
