# US 2019/0244603 A1 · Methods and Systems for Intent Detection and Slot Filling in Spoken Dialogue Systems

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

- Number: US 2019/0244603 A1
- Document ID: US20190244603A1
- Status: Published application
- Kind: A1
- Application number: 15/962,900
- Filed: 2018-04-25
- Published: 2019-08-08
- Earliest claimed: 2018-02-06
- Assignees: Robert Bosch GmbH
- Applicants: Robert Bosch GmbH
- Inventors: Pongtep Angkititrakul; Raphael Schumann
- Web page: https://patentgrep.com/patent/US20190244603A1

## CPC

- G10L 15/16: using artificial neural networks
- G06F 40/30: Semantic analysis
- 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 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/09: Supervised learning
- G10L 15/1815: {Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning}
- G10L 15/1822: {Parsing for meaning understanding}
- G10L 15/22: Procedures used during a speech recognition process, e.g. man-machine dialogue
- G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A method for spoken language understanding (SLU) includes generating a first encoded representation of words from a user based on an output of a recurrent neural network (RNN), generating an intent label corresponding to the words based on an output of a first RNN decoder based on the first encoded representation, generating a corrected plurality of words based on an output of a second RNN decoder based on the first encoded representation and the intent label, generating a second encoded representation corresponding to the plurality of corrected words using the RNN encoder based on the plurality of corrected words, and generating a machine-readable dialogue phrase that includes at least one word in the plurality of corrected words assigned to at least one slot based on an output of a third RNN decoder based on the second encoded representation of the plurality of corrected words and the intent label.

## Claims

### Claim 1

A method for operation of a spoken language understanding (SLU) system comprising:
receiving, with a processor in the SLU system, a plurality of words in a text phrase;
generating, with the processor, a first encoded representation of the plurality of words based on an output of a recurrent neural network (RNN) encoder using the plurality of words as inputs to the RNN encoder;
generating, with the processor, an intent label corresponding to the plurality of words based on an output of a first RNN decoder using the first encoded representation as input to the first RNN decoder;
generating, with the processor, a corrected plurality of words based on an output of a second RNN decoder using the first encoded representation and the intent label as inputs to the second RNN decoder, the corrected plurality of words including at least one word that is different than one of the plurality of words in the text phrase;
generating, with the processor, a second encoded representation corresponding to the plurality of corrected words based on another output of the RNN encoder using the plurality of corrected words as inputs to the RNN encoder; and
generating, with the processor, a machine-readable dialogue phrase that includes at least one word in the plurality of corrected words assigned to at least one slot in the machine-readable dialogue phrase based on an output of a third RNN decoder using the second encoded representation of the plurality of corrected words and the intent label as inputs to the third RNN decoder.

### Claim 2 (depends on 1)

The method of claim 1, further comprising:
receiving, with an audio input device, a spoken input phrase; and
generating, with the processor, the plurality of words in the text phrase using automated speech recognition based on the spoken input phrase.

### Claim 3 (depends on 1)

The method of claim 1, wherein the RNN encoder generates the first encoded representation corresponding to the plurality of words in the text phrase based on a first vector of hidden states generated in the RNN encoder.

### Claim 4 (depends on 3)

The method of claim 3, wherein the RNN encoder generates the second encoded representation corresponding to the plurality of corrected words based on a second vector of hidden states generated in the RNN encoder.

### Claim 5 (depends on 1)

The method of claim 1, further comprising:
identifying, with the processor, the machine-readable dialogue phrase in a database stored in a memory based on the intent label.

### Claim 6 (depends on 1)

The method of claim 1, wherein the processor implements the RNN encoder as a bidirectional RNN, the first RNN decoder as a unidirectional RNN, the second RNN decoder as another unidirectional RNN, and the third RNN decoder as another unidirectional RNN.

### Claim 7 (depends on 6)

The method of claim 6, wherein the processor implements the first RNN decoder including an attention mechanism to update a context of the RNN decoder during generation of the output of the first RNN decoder.

### Claim 8 (depends on 1)

The method of claim 1 further comprising:
generating, with an output device in the SLU system, an output in response to the machine-readable dialogue phrase including the at least one word in the plurality of corrected words assigned to the at least one slot.

### Claim 9

A system for spoken language understanding (SLU) comprising:
a memory configured to store:
a recurrent neural network (RNN) encoder;
a first RNN decoder;
a second RNN decoder; and
a third RNN decoder;
an input device; and
a processor operatively connected to the memory and the input device, the processor being configured to:
receive a plurality of words in a text phrase from the input device;
generate a first encoded representation of the plurality of words based on an output of the RNN encoder using the plurality of words as inputs to the RNN encoder;
generate an intent label corresponding to the plurality of words based on an output of the first RNN decoder using the first encoded representation as input to the first RNN decoder;
generate a corrected plurality of words based on an output of the second RNN decoder using the first encoded representation and the intent label as inputs to the second RNN decoder, the corrected plurality of words including at least one word that is different than one of the plurality of words in the text phrase;
generate a second encoded representation corresponding to the plurality of corrected words based on another output of the RNN encoder using the plurality of corrected words as inputs to the RNN encoder; and
generate a machine-readable dialogue phrase that includes at least one word in the plurality of corrected words assigned to at least one slot in the machine-readable dialogue phrase based on an output of the third RNN decoder using the second encoded representation of the plurality of corrected words and the intent label as inputs to the third RNN decoder.

### Claim 10 (depends on 9)

The system of claim 9 wherein the input device is an audio input device and the processor is further configured to:
receive a spoken input phrase with the audio input device; and
generate the plurality of words in the text phrase using automated speech recognition based on the spoken input phrase.

### Claim 11 (depends on 9)

The system of claim 9, wherein the RNN encoder generates the first encoded representation corresponding to the plurality of words in the text phrase based on a first vector of hidden states generated in the RNN encoder.

### Claim 12 (depends on 11)

The system of claim 11, wherein the RNN encoder generates the second encoded representation corresponding to the plurality of corrected words based on a second vector of hidden states generated in the RNN encoder.

### Claim 13 (depends on 9)

The system of claim 9, the memory being further configured to store:
a database that stores an association between a plurality of intent labels and a plurality of machine-readable dialogue phrases; and
the processor being further configured to:
identify the machine-readable dialogue phrase in the database based on the intent label that is generated based on an output of the first RNN decoder.

### Claim 14 (depends on 9)

The system of claim 9, wherein the memory stores the RNN encoder as a bidirectional RNN, the first RNN decoder as a unidirectional RNN, the second RNN decoder as another unidirectional RNN, and the third RNN decoder as another unidirectional RNN.

### Claim 15 (depends on 14)

The system of claim 14, wherein the memory stores the first RNN decoder including an attention mechanism to update a context of the RNN decoder during generation of the output of the first RNN decoder.

### Claim 16 (depends on 9)

The system of claim 9, further comprising:
an output device; and
the processor being operatively connected to the output device and further configured to:
generate an output with the output device in response to the machine-readable dialogue phrase including the at least one word in the plurality of corrected words assigned to the at least one slot.
