Using the Natural Language Understanding Capability of ChatGPT

Last Updated on July 3, 2023

ChatGPT as a Large Language Model, is well-known for understanding human languages. Instead of asking ChatGPT for an answer you don’t know, you can make it work on existing information while leveraging the natural language understanding (NLU) capability. In this post, you will learn

How to make ChatGPT produce a summary from a long text
How to make ChatGPT extract key information from a passage

Let’s get started.

Using the Natural Language Understanding Capability of ChatGPT.
Picture generated by the author using Stable Diffusion. Some rights reserved.

Overview

This post is divided into three sections; they are:

You AI Secretary
Extracting Unstructured Information
How to Use ChatGPT for NLU

Your AI Secretary

Imagine you have a meeting transcript with conversations between different people. The transcript would be a great way to check how a conclusion is reached by tracking the discussion, but meeting minutes should be easier to read the decision. You can ask ChatGPT to produce the meeting minute from a transcript.

Let’s consider the transcript here. An excerpt is as follows:

>> GOOD MORNING, EVERYBODY. I’M MICHAEL STERN, DELEGATE FOR COMPTROLLER BRAD LANDER AND CHAIR OF THE FINANCE COMMITTEE OF THE NEW YORK CITY TRANSITIONAL FINANCE AUTHORITY. AS A REMINDER, PLEASE KEEP YOUR FILMS MUTED UNLESS SPEAKING. I’LL NOW CALL TO ORDER A MEETING OF THE FINANCE COMMITTEE. RITA, COULD YOU PLEASE CALL THE ROLL?

…
>> THANK YOU, EVERYBODY. HAVE A GREAT WEEKEND.
>> HAVE A GREAT WEEKEND.
>> THANK YOU.

This is not a long meeting, but the transcript spans several pages. You can summarize the transcript into 1-page minutes using the following prompt:

As an experienced secretary, please summarize the meeting transcript below to meeting minutes, list out the participants, agenda, key decisions, and action items.

>> GOOD MORNING, EVERYBODY. I’M MICHAEL STERN, DELEGATE FOR COMPTROLLER
BRAD LANDER AND CHAIR OF THE FINANCE COMMITTEE OF THE NEW YORK CITY
TRANSITIONAL FINANCE AUTHORITY. AS A REMINDER, PLEASE KEEP YOUR FILMS
MUTED UNLESS SPEAKING. I’LL NOW CALL TO ORDER A MEETING OF THE FINANCE
COMMITTEE. RITA, COULD YOU PLEASE CALL THE ROLL?

…

>> THANK YOU, EVERYBODY. HAVE A GREAT WEEKEND.
>> HAVE A GREAT WEEKEND.
>> THANK YOU.

In the prompt, you provided the transcript and told ChatGPT what to include in the output to ensure all key details are included. The following is what ChatGPT produced:

Meeting Minutes

Meeting: Finance Committee Meeting
Date: [Insert Date]
Time: [Insert Time]
Location: [Insert Location]

Participants:
– Michael Stern (Delegate for Comptroller Brad Lander and Chair of the Finance Committee)
– David Womack (Representing Director Jacques Jiha)
– Diana Beinart (Representing Commissioner Preston Niblack)
– Louann Koziol (Representing Commissioner Thomas Foley)
– Hector German (Representing Speaker Adrienne Adams)

Agenda:
1. Approval of the minutes of the committee’s last meeting.
2. Resolution recommending the Board of Directors approve the 125th series resolution authorizing the issuance of up to 1.6 billion of fiscal 23 series D future tax-secured bonds.

Key Decisions:
1. Approval of the minutes of the committee’s last meeting.
– The minutes were reviewed and approved unanimously.

2. Resolution recommending the Board of Directors approve the 125th series resolution authorizing the issuance of up to 1.6 billion of fiscal 23 series D future tax-secured bonds.
– The resolution was presented, and the expected breakout of the subseries was provided.
– The committee members voted in favor of the resolution unanimously, and it was approved.

Action Items:
1. None mentioned in the transcript.

Adjournment:
– The Finance Committee meeting was adjourned.

—

Meeting: Board of Directors Meeting
Date: [Insert Date]
Time: [Insert Time]
Location: [Insert Location]

Participants:
– David Womack (Delegate for Budget Director Jacques Jiha and Chair of the Board of Directors)
– Diana Beinart (Representing Commissioner Preston Niblack)
– Michael Stern (Representing Comptroller Brad Lander)
– Louann Koziol (Representing Commissioner Thomas Foley)
– Hector German (Representing Speaker Adrienne Adams)

Agenda:
1. Resolution to allow remote meetings during a state of emergency related to COVID-19.
2. Approval of the minutes of the last board meeting.
3. Approval of the 125th series resolution authorizing the issuance of up to 1.6 billion of fiscal 2023 series D future tax-secured bonds.

Key Decisions:
1. Resolution to allow remote meetings during a state of emergency related to COVID-19.
– The resolution was proposed to allow remote meetings while a state of emergency related to COVID-19 is in effect in the city.
– The board members voted in favor of the resolution, and it was approved.

2. Approval of the minutes of the last board meeting.
– The minutes were reviewed and approved unanimously.

3. Approval of the 125th series resolution authorizing the issuance of up to 1.6 billion of fiscal 2023 series D future tax-secured bonds.
– The resolution was presented, and the committee’s recommendation was mentioned.
– The board members voted in favor of the resolution unanimously, and it was approved.

Action Items:
1. None mentioned in the transcript.

Adjournment:
– The Board of Directors meeting was adjourned.

Note: The transcript provided does not include specific details about the meeting such as dates, times, and locations. Please fill in the missing information based on the context and requirements of the organization.

The minutes are much tidier than the transcript and easier to read. One particular fact is that this transcript is two meetings scheduled as one and you can be aware of that from the minutes. However, you may still need to polish it if you wish it in a particular format. Some information that should appear in the minute but not available from the transcript is replaced with a placeholder that you need to fill in. This shouldn’t be very difficult compared to summarizing the transcript.

ChatGPT can do this job easily because it is trained to understand language, and it can extract key information. You can make it a helpful secretary with an appropriate prompt to instruct ChatGPT to reproduce the key information in a particular format.

Nevertheless, some meetings are lengthy. The transcript may be very long and would hit the context length limit in ChatGPT. In that case, ChatGPT will refuse to proceed. You may want to break down the transcript into multiple chunks and let ChatGPT summarize each first. Then combine the summarized result to generate the minutes.

Extracting unstructured information

Because ChatGPT can understand text, you can use it to extract unstructured information. Unstructured information is those immersed in paragraphs of texts, contrary to lists and tables. Since it has no trivial pattern, it isn’t easy to search.

You can leverage the language understanding capability of ChatGPT to help you extract unstructured information and produce a list. Let’s consider the following example:

Read the following text, and answer the following:

1. What is the latest passenger count?
2. What is the date referenced for this volume?
3. How many passengers it has before?
4. What was the date referenced for the previous volume?

Text:

The Port Authority of New York and New Jersey today announced that for the first time since the start of the COVID-19 pandemic, monthly air passenger levels across the agency’s regional airports surpassed pre-pandemic volumes.

In November, the Port Authority’s airports handled a total of 11 million passengers, which was 102 percent of pre-COVID November 2019 volumes. Strong demand for air travel in November was driven primarily by domestic travelers during Thanksgiving, when 8 percent more holiday passengers flew to U.S. destinations compared to Thanksgiving 2019.

The Port Authority also forecasts demand for air travel will remain high through the remainder of 2022, particularly during the 11-day period beginning Friday, Dec. 23, 2022, to Monday, Jan. 2, 2023.

The evident return to air travel in the bistate region caps a year of accomplishments and openings for the Port Authority’s redevelopment projects at its three major regional airports. It is a stark contrast to the nadir of the pandemic in 2020, when volumes at the three major airports declined steeply to just a quarter of its pre-pandemic volumes.

“We are extremely gratified to see our airports at pre-pandemic levels after more than two years of a global pandemic that left us homebound and filled with wanderlust,” said Port Authority Chairman Kevin O’Toole. “Throughout this year, we have seen the steady return of airport passengers and also the pleasant surprise on their faces when they see the brand-new, world-class air facilities that await them. We look forward to welcoming millions more passengers in the coming weeks as people celebrate the holidays with loved ones.”

“The fact that November 2022 aviation volumes surpassed pre-pandemic November 2019 numbers demonstrates that our historic investments in the region’s airports are critically important,” said Port Authority Executive Director Rick Cotton. “Our goal is to provide air passengers who travel through our airports with a world-class 21st century travel experience, and to contribute to the growing resurgence in the region’s economy post-COVID.”

LaGuardia Airport reached substantial completion earlier this year of its $8 billion public-private transformation, making it the first new major airport built in the United States in the last 25 years. The new airport was built in phases to ensure that it remained fully operational and never lost capacity throughout construction. This year all passenger facilities at the new Terminal B opened, as well as Delta’s new Terminal C and the newly rebuilt roadway network. The new airport has won praise from passengers for its 21st century design, amenities and customer experience, and awards and accolades from respected industry groups worldwide.

Newark Liberty International Airport also celebrated a key milestone in November 2022 with the ribbon-cutting of its new Terminal A, which features approximately 1 million square feet of space and 33 gates to handle larger aircraft, as well as a modernized common-use check-in area, security and baggage claim areas expected to efficiently handle an estimated 13.6 million passengers a year. The Port Authority also announced a world-class master planner to lead the transformation of the entire airport. The new Terminal A is expected to open for service in early 2023.

At John F. Kennedy International Airport, the Port Authority reached financial close on two new terminals – Terminal One and Terminal 6 – and completed the expansion of Terminal 8 with approximately 130,000 square feet of additional and refurbished space. The expansion marked the first step in the transformation of JFK Airport and allowed British Airways to move from its long-time home in the 50-year-old Terminal 7 to co-locate with American Airlines in Terminal 8. Terminal 7 will be demolished to clear space for a new Terminal 6 that will begin construction early next year. The $1.5 billion expansion of Terminal 4 is also underway, having broken ground in December 2021.

The text is from a press release obtained here.

The result from ChatGPT is as follows:

1. The latest passenger count mentioned in the text is 11 million passengers.
2. The date referenced for this volume is November.
3. The text does not provide information on the number of passengers before the COVID-19 pandemic.
4. The date referenced for the previous volume is not mentioned in the text.

The distilled result is easier to extract, and you may transfer it to another use. Also note that when the provided text does not contain the information you were looking for, it will tell you so. This is because, in your prompt, you asked it to answer based on the text only.

How to Use ChatGPT for NLU

ChatGPT as a language model should understand text fairly easily. However, to leverage its natural language understanding capabilities, you should be careful not to allow it to hallucinate. Hallucination happens when you allow it to generate text out of context. This can be avoided when you explicitly mention using only the provided information for the task in the prompt, but nothing else.

Even so, bear in mind that ChatGPT may not understand everything well. Should there be some information out of context, ChatGPT would not know. One particular example is domain-specific abbreviations and jargon. If there is information not very straightforward, ChatGPT may also fail to extract it. You may try to regenerate the second example above for a few times. ChatGPT sometimes cannot find the exact passenger volume since the text expressed it as a percentage of another volume.