Analyse de texte

Ingénierie des prompts avec l'API OpenAI

Fouad Trad

Machine Learning Engineer

Analyse de texte

Analyser un texte afin d'en extraire des informations
- Classification des textes
- Extraction d'entités
Solliciter conseils juridiques si utilisation données clients

Image showing how text analysis extracts information such as sentiment, emotion, tone, entities, etc. from a given text.

Classification des textes

Attribuer des catégories au texte
Exemple : analyse des opinions

Image showing the classes of sentiment analysis as positive, negative, or neutral.

Catégories spécifiées

Indiquer les catégories de classification si elles sont connues
Indiquer les exigences relatives au résultat

text = "I bought your XYZ Smart Watch and wanted to share my positive experience. 
Impressed with its sleek design, comfort, and touchscreen usability."


prompt = f"""Classify the sentiment of the text delimited by triple backticks as 
positive, negative, or neutral. Give your answer as a single word:
           ```{text}```"""

print(get_response(prompt))

positive

Catégories non spécifiées

Le modèle utilise ses connaissances lorsque les catégories ne sont pas spécifiées

text = "I bought your XYZ Smart Watch and wanted to share my positive experience. 
Impressed with its sleek design, comfort, and touchscreen usability."

prompt = f"""Classify the sentiment of the text delimited by triple backticks. 
             Give your answer as a single word.
           ```{text}```"""
print(get_response(response))

positive.

Pour certains problèmes ouverts, cela pourrait ne pas être efficace

Plusieurs catégories

Le texte peut être classé dans plusieurs catégories
Définir un nombre maximal de catégories si elles ne sont pas connues

text = "I bought your XYZ Smart Watch and wanted to share my positive experience. 
Impressed with its sleek design, comfort, and touchscreen usability."

prompt = f"""Identify emotions used in this text. Don't use more than 3 emotions. 
             Format your answer as a list of words separated by commas:
           ```{text}```"""

print(get_response(prompt))

impressed, positive, comfortable

Extraction d'entités

Extraction d'entités spécifiques à partir d'un texte
Exemples : noms, lieux, organisations, dates

Icon showing how entity extraction is about extracting specific entities from a given text

Extraction d'entités : préciser les entités

Indiquer les entités à extraire
Indiquer le format de sortie

text = "The XYZ Mobile X200: a sleek 6.5-inch Super AMOLED smartphone with a 48MP 
triple-camera, octa-core processor, 5000mAh battery, 5G connectivity, and Android 
11 OS. Secure with fingerprint and facial recognition. 128GB storage, expandable up 
to 512GB."

prompt = f"""Identify the following entities from the text delimited by triple backticks:
             - Product Name



           ```{text}```"""
print(get_response(prompt))

Extraction d'entités : préciser les entités

Indiquez les entités à extraire
Veuillez indiquer le format de sortie

text = "The XYZ Mobile X200: a sleek 6.5-inch Super AMOLED smartphone with a 48MP 
triple-camera, octa-core processor, 5000mAh battery, 5G connectivity, and Android 
11 OS. Secure with fingerprint and facial recognition. 128GB storage, expandable up 
to 512GB."

prompt = f"""Identify the following entities from the text delimited by triple backticks:
             - Product Name
             - Display Size


           ```{text}```"""
print(get_response(prompt))

Extraction d'entités : préciser les entités

Indiquez les entités à extraire
Veuillez indiquer le format de sortie

text = "The XYZ Mobile X200: a sleek 6.5-inch Super AMOLED smartphone with a 48MP 
triple-camera, octa-core processor, 5000mAh battery, 5G connectivity, and Android 
11 OS. Secure with fingerprint and facial recognition. 128GB storage, expandable up 
to 512GB."

prompt = f"""Identify the following entities from the text delimited by triple backticks:
             - Product Name
             - Display Size
             - Camera Resolution

           ```{text}```"""
print(get_response(prompt))

Extraction d'entités : préciser les entités

Indiquez les entités à extraire
Veuillez indiquer le format de sortie

text = "The XYZ Mobile X200: a sleek 6.5-inch Super AMOLED smartphone with a 48MP 
triple-camera, octa-core processor, 5000mAh battery, 5G connectivity, and Android 
11 OS. Secure with fingerprint and facial recognition. 128GB storage, expandable up 
to 512GB."

prompt = f"""Identify the following entities from the text delimited by triple backticks:
             - Product Name
             - Display Size
             - Camera Resolution
             Format the answer as an unordered list.
           ```{text}```"""
print(get_response(prompt))

Extraction d'entités : préciser les entités

Product Name: XYZ Mobile X200
Display Size: 6.5-inch
Camera Resolution: 48MP triple-camera

Extraction d'entités à l'aide de prompts few-shot

Pour les structures complexes

ticket_1 = "Hello, I'm Emma Adams. I'd 
like to ask about my reservation with 
the code CAR123. 
You can reach me at +123456 if needed."

ticket_2 = "This is Sarah Williams. 
I would like to request some information
regarding my upcoming flight with 
reservation code FLIGHT987. Thank you."

entities_1 = """
* Customer Details:
  - Name: Emma Adams
  - Phone: +123456
* Reservation Details:
  - Reservation Code: CAR123"""
entities_2 = """
* Customer Details:
  - Name: Sarah Williams
* Reservation Details:
  - Reservation Code: FLIGHT987"""

Extraction d'entités à l'aide de prompts few-shot

ticket_3 = "Hello, I'm David Brown (CUST123). I need assistance with my reservation under 
the code HOTEL456. There are some questions and issues related to my upcoming stay that 
require your attention."


prompt = f"""Text: {ticket_1} -> Entities: {entities_1}
             Text: {ticket_2} -> Entities: {entities_2}
             Text: {ticket_3} -> Entities: """
print(get_response(prompt))

* Customer Details:
  - Name: David Brown
  - Customer ID: CUST123
* Reservation Details:
  - Reservation Code: HOTEL456

Passons à la pratique !

Ingénierie des prompts avec l'API OpenAI