Insight from chatbots and WhatsApp with NLP

As organizations open new client interaction channels, unstructured information capture grows (text, voice, image, etc.). Most chat systems allow for parametrization of instruction sequences but do not inform the business about the conversation or narrative within the chat. This blog seeks to expose some ideas from our consulting experience to make omnichannel an inexhaustible source of insights for the business.

The table below shows some questions our clients have asked with relation to the chat’s and WhatsApp's information; it also shows what has been done with the answers

Usage case

Business questions

Actions are taken after obtaining an answer

User Experience
  • What is the sequence of conversation? How does it begin, mature and end?
  • What is the client’s emotionality throughout the conversation? What triggers positive or negative emotions?
  • Improve funnels and bottlenecks in the bot’s instruction tree.
  • Detect negative emotional states in advance to accelerate short answers or transition from machine to a human agent.
  • Isolate information to feed models and predict client churn.
Audits
  • What are the main needs the user seeks to solve during the chat?
  • Does the agent respond appropriately? (If a human agent)
  • Does the bot have a parameterized response? (if an artificial agent)
  • Customer service clinic.
  • Enable new functions in the channel.
  • Expedite routes and create quick access to specific needs.
  • Improve channel positioning and performance
Innovation
  • What user questions are not parameterized in the bot?
  • Unresolved needs?
  • Develop new functions, products or services within the chat.

When looking at them closely, the chat’s information has several complexities:

    • A sequence defined by questions and answers very similar to a Ping-Pong game.
    • Short answers that resemble a telegram or tweet.
    • Excessive courtesy, specifically from the agent, makes the conversation’s mood hard to detect.

Our consulting process is based on integrating linguistic methodologies to artificial intelligence as shown below:

 
    1. Data integration: Creation of a text datamart that integrates the chats in a sequence, which makes it easier to access enriched text. A text datamart disaggregates the various orthographical forms in a word into tokens. People often do not chat with proper grammar and the absence of an accent or changing an s to a c can change the meaning of a word.
    2. Sample design for training: Sample conversation sets are required to train the model. Given that channels evolve and add functionality over time and the client needs change, the sample design must include conversations from different months, moments during the day, days of the week, etc. This also helps to consider the possible seasonal effects that may impact the model’s narrative context.
    3. Text mining: Based on the chat set, we carry out parsing and filtering to isolate keywords that appear frequently and are associated between them. Synonyms are cleaned and a list of keywords associated with the business is created so signal detection is honed. This allows extracting core themes.
    4. Creation of taxonomies: Using linguistic heuristics and methodologies, we crate thematic taxonomies associated with the needs of the business. This helps classify client questions, products, requirements, services, agent responses and others.
    5. Association rules: After classifying the text into key taxonomies for the business accordingly, we analyze the conversation sequence: how it begins and how it ends. With this in mind, rules of association help identify frequent sequences with a high probability of occurrence, which allows to segment conversations according to narrative sequences and create interpretations based on the journey.
    6. Context analysis: Using boolean rules and advanced linguistics, taxonomies become automatic classifiers based on context, helping to determine feelings, irony, sarcasm and reproach to finish with a complete textual analysis.

We have had the opportunity to support various organizations in their unstructured data analytics challenges for chats. It will be our pleasure to share these experiences with you, let’s meet.

When looking at them closely, the chat’s information has several complexities:

    • A sequence defined by questions and answers very similar to a Ping-Pong game.
    • Short answers that resemble a telegram or tweet.
    • Excessive courtesy, specifically from the agent, makes the conversation’s mood hard to detect.

Our consulting process is based on integrating linguistic methodologies to artificial intelligence as shown below:

 
    1. Data integration: Creation of a text datamart that integrates the chats in a sequence, which makes it easier to access enriched text. A text datamart disaggregates the various orthographical forms in a word into tokens. People often do not chat with proper grammar and the absence of an accent or changing an s to a c can change the meaning of a word.
    2. Sample design for training: Sample conversation sets are required to train the model. Given that channels evolve and add functionality over time and the client needs change, the sample design must include conversations from different months, moments during the day, days of the week, etc. This also helps to consider the possible seasonal effects that may impact the model’s narrative context.
    3. Text mining: Based on the chat set, we carry out parsing and filtering to isolate keywords that appear frequently and are associated between them. Synonyms are cleaned and a list of keywords associated with the business is created so signal detection is honed. This allows extracting core themes.
    4. Creation of taxonomies: Using linguistic heuristics and methodologies, we crate thematic taxonomies associated with the needs of the business. This helps classify client questions, products, requirements, services, agent responses and others.
    5. Association rules: After classifying the text into key taxonomies for the business accordingly, we analyze the conversation sequence: how it begins and how it ends. With this in mind, rules of association help identify frequent sequences with a high probability of occurrence, which allows to segment conversations according to narrative sequences and create interpretations based on the journey.
    6. Context analysis: Using boolean rules and advanced linguistics, taxonomies become automatic classifiers based on context, helping to determine feelings, irony, sarcasm and reproach to finish with a complete textual analysis.

We have had the opportunity to support various organizations in their unstructured data analytics challenges for chats. It will be our pleasure to share these experiences with you, let’s meet.

South America

Mexico and Central America


 

 

Artículos Relacionados

Natural Language Processing for marketing executives

The first 20 years of this century were focused on developing technologies for managing and...

CONTINUAR LEYENDO

How to understand artificial intelligence - Machine learning?

At SINNETIC, our cognitive services unit develops analytical and artificial intelligence models...

CONTINUAR LEYENDO

Data Science: Customer experience and predictive collection

According to our Consumer Pulse observatory, 63% of interviewees report a reduction in their...

CONTINUAR LEYENDO