Artificial Intelligence in VoC
Subject detection by AI
Elementary subjects
Extraction of Key Topics
We analyze each verbatim to identify the key points and sentiments expressed. This enables us to identify recurring motifs.
For certain formats, such as chat conversations, emails or audio calls, we analyze each message individually, while taking the whole conversation into account. This enables our model to understand the overall context and interpret each interaction consistently.
Take, for example, the following verbatim:
Verbatim :
Extraction :
- Positives: Pleasant staff, good products
- Negatives: Expensive products, not very interesting loyalty program
This extraction is applied systematically to each verbatim in order to clearly distinguish between positive and negative aspects.
Example on several verbatims from the same data source:
Clustering
Once we've extracted the key points, we group them into clusters to form what we call "elementary topics".
This approach makes it possible to bring subjects together, despite the diversity of verbatim formulations, while retaining a fine granularity. This allows us to reflect the nuances expressed by customers.
In addition, elementary topics are generated automatically and in a scalable way. As new verbatims are integrated, our system identifies existing topics or creates new ones.
This process guarantees a dynamic, adaptive analysis, capable of tracking the emergence of new topics in real time.
Example of topic grouping and creation:
Classification plan
The elementary subjects are then organized according to business categories predefined by the customer, described in"classification plans".
Each category includes :
- Name: Category designation
- Description/Classification rule: Specific classification criteria
- Examples (optional) : Illustrations to clarify the application
These elements help our model to increase its accuracy and to understand the specificities of often complex business subjects.
Elementary topics are then mapped onto these business categories, thus integrating all verbatims into these business topics.
Example of a mapping of elementary subjects in a filing plan:
Partial or overall feeling
Types of feelings
To analyze feelings, we distinguish two categories:
- Verbatim Global Sentiment: Inferred by a model that evaluates the whole text.
- Partial feelings: associated with the elementary topics identified in the verbatim sub-sections.
For simple verbatim, such as text, we exclusively use the overall sentiment to assess the general feeling. On the other hand, for formats such as chat conversations, audio calls and email exchanges, we employ a distinct approach by analyzing each message individually.
Analysis process for Conversations
In these conversational formats (chat conversation, email, audio call...), we evaluate the overall sentiment of each message individually. To determine the sentiment of the entire conversation, we employ the following process:
- Positive: If all the messages in the conversation are positive, the sentiment of the discussion is positive.
- Negative: If all messages display a negative sentiment, then the conversation is deemed negative.
- Mixed: The coexistence of positive and negative feelings among messages results in an overall mixed feeling.
- Neutral: The absence of any particular sentiment in the messages results in an overall neutral sentiment.
AI summary
Our dashboard features advanced automatic summary generation. Once filters have been applied (e.g. sentiment, subject, date, etc.), we synthesize the selected verbatims to create a relevant summary.
Summary features
- Topic highlighting : The summary highlights the main topics emerging from the filtered verbatims.
- Key Phrase Quotes: Significant extracts from customers are quoted to illustrate crucial points.
- Identifying Repeating Patterns : Frequently mentioned points are underlined, taking into account the importance of the filters applied.
Résumé adaptability
The summary is automatically generated in real time each time a filter is modified on the dashboard or new verbatims are added, ensuring that the analysis is always up to date and synchronized with the filters.