Success Story - en
NLP Contact Centers

Semantic segmentation of Contact Center dialogs

Description: build automation scripts of dialogue operators of the contact centre with customers and the subsequent automated evaluation of the quality of the operators in the contact center requires a categorization of a set of dialogues between operators and customers, highlighting the main topics of the recognition sequence in the conversations between operators and callers. It is planned to use text segmentation into semantically related parts using thematic modeling models.

Context: the contact center employs a large number of operators, you need to understand the quality of their work, since their effectiveness and sales efficiency directly depends on their following the script. Validation of manually - quite a long and expensive process, it is necessary to automate it and to propose a methodology of assessing quality.

Decision: a lot of topics were built based on replicas inside the CC dialogues. A model was built that decomposes the entire new dialog into replicas, and marks the replicas with appropriate themes. Then a "map" of dialogs with the highest quality sales was built - an ideal operator script. For new dialogues, the sequence of replicas was checked against the ideal map and the" quality " of the dialogue was measured.

Results:
Selected topics: 41 topics;
The quality of the allocation: 75% accuracy on average;
Integration into the Bank's BI software quality assessment of operators;
Increase the conversion rate and the percentage of successful dialogues.

To build the model, we used:
Dialogs between contact center operators and the client;
Information about the success of the dialog;
A priori expert knowledge of topics;
Assessor markup.

Simulation result:
Set of themes for operators and clients ' dialogs;
Transition graph between topics for successful and unsuccessful dialogues;
The thematic segmentation tool of replicas of the dialogue.

Customer: Finance, Banking

Technology stack: Thematic modeling, Syntactic relations, Neural network models of text segmentation, BigARTM, PyTorch, gensim, nltk, Python.