Please note: The algorithm descriptions in English have been automatically translated. Errors may have been introduced in this process. For the original descriptions, go to the Dutch version of the Algorithm Register.

PolyAI Voicebot

PolyAI is a voicebot that can communicate with a citizen based on natural speech recognition. Poly AI uses an algorithm to recognise the subject of a question asked.

Last change on 2nd of September 2024, at 7:43 (CET) | Publication Standard 1.0
Publication category
Impactful algorithms
Impact assessment
DPIA
Status
In use

General information

Theme

Field not filled in.

Begin date

2024-01

Contact information

algoritmen@amsterdam.nl

Responsible use

Goal and impact

PolyAI is a voicebot that can communicate with a citizen based on natural speech recognition. Poly AI uses an algorithm to recognise the subject of a question asked.

This topic is used to initiate one of the following actions:

  • Transferring to the appropriate department
  • Providing a predefined answer to frequently asked questions.


The municipality of Amsterdam is engaged in a programme to improve contact with Amsterdammers. In doing so, the Amsterdam municipality believes that interaction with the Amsterdammer should be personal, simple and accessible. One of the ways the municipality of Amsterdam does this is by bringing together the information in all our 'channels' (such as phone, mail, website, social media, counter) so that (regardless of how you contact us) you always get the same information. Using the voicebot also contributes to these goals. This ensures that staff have more time for personal contact with citizens, where needed.


Considerations

The General Data Protection Regulation (GDPR) defines tasks in the public interest as tasks related to the exercise of public authority or tasks performed by a body entrusted with a public task. Services provided by the municipality fall within this framework.

Poly AI is deployed within the Contact Centre Amsterdam (CCA) and replaces Poly AI replaces the telephone choice menu.

The algorithm does not use personal data. Poly AI does use transcripts of calls. If the citizen mentions personal data (e.g. name) in a call, there may be identified personal data. This data is not used for determining intent. This data will be destroyed after 28 days

Consideration has been given to countering unequal treatment (due to the voicebot not 'understanding' certain groups of people):

  • Voice recognition is trained based on hundreds of millions of examples from public online discussions.
  • The voicebot asks citizens whether they have correctly recognised the subject before transferring.
  • If no subject is recognised, the caller is transferred to a contact centre employee.


Human intervention

The bot redirects the caller. If this is not done correctly, it is immediately corrected by the relevant agent.

Risk management

The algorithm is classified as limited risk: no personal data are used, there is no generative AI, the result of the algorithm is not used for decision-making or risk assessment.



Link to Processing Index

https://www.amsterdam.nl/privacy/verwerkingsregister/

Impact assessment

Data Protection Impact Assessment (DPIA)

Operations

Data

The algorithm ensures that the voicebot can generalise beyond the training data provided and correctly classify similar data.

Training is done using a combination of public data and data manually taken from transcripts of real callers.

Speech recognition is trained based on hundreds of millions of examples from public online discussions.

"Intentions" are trained with a combination of data. Initially, the PolyAI development team creates a list of phrases to trigger a specific intention. For example, "I want to apply for a passport" or "I need a new passport". The voicebot can generalise beyond these examples, so a caller does not have to use these phrases word for word to trigger the intention. In the test and early follow-up period, the voicebot will have to deal with data it does not recognise. The PolyAI development team uses this additional input to supplement the original training data and enhance the performance of each intention.

Technical design

The self-learning aspect concerns learning to better recognise the language of citizens. Forwarding is done on the basis of a routing table managed by the municipality of Amsterdam. This is not self-learning.


The algorithm is PolyAI's own ConveRT model. (see also this https://arxiv.org/pdf/1911.03688).

External provider

Private