System and method for detecting chatbots

Abstract

Several embodiments determine whether an entity involved in an online conversation includes a chatbot. Challenge messages are constructed by determining plausible conversation continuations using a generative language model according to the current context of the ongoing conversation, and then distorting each continuation. The challenges are added as new messages to each conversation. The chatbot detector may then determine whether the conversation partner is a chatbot based on the similarity between the response to the challenge received from each conversation partner and an artificially generated plausible response to an undistorted conversation continuation. The distortion applied to construct the challenges is designed to send the chatbot on a significantly different trajectory.

Claims

1. A computer system comprising at least one hardware processor, wherein the at least one hardware processor is This involves applying a generative language model to generate the continuation of a substitute conversation and substitute responses, The continuation of the aforementioned substitute conversation includes a predicted follow-up to an ongoing online conversation, which includes a sequence of electronic messages. The substitute response includes an anticipated response from the conversation partner to the continuation of the substitute conversation. Applying a generative language model to generate the continuation of a substitute conversation and substitute responses, In response to this, the continuation of the aforementioned substitute conversation is distorted to create a challenge message, Adding the aforementioned challenge message to the ongoing online conversation, In response to receiving a response from the other party, including a response to the challenge message, it is determined whether the other party includes a robot, according to the similarity between the other party's response and the substitute response. A computer system configured to do so.

2. A computer system according to claim 1, wherein at least one hardware processor is configured to determine whether the conversation partner includes the robot, based on the result of comparing a similarity scale with a predetermined threshold, the similarity scale quantifies the similarity between the conversation partner's response and a substitute response.

3. A computer system according to claim 1, wherein the at least one hardware processor is configured to determine whether the conversation partner includes the robot, further according to the similarity between the other party's response and another substitute response including a predicted response to the challenge message from the conversation partner.

4. The computer system according to claim 3, wherein at least one hardware processor is configured to determine whether the conversation partner includes the robot, based on the result of comparing a first similarity scale with a second similarity scale, the first similarity scale quantifies the similarity between the partner's response and the substitute response, and the second similarity scale quantifies the similarity between the partner's response and the other substitute response.

5. A computer system according to claim 1, wherein distorting the continuation of the substitute conversation includes an item selected from a set of items including replacing a selected token in the continuation of the substitute conversation with an alternative token, adding a token to the continuation of the substitute conversation, rephrasing the continuation of the substitute conversation as a question, and rephrasing the continuation of the substitute conversation to change the emotion of the continuation of the substitute conversation.

6. The computer system according to claim 1, wherein the at least one hardware processor is In response to the creation of the aforementioned challenge message, the system determines whether the challenge message satisfies quality conditions based on the similarity between the challenge message and the continuation of the substitute conversation. Accordingly, the challenge message will be added to the ongoing online conversation only if the challenge message satisfies the quality conditions. A computer system further configured to do the following.

7. A computer system according to claim 6, wherein at least one hardware processor is configured to determine whether the challenge message satisfies the quality conditions according to the result of comparing a similarity scale with a predetermined threshold, the similarity scale quantifies the similarity between the challenge message and the continuation of the substitute conversation.

8. A computer system according to claim 6, wherein the at least one hardware processor is configured to determine whether the challenge message satisfies the quality conditions, further according to the similarity between the substitute response and another substitute response including another predicted response from the conversation partner to the challenge message.

9. The computer system according to claim 1, wherein the ongoing online conversation includes selected items from a group including the exchange of messages performed via an instant messaging application running on the computer system, a sequence of messages posted to an online forum, and a sequence of messages posted to a social media page.

10. A computer system according to claim 1, wherein applying the generative language model includes sending an encoding of a fragment of the ongoing online conversation to a remote chatbot, and in response receiving the continuation of the substitute conversation or the substitute response from the remote chatbot.

11. Using at least one hardware processor of a computer system, This involves applying a generative language model to generate the continuation of a substitute conversation and substitute responses, The continuation of the aforementioned substitute conversation includes a predicted follow-up to an ongoing online conversation, including a sequence of messages. The substitute response includes an anticipated response from the conversation partner to the continuation of the substitute conversation. Applying a generative language model to generate the continuation of a substitute conversation and substitute responses, In response to this, the continuation of the aforementioned substitute conversation is distorted to create a challenge message, Adding the aforementioned challenge message to the ongoing online conversation, In response to receiving a response from the other party, including a response to the challenge message, it is determined whether the other party includes a robot, according to the similarity between the other party's response and the substitute response. A chatbot detection method that includes the steps of doing the following.

12. A method according to claim 11, comprising the step of determining whether the conversation partner includes the robot, based on the result of comparing a similarity scale with a predetermined threshold, wherein the similarity scale quantifies the similarity between the conversation partner's response and a substitute response.

13. A method according to claim 11, further comprising the step of determining whether the conversation partner includes the robot, according to the similarity between the other party's response and another substitute response including a predicted response from the conversation partner to the challenge message.

14. A method according to claim 13, comprising the step of determining whether the conversation partner includes the robot, based on the result of comparing a first similarity scale with a second similarity scale, wherein the first similarity scale quantifies the similarity between the conversation partner's response and the substitute response, and the second similarity scale quantifies the similarity between the conversation partner's response and the other substitute response.

15. A method according to claim 11, wherein distorting the continuation of the substitute conversation includes an item selected from a set of items including replacing a selected token in the continuation of the substitute conversation with an alternative token, adding a token to the continuation of the substitute conversation, rephrasing the continuation of the substitute conversation as a question, and rephrasing the continuation of the substitute conversation to change the sentiment of the continuation of the substitute conversation.

16. A method according to claim 11, wherein the at least one hardware processor is used In response to the creation of the aforementioned challenge message, the system determines whether the challenge message satisfies quality conditions based on the similarity between the challenge message and the continuation of the substitute conversation. Accordingly, the challenge message will be added to the ongoing online conversation only if the challenge message satisfies the quality conditions. A method that includes additional steps to do so.

17. A method according to claim 16, comprising the step of determining whether the challenge message satisfies the quality conditions according to the result of comparing a similarity scale with a predetermined threshold, wherein the similarity scale quantifies the similarity between the challenge message and the continuation of the substitute conversation.

18. A method according to claim 16, further comprising the step of determining whether the challenge message satisfies the quality conditions, according to the similarity between the substitute response and another substitute response including another predicted response from the conversation partner to the challenge message.

19. A method according to claim 11, wherein the ongoing online conversation includes selected items from a group including an exchange of messages performed via an instant messaging application running on the computer system, a sequence of messages posted to an online forum, and a sequence of messages posted to a social media page.

20. A method according to claim 11, wherein applying the generative language model includes using the at least one hardware processor to send an encoding of the fragment of the ongoing online conversation to a remote chatbot, and in response to receiving the continuation of the contended conversation or the contended response from the remote chatbot.

21. A non-temporary computer-readable medium for storing instructions, wherein the instructions are executed by at least one hardware processor of a computer system. This involves applying a generative language model to generate the continuation of a substitute conversation and substitute responses, The continuation of the aforementioned substitute conversation includes a predicted follow-up to an ongoing online conversation, including a sequence of messages. The substitute response includes an anticipated response from the conversation partner to the continuation of the substitute conversation. Applying a generative language model to generate the continuation of a substitute conversation and substitute responses, In response to this, the continuation of the aforementioned substitute conversation is distorted to create a challenge message, Adding the aforementioned challenge message to the ongoing online conversation, In response to receiving a response from the other party, including a response to the challenge message, it is determined whether the other party includes a robot, according to the similarity between the other party's response and the substitute response. A non-temporary computer-readable medium that enables the aforementioned computer system to perform the operation.

Images