Conversational AI:

Technical & Ethical

Perspectives
Nayeemuzzaman(Roll: 02)
Farhan Mahmud (Roll: 16)
 Conversational Agents
• Conversational agents are computer systems
designed to converse with humans trying to mimic
human interactions as closely as possible.

• Two aspects when designing conversational agents:

(i) Technical perspective: Methods to design the

agent to efficiently recreate human interactions
by generating appropriate responses.

(ii) Ethical perspective: Rules and considerations to

to avoid harmful effect and abuse of CAs.
 Background & Related Work
• There are diverse researches regarding what algorithms to use for proper response
from chatbots. These algorithms use machine learning, deep learning, natural
language processing etc. to give responses from millions of conversational data
from users in social media, forums etc. This gives a summary of prevalent
researches.

• Not many works have been published regarding the ethical aspects of general
purpose chatbots. The few resources regarding this aspect is mentioned when the
data from them is used.
 Research Questions

This research paper tries to answer mainly two

questions:

• What response methodologies should be used to

generate appropriate responses from chatbots?

• What are the ethical considerations when creating

chatbots available for public use?
 Response Methodologies

Three types of methodologies for generating responses:

• Retrieval Based Conversation System

The system searches for candidate responses by matching metrics. The
retrieval system learns the metrics using deep learning techniques. A
repository is used to store ’responses’ for different situations which introduces
additional overhead.

• Generation Based Conversation System

Here a model built using deep learning techniques and trained using some
dataset to generate a response from a user query. As we are not using any
data repository, the response is not guaranteed to be meaningful. But it saves
from the overhead of searching for a response from a repository.

• System Ensemble
Uses both methods. Candidate responses are merged into a list from both
retrieve based system and generate based system and then the best outcome
is selected as the response for the user.
 Contextual, Diverse & Proactive Responses

• Context Based Response

Context is important to multi-turn conversations. In creating a
response model, adding previous utterance weights into it. Most
real life conversations are contextual and the appropriate response
depends also on the previous occurrences. If the conversation is too
long, context segmentation should be used to keep the relevance.

• Diverse Response
Normal models tend to give similar responses to a high portion
of the query domain. To solve this problem, the suggestion is that
we add some diverse factors in training our model so that
the candidate response generated from similar queries can be
diverse as much as possible.

• Proactive Conversations
Proactive conversation leads to a much better experience
conversing with CAs. It can continue the conversation when
it comes to a halt. One way is to add proactive suggestions to
response retrieval
 Ethical Issues

• Plurality of approaches
The desired response in a conversation will vary based on the target user
group, application domain. So, detection of the user group and adapting a
a plurality of approach is imperative to address the ethical concerns
properly.

• Trust and Transparency

Transparency about CAs goals and capabilities helps users to make informed
choices which can prevent users from self harming themselves. Users
expectations should not be taken advantage of. There should be trust
between the developers and target users.

• Privacy
Privacy is a major issue regarding CAs. User may disclose sensible
information which may be used for harm. It’s a legal issue as well(varies
geographically). The suggestion is to see legal requirements as a baseline.
 Ethical Issues(Cont’d)

• Agent Persona
Agent persona raises another issue regarding ethical issues of using CAs.
Research has shown that agent persona can directly affect creating stereotypes.
The suggestion is to developers to make agent’s persona to be androgynous and
neutral.

• Anthropomorphism and sexualization

Anthropomorphism and sexualization issues towards CAs is a rising
phenomenon. A part of user groups tend to be verbally sexual towards the
agents. The suggestion is to handle the utterance and response with some
strategies that might include no response, neutral response, in-kind response or
escalation to a human agent
 Conclusion

• There are abundant researches being conducted regarding general

purpose conversational agents. We have tried to give a summary of
literature regarding methods used to construct chatbots used for
general conversations.

• Also we have focused on the ethical part regarding the deployment of

conversational agents to the general public. There are no universal
solutions to these questions but the developers must keep certain
responsibilities in mind when designing chatbots.
 Acknowledgements
• [Bahdanau et al., 2015] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine
translation by jointly learning to align and translate. In ICLR’15, 2015.

• [Banerjee and Lavie, 2005] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for
mt evaluation with improved correlation with human judgments. In Proceedings of the acl
workshop on intrinsic and extrinsic evaluation measures for machine translation and/or
summarization, pages 65–72, 2005.

• [Higashinaka and others, 2014] Ryuichiro Higashinaka et al. Towards an open domain
conversational system fully based on natural language processing. In COLING’14, pages 928–939,
2014.

• [Hochreiter and Schmidhuber, 1997] Sepp Hochreiter and J¨urgen Schmidhuber. Long short-term memory.
Neural computation, 9(8):1735–1780, 1997.