To the user, chatbots seem to be “intelligent” due to their informative skills. However, chatbots are only as intelligent as the underlying database.
– Peter Gentsch 

Premise

Training a Dyadic chatbot as a community member in Python to interact with the African Diaspora with the right tone of voice.

Synopsis

After working for several years in advertising, public relations, and modeling I noticed that there are often only a limited number of creatives in companies with a non-Western background, which sometimes gave me the feeling of having a lack of connection with my colleagues. Here I saw an opportunity for developing a chatbot as a mentor that helps creatives from African Diaspora to build small communities where they can collaborate with each other.

Introduction

May 25, 2020. “I can’t breathe, no justice, no peace, Black Lives Matter (BLM)“.  Some words that we heard a lot after the death of George Floyd (1973 – 2020).  The death turned not only all of America upside down, but all of the world (NOS, 2020) (Hill, et al., 2020). Since the worldwide demonstrations from the BLM movement and others after this violent event, I see a lot of shift when it comes to racial discrimination in different ways. Myself, as well as the people around me with a non-Western background, are making a case for equality in the workplace, for example. Research of Sociaal Cultureel Planbureau (SCP) showed that on average only 12.6 percent of the workforce of Dutch companies in the years 2017 and 2018 consisted of people with a non-Western migration background. Taking Maslow’s hierarchy of needs into consideration here where the hierarchies of motivation are classified by rank starting from the primary need of life namely, physical needs, safety needs, love and reward, esteem, and self-actualization. They have to be fulfilled in corresponding order, there can be a lack of love and belonging, due lack of connection on the work floor.

Approach

During this project my goals to improve my general programming skills by programming a chatbot, and get a clear insight into how people feel “safe” on the platform by researching how they communicate and interact with each other.

Research phase

It is not possible for humans to do a large number of interviews without potential biases. That is why several people are trying to build and experiment with chatbots as interviewers to collect data.  In order to find out in which way people talk to each other, I assume that the way of talking and norms values between the African diaspora is different than when they are mixed with other cultures (and in this case, especially the Dutch culture) (Zhou et al., 2019).  Chatbots have a lot of potential successes when it comes to engaging in one-on-one conversations. Much progress has been made in the field of language processing used in chatbots, this ensures that communication becomes more and more natural. According to the research of Seering et al., there is way more possible than just one-to-one conversations with a chatbot. Namely, chatbots could help with social interactions, in online communities (Seering et al., 2019).

In this research, I distinguish two types of chatbots. Dyadic chatbots: chatbots that communicate on-on-one with users. They can communicate with users, with real-time interactions. Dyadic chatbots can react to what the user is saying to them. So they can ask questions or can give answers to questions. Multiparty-based chatbots: chatbots that can interact and communicate with multiple users at the same time.  This form of chatbots is often seen on platforms where groups are created, e.g. Facebook Groups, Subreddits, Twitch channels, and Slack Groups (Shewan, 2021).

• For my platform, I want to create a multiparty-based chatbot that can interact in a group, as a mentor during projects. But first I will focus on creating a Dyadic chatbot, and then do further research into how to train a dyadic chatbot to become a multiparty-based chatbot.
• Taken into account is that chatbots have a hard time talking about race and knowing when to react, these are points that should be taken into account in the developing phase (Schlesinger et al., 2018).

Exploration phase

The first step is to clarify the scope of my project. After feedback about the scope of the research, it has to be specified. This is done by answering the following questions (see below): How could the platform be more data-driven? And for which aspect can I use the chatbot? In this first phase, I’m going to focus on what do I want to test? before I’m really going to start to create things and bots. Having an interview with Coco Olakunle and Kemo Camara helped me with creating a new research scope:

• Going to test the tone of voice by scraping data.
  

Data-scraping

I looked at open data sources like Instagram and Twitter to scrape data on the following hashtags #BlackDesigners #BlackCreatives #AfricanDesginers. But data-scraping from these social media platforms doesn’t seem like the best method to get more insight into how they interact with each other, because the topics and conversations were too random. Then I got the idea to study Slack Groups in order to find out how the community is talking to each other. Shillington published an article that lists multiple creative Slack Groups. ‘Where are the Black Designers?’ is one of the groups that caught my intention. This is a group where black designers motivate and support each other on a professional level.  This group is focused on the American market however, it gives valuable information on how the target group interacts. This group could be scraped in order to create my own chatbot.

Protoyphing OpenAI & Chatfuel
Some ready-to-use chatbots were investigated like the one from OpenAi API, however, a paid subscription is needed to use this. Further research showed that there were no usable chatbots. I dived into software like Power BI and Chatfuel. But they showed very limited adaptability. So I decided that I had to focus on creating a chatbot from scratch.  This will limit my process since I will probably not be able to create an advanced chatbot from scratch. On the other hand, this will give me more insight into the back-end part of the chatbot.

Prototyping Python

First, the chatbot is trained, the packages that are used are nltk (Natural Language Toolkit), json, numpy, and keras. The nltk package contains a function needed for cleaning up text data and preprocess it before it can be used in algorithms and keras is the deep learning package that is used. The input for this model is a .json file that contains possible inputs and corresponding outputs, as shown below.

So, I started training the chatbot with the created .json file. Creating a bag of words using nltk, nltk turns human language into computer language. In this stage, I make use of a test and training set. With the training set, we train if a certain word comes gives a certain value out the .json file as output. To improve the .json file 5 iterations are made to train the chatbot. Finally, tkinter and the resulting model are used to create an interface where the user can interact with the bot. Figure 4 shows the two scripts on the left, the first for training the model and the middle for creating the chatbot, on the right the final chatbot is shown.

A built-in function of tkinter is also used to store the resulting conversation, for further analysis and topic modeling. This was quite some work since the online reporting on this package was outdated. This function is shown below.

Topic Modelling

Topic modeling is done in order to get a better insight into the topic that the community talks about. This way the chatbot could be further improved. Due to the lack of a large test group, this was not fulfilled. However, a first attempt is made to apply topic modeling to a merged textfile, including all the saved chats. This was not successful up to this point, I want to further work on it to implent it my final thesis research.

Conclusion

I developed new skills within data science as it comes to general programming, by experimenting with a dyadic chatbot. The approach for this research was to get an insight into how people interact with each other to find the right tone of voice for the chatbot. Finding the right tone of voice has unfortunately not been successful due to the lack of user-testing and scrapping data. But preparing a chatbot that can interact, store and label data is ready to use. Also, it is no longer necessary to make the chatbot multidisciplinary, because this research has given me new inspiration for using the chatbot in later thesis research. The chatbot will not be a community member, but a mentor used as a conversation tool that will store, label and process the data in a recommending system.

References

Michelle X. Zhou, Carolyn Wang, Gloria Mark, Huahai Yang, and Kevin Xu. 2019. Building real-world chatbot interviewers: Lessons from a Wizard-of-Oz field study. In Joint Proceedings of the ACM IUI 2019 Workshops, Los Angeles, USA, March 20, 2019, 6 pages.

12 Creative Slack Groups You Should Be Following. (2020, November 2). Shillington Design Blog. https://www.shillingtoneducation.com/blog/12-creative-slack-groups-following/

Hill, E., Tiefenthäler, A., Triebert, C., Jordan, D., Willis, H., & Stein, R. (2021, March 18). How George Floyd Was Killed in Police Custody. The New York Times. https://www.nytimes.com/2020/05/31/us/george-floyd-investigation.html

NOS. (2020, June 6). Dit is het verhaal achter de Black Lives Matter-beweging. https://nos.nl/op3/collectie/13842/artikel/2336375-dit-is-het-verhaal-achter-de-black-lives-matter-beweging

Schlesinger, A., O’Hara, K. P., & Taylor, A. S. (2018). Let’s Talk About Race. Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, 1. https://doi.org/10.1145/3173574.3173889

Seering, J., Luria, M., Kaufman, G., & Hammer, J. (2019). Beyond Dyadic Interactions. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, 1–13. https://doi.org/10.1145/3290605.3300680

Shewan, D. (2021, January 26). 10 of the Most Innovative Chatbots on the Web. WordStream. https://www.wordstream.com/blog/ws/2017/10/04/chatbots

Where are the Black Designers. (2020). Where are the black designers? Https://Wherearetheblackdesigners.Com/About. https://wherearetheblackdesigners.com