Images

The gender bias in translation machines

PREMISE

Using a chatbot  to increase awareness of the gender bias in the text translated by translation machines

SYNOPSIS

Gender bias in machine translation is a critical issue that doesn’t receive sufficient attention in my opinion, and the project aims to address this gap. The project aims to address this issue of gender bias in machine translation since most translation machine users are not aware of this bias in their translater text. By creating a chatbot prototype it is aimed to inform those users about this problem. The chatbot is designed to create awareness, inform and educate users about gender bias in machine translation and how it can perpetuate gender stereotypes.

The need to address gender bias in machine translation is critical as it can impact the perception of gender roles and reinforce negative stereotypes. Therefore, it is important to equip users of these tools with the knowledge and awareness to identify and mitigate gender bias. The project’s target audience is individuals who use machine translation tools, particularly students who are likely to encounter the issue in their academic work.

To create the chatbot prototype, the project used Voice Flow. By involving stakeholders and users feedback was generated throughout the process to improve the chatbot functionality.

The chatbot was designed to be user-friendly and engaging, that informed users about Gender bias.

SYNOPSIS

Gender bias in machine translation is a critical issue that doesn’t receive sufficient attention in my opinion, and the project aims to address this gap. The project aims to address this issue of gender bias in machine translation since most translation machine users are not aware of this bias in their translater text. By creating a chatbot prototype it is aimed to inform those users about this problem. The chatbot is designed to create awareness, inform and educate users about gender bias in machine translation and how it can perpetuate gender stereotypes.

The need to address gender bias in machine translation is critical as it can impact the perception of gender roles and reinforce negative stereotypes (Savoldi et al., 2021). Therefore, it is important to equip users of these tools with the knowledge and awareness to identify and mitigate gender bias. The project’s target audience is individuals who use machine translation tools, particularly students who are likely to encounter the issue in their academic work.

To create the chatbot prototype, the project used Voice Flow. By involving stakeholders and users feedback was generated throughout the process to improve the chatbot functionality.

The chatbot was designed to be user-friendly and engaging, that informed users about Gender bias.

SUBSTANTIATION

Gender bias is a pervasive problem in the translation world, especially when using translation machines. Translation machines rely on machine learning algorithms to generate translations, and these algorithms are often trained on large amounts of pre-existing translations. However, these translations can themselves be biased and perpetuate gender stereotypes, which can lead to gender bias in automatic translation (Pinnis, 2020).

One of the most common forms of gender bias in machine translation is the use of gender language. For example, some languages have gender-specific nouns, assigning a masculine or feminine gender to certain words. Translation machines can easily reinforce these gender-specific language norms by consistently translating words in a gender-specific way, even when the source text is gender-neutral.

Another form of gender bias in machine translation is the use of gender-specific pronouns. In some languages, pronouns are gender-specific, and translation machines may use the wrong pronoun based on the perceived gender of the subject of the sentence. This can lead to confusion and even offense, especially when translating content related to gender identity or sexual orientation (Chunyu, 2002).

Translation machines can also perpetuate gender stereotypes by using gender language to describe certain occupations or roles. For example, certain occupations in English are traditionally associated with men, such as “firefighter” or “policeman,” while others are associated with women, such as “nurse” or “teacher.” Translation machines may unintentionally reinforce these stereotypes by using gender language when translating texts about these occupations (Marcelo O. R. Prates, 2019).

In short, gender bias in that needs to be addressed. Translation companies and developers should ensure that their machine learning algorithms are trained for gender-neutral translation and adopt inclusive language policies. In addition, end users should be aware of the potential gender bias in machine translations and take steps to check the accuracy of these translations before they are used (Savoldi et al., 2021). Ultimately, it is up to everyone in the translation industry to work toward more inclusive and unbiased translations.

PROJECT

During my project, I initially planned to work on recommendation systems but my interest was piqued when I attended the Chatbot workshop. I became fascinated by the potential of chatbots to educate people on gender bias in machine learning in an interactive manner. To determine the best chatbot program for my project, I experimented with four different platforms, namely VoiceFlow, Crisp, chatbot, and Dialogflow. After creating a few basic chat flows on each of them, I ultimately chose VoiceFlow due to its user-friendly interface, expansive capabilities, and appealing aesthetic.

The primary objective of my project is to create an engaging chatbot that educates users about gender bias in machine translation. To achieve this, my target audience is students who frequently use machine translation throughout their academic career but may not be aware of the potential for gender bias. I believe that students are open-minded and eager to learn about this important topic.

 

 

ITERATIONS

First iteration – Story line and first flow

Using a flowchart, the flow for the conversation for the prototype was visualized. After this visualization multiple chatbots were tested. In voiceflow the first flow was created. AI was enabled during the first iteration. When the chatbot didn’t have a perfect match it generated an outcome using AI.

Insights

  • AI gave most of the answers but therefor gave answers and reacted to questions that where outside of the scope.
  • Testers missed a fun element or something that would be more engaging
  • Testers had no idea on the problem of gender bias in machine translation. They liked that the chatbot gave them new insights.

Second iteration – Story flow improvements

During the first iteration AI answered most of the questions. The focus of the second interation was improving the answers and most of all recognizing the users answers. After adding more input and training the chatbot I observed that all answers were given at once. If the user asked for tips, it gave 8 tips at the same time, after some tweeking it gives one answer and when you ask again or ask for another tip it will give a new one. In this way the user is not overloaded with information and gets short answers.

Insights

  • Voice flow is really sensitive, if it is not 100% a match it doesn’t recognize the question.
  • AI was turned of to keep better track on the conversation
  • When giving one answer at the time, users will not be overloaded with information
  • Users are missing a fun element, something like a quiz.

 

Third iteration – Interaction improvements

During the third and last iteration the focus was on  creating  it was possible to ask questions  AI answered most of the questions. The focus of the second interation was improving the answers and most of all recognizing the users answers. After adding more input and training the chatbot I observed that all answers were given at once. If the user asked for tips, it gave 8 tips at the same time, after some tweeking it gives one answer and when you ask again or ask for another tip it will give a new one. In this way the user is not overloaded with information and gets short answers.

Insights

  • Voice flow is really gives a lot of possibilities, and recognizes a lot of questions.
  • AI can give helpfull information when the matching scores low.
  • When the same question is asked it can give different answers.
  • Users can test if they recognize gender bias with some small examples.
  • Testers liked the answers, and thought the chatbot was a good addition to a machine translator.
  • Testers liked the balance between engaging and formality of the chatbot.

CONCLUSION

This chatbot was developed to help create awareness and inform machine translator users by using Voiceflow. Based on the survey results, it can be said that chatbots help create awareness by users, almost all of the participants where not aware of gender bias in machine translation before conducting in this research. In most cases, they found the chatbot engaging, informative and helpful. There is room for improvement of the prototype. One improvement that would be really beneficial would be options in the form of buttons, but Voiceflow doesn’t has the option at this moment. The chatbot would be able to easier understand and have better interactions. The interface needs to be less confusing and it would make the conversation easier. A second improvement would be to give feedback, so users can give feedback on answers but can also give ideas for extra  information.

LITERATURE

  • Chunyu, K., Haihua, P., & Webster, J. J. (2002). Example-based machine translation: A new paradigm.
  • Marcelo O. R. Prates, P. H. A. L. C. L. (2019). Assessing gender bias in machine translation: a case study with Google Translate. https://doi.org/https://doi.org/10.1007/s00521-019-04144-6
  • Pinnis, M., & Šauperl, A. (2020). Translating gender: Machine translation and gender-neutral language. Journal of Language and Sexuality. 147-174. https://doi.org/10.1075/jls.20007
  • Savoldi, B., Gaido, M., Bentivogli, L., Negri, M., & Turchi, M. (2021). Gender Bias in Machine Translation. Transactions of the Association for Computational Linguistics, 9, 845-874. https://doi.org/10.1162/tacl_a_00401

 

Read more "The gender bias in translation machines"

NFT Memes Gallery

Premise

Memes have been a big part of online ecosystem for years and  integration with NFT technology might convert them into a digital historic archive and give their creators monetization opportunities.

Synopsis

Semester “C” weeks I will never forget. On the 24th of February, 2022 Vladimir Putin had turned the worst page in modern Russian history. Since then, we have been observing the most horrifying, ugly, and powerful cases of propaganda and fake news on both sides of the conflict. Being half-Russian/half-Ukrainian and having cousins in both countries, I see enormous polarization in opinions to the extent that people are unable to speak or listen.
Although there’s a striking phenomenon – one thing is still considered by both sides more or less equally – humor that comes in the form of MEMES. Close ethnically, Russians and Ukrainians still find very similar jokes funny, re-post them and react with mitigation of hostility.

That made me think that MEMES could be a great mirror of society, modern-day chronicles, which tend to disappear quickly. The idea is to find ways to save such newsreels, not only in connection to this ugly war but generally as a chronicle of the history, and give the authors ways to monetize them. An excellent way for it is NFT technology, and the reasons why are stated in the next section.

Substantiation

When it comes to Memes, the topic is familiar to an average internet user and perfectly fits into a concept of “shared economy,” where digital content is created, shared, re-posted, and consumed without such restrictions on copyrights and ownership, NFT technology provides significant advancement toward claiming ownership for both, tangible and intangible assets, in a more transparent, secure, and concrete form, thanks to being derived from blockchain technology.

Smart contracts for NFTs can ensure that money and assets change securely (Dowling, M., 2021) and that parties are clear about the content of agreements, reducing the need for middle-agents (Morkunas, V., Paschen, J., Boon E , 2019), so NFTs as a technology simplify the process of converting assets to tokens to facilitate ease of movement within the legal ecosystem.

There are some proven assumptions (Angelis. J., and Da Silva. E., 2019) that different NFT standards have a great chance to replace processes human-centered guarantors of authenticity via lawyers and escrow agents in industries such as property and vehicle sales.

Provable Scarcity:
Either real worlds objects or virtual(digital) objects derive value and capitalization from their scarcity(Fairfield.J.,2021). Each NFT asset can be tracked on the blockchain along with its unique details. It is possible to parse the chain data for all assets in existence and group assets by traits. So, users can independently verify collectible rarities and quantities, 100% uniqueness, and lack of duplicates now and in the future.
And unlike cryptocurrencies (such as Bitcoins or Ethers), which are “fungible” and “interchangeable”, “non-fungibility” makes an NFT a unique and scarce asset.

Freely accessible by major players within the industry:
NFT marketplaces and crypto markets, where NFTs are being traded, are pretty transparent about their items and collections and freely provide APIs to work with. As an example – the largest NFT marketplace OpenSea API:


A gap in the Market:
There’s an extensive number of channels where NFTs of all kinds are being traded. OpenSea.io (Etherium, Polygon, now Solana), Rarible(Etherium, Flow, Tezos), Nifty Gateway(Etherium), SolanArt(Solana), Axie(Robin), Binance NFT MarketPlace(BSN), Decentraland(Etherium, MANA), NBA Top Shot (Flow), SportsCoin(Flow) and other platforms are either generally NFT or specialize in rare/sports/in-games/music items. The great news is that none of them specializes in or has a dedicated NFT Memes section.

Immutable Ownership:
NFT items have a unique immutable reference written to the Blockchain. Once ownership is transferred, it’s recorded within the smart contract and can’t be edited; only the next performed transactions could be added with that NFT.

Possible limitations and barriers to entry: miscellaneous NFT standards.
When it comes to technical aspects of NFT, different industry technical standards should be taken into consideration to influence cost formation in NFT marketplaces and NFT minting.
NFT originated from the Ethereum blockchain and started using a non-fungible token standard ERC-721 in 2018 (Ethereum official documentation, 2021).
ERC-721 is widely used in many original NFTs with high capitalization and implements an API for single tokens within Smart Contract.
Later ERC-1155(2019) standard is rather used for collections of NFTs or as a combination of ERC-721 and ERC-20(a standard for fungible tokens). ERC-1155 works for all types of assets: fungible and non-fungible.
Like many pioneering technologies, ERC standards are not just widely used but with the highest transaction and gas fees.
Early adopters of the technology, such as Binance Smart Chain(BEP-721 and BEP-1155), Tezos (TZIP-12), Flow(fast, low-cost transactions, ideal for dApps environment, like NFT marketplaces and crypto games.), TRON(TRC-721 and TRC-1155), offer their own standards extending NFT applicability to many different platforms and, due to later advanced development, offer significantly lower transactional and gas fees.
So, high gas and transaction fees will be the main challenge and unpleasant surprise for an unprepared creator of a meme.

Monetization opportunities for content (Memes) creators are undoubted, see the graph below (Buchholz, K., 2021)

 

Prototypes & Experiments

Iteration #1

I have started my project with an open interview to find out if there would be some interest in such service, which brought some stunning insights. Not many interviewed understand:
1) why NFT technology could/should be used
2) what NFT is
3) if NFT is anything more than just badly overprized .jpeg images

Solution #1

The solution to the problem is either educating potencial Memes creators or “hiding” the implementation part deep into the solution, making it as simple and user-friendly as possible.

An application with “hidden” implementation of the parts, where users should be choosing cryptotokens, marketplaces, platforms for their memes and the process of converting them into an NFT itself, has been chosen as a solution.

 

Iteration #2

The simplest prototypes was tested on potencial users with options to convert a meme into an NFT with the easiest possible way.
To navigate withing the Gallery the users are able to see “Latest” added items, “Channels” to communicate in Telegramm chats, form “Groups” and “Filter” items in the gallery.
An NFT could be converted from a picture or a photo and after to be [$] placed into external marketplaces for monetisation and [#] tagged to be shared in social media.


Solution #2
The initial idea of a “gallery for novice users” turned into a practical and functional “converter” from traditional formats into NFT-format with granting a ownership priviliges.
(as a simple example, given by an interviewee – “Instagram” has 3 buttons – I use it, “FB” is too cumbersome, stopped using it.)
The feedbacks were favourable as the app dosn’t overwhelm NFT-unsavvy  user with unnecessary functionality and information.

The main challenge for the next part is to:
1) continue keeping the app as simple as possible with “hidden” from th euser technical details
2) come up with the most resource and cost-efficient way of publishing NFT memes
3) to find the best revenue model and other means of the monetisation for the Converter.

Iteration #3

At this stage of the prototype the user receives the NFT “ownership”(the link on the screen), can see the platform where the NFT is traded, [$]the wallet, observe views/likes/comments from other platforms about his/her creation.

Solution #3
The main objective – to keep the solution as simple as possible for unsavvy NFT user is completed.
After a long consideration the best revenue model turned to be a comission from the first sale of the NFT(on top of the marketplace comission), if the sale happens. Gas and Transaction costs the Gallery/Converter has to cover itself and should not bother the authors of the NFT Memes.

Conclusion:

https://www.figma.com/proto/y415bimKGqXZsG5DsUI5MF/NFT?page-id=0%3A1&node-id=7%3A117&viewport=241%2C48%2C0.23&scaling=scale-down&starting-point-node-id=7%3A43

The original idea of an NFT Memes Gallery, which main purpose would be keeping records of history in such novel and witty way as Memes, together with stimulating creativity by paying the authors (with instruments, provided by emerging NFT technology), evolved into a user-friendly Memes-to-NFT_Memes Converter.
User-Friendliness can’t be a trade-off. This trade-off became apparent during the iterations of User Research.

Reference List

Fairfield.J.(2021), “Tokenized: The Law of Non-Fungible Tokens and Unique Digital Property.” Indiana Law Journal, Forthcoming,
Available at SSRN: https://ssrn.com/abstract=3821102

Angelis. J., and Da Silva. E., (2019)”Blockchain adoption: A value driver perspective”, Science Direct
https://www.sciencedirect.com/science/article/pii/S0007681318302088

Dowling, M. (2021). “Is non-fungible token pricing driven by cryptocurrencies?” Finance Research Letters.
Advance online publication. https://doi.org/10.1016/j.frl.2021.102097

V.J. Morkunas, J. Paschen, E. Boon (2019) “How blockchain technologies impact your business model”
Business Horizons, 62 (3) (2019), pp. 295-306, 10.1016/j.bushor.2019.01.009
Zhang, Y and Cheng, H. Kenneth, (2022) “How to Sell your Crypto Art? Evidence from Non-fungible Token Art Drops”   https://ssrn.com/abstract=4047023Mazur, M., (2021) “Non-Fungible Tokens (NFT). The Analysis of Risk and Return” https://ssrn.com/abstract=3953535 or http://dx.doi.org/10.2139/ssrn.3953535Ethereum official documentation. ERC -721 Non-fungible token standard.
https://ethereum.org/en/developers/docs/standards/tokens/erc-721/
https://ethereum.org/en/developers/docs/standards/tokens/erc-1155/Sportsicon official documentation https://docs.sportsicon.com/whitepaper/overview-of-sportsiconSoRare platform official documentation (2022)
https://help.sorare.com/hc/en-us/categories/360003699737-Getting-started-

Buchholz, K. (2021) “Memes-Turned-NFTs Earn Big Bucks” https://www.statista.com/chart/24814/meme-nft-auction-prices/?utm_source=Statista+Newsletters&utm_campaign=ef0d341d89-All_InfographTicker_daily_COM_PM_KW17_2021_We_COPY&utm_medium=email&utm_term=0_662f7ed75e-ef0d341d89-315801217

 

Read more "NFT Memes Gallery"

PREMISE

Detecting the emotional state of the employees with the computer vision technique and applying music therapy to increase productivity and efficiency of the employees.

SYNOPSIS

During my term as an industrial engineer, the biggest problem of companies was that the productivity and efficiency of their employees were not stable. This is due to the fact that the emotional status of the employees varies on different days or within the same day. Music was one of the methods used to change the emotional states of me and other employees in such situations. Employees apply music therapy to themselves, both consciously and unconsciously. In this study, in order to increase the awareness of the employees’ emotional status, a song recommendation music therapy will be applied that will make the employees aware of their emotions and change their current emotion status, with the emotion status detection model to be developed with computer vision.

PREFACE

Music is an ubiquitous piece of art that has an important place in people’s lives and can be found at any time of the day in order to change their thoughts, emotions and distract their attention. While people are listening to music, their minds are imprinting the music into their minds with their current emotions without being aware of it, and whenever they listen to that song, the mind invokes these emotions without the person being aware of it (Jäncke, 2008). Because of this feature of the human mind, music therapies have been started to be applied as well as psychotherapies. While psychotherapies focus on changing people’s perspectives on their problems, music therapies are a type of therapy used to improve people’s emotional states (Mao, 2022). The emotional state is one of the factors that affect people’s productivity and efficiency. For example, while rap music makes people feel more positive, depressive music makes people feel melancholy.

Employees cannot work as productively or efficiently as they would like when they are upset, stressed, or frustrated. In other words, employees who are not happy while working have difficulty showing their performance (Mao, 2022). In addition, since there is no department in the companies regarding the psychological or emotional states of the employees, the employees try to cope with this problem on their own (Zhou, 2018). To give an example, when the famous scientist Albert Einstein could not find enough inspiration to write, he changed his current emotion by playing the violin and returned to writing again (Mao, 2022). The aim of this project is to determine the unhappy emotions of the employees during the day and to recommend music in order to turn their current emotional status into happiness with music therapy.

PROTOTYPING

PREPARATIONS

In order to understand and apply how Computer Vision technology works in practice, I did research on medium, google scholar, and various sources. My technical background and level of knowledge made it easier for me to obtain information on how this technology can be applied. As a result of the research I have done, I learned that CV technology is a method that is formed by the combination of deep learning and machine learning algorithms and is used in fields such as object detection, motion tracking, action recognition, human emotion recognition, and that it can be created both with and without coding. Since I preferred the non-coding method in this study, I used the Lobe application, and I watched videos related to this subject on YouTube in order to understand how this application works. In order to train the Emotional status recognition CV model, I first started by scraping the photos on Google images. I gathered the image data collected from hereunder 2 different labels happy and unhappy. The gender and emotion distribution of the image data is shown in Graph -1. However, many factors such as the light level in the photographs, the accessories on the people, gestures, and mimics affect the learning of the system while applying emotions. In order to prevent erroneous learning of the system, I included grayscale and color photos, photos taken both outdoors and indoors, and photos with and without accessories in both emotion data sets. Thus, I have ensured that the system understands emotional states as accurately as possible. It is important for the system to understand the emotional states correctly in order to correctly recommend the music to be recommended. I scrapped the songs to be used in music therapy and the features of the songs from the Spotify API using python and collected them in a csv file. Figure 1 shows how song features are gathered from Spotify API.

 

Iteration -1

Although the accuracy value of the model trained using the Lobe is 92%, the predictions made during the test phase do not fully reflect reality. The system cannot define an Asian person as happy during the testing phase. In my opinion, the reason for this is that his eyes get smaller when he smiles and his face looks like a sad emotion is formed. In order to solve this problem, images of Asian people’s emotions were included in the model and retrained. Also, the system cannot fully understand the unhappy emotion of people whose facial features are not obvious. In order to solve this problem, the data set has been retrained by adding data with this face feature to the system. The new accuracy value created at the end of these processes is 96%.

Before Improvement                                             After Improvement           

Iteration -2

In order to be able to change the emotional state of people, 2 playlists were created to understand if the recommended song is more effective than the songs they have listened to or not listened to before. One of the playlists consists of songs that the users may have heard before, as it is taken from the Spotify popular songs playlist, while the other consists of newly released songs. Before the test, the participants were offered songs from both lists, and the emotional status of the participants was determined by the CV model. If the song from the popular song’s playlist has not been listened to by the user, then he/she would be asked to say a song they like, and the user test continued with this song. Following the user testing, a survey was given to the users, and they said that the songs they listened to before were more effective in changing their emotional status. The survey includes questions such as felt emotional state, the emotional state according to Lobe, listened to popular and newly released songs if there was a change in their emotional state afterward, and which song helped them. I made face to face usability test with 10 people and, the distribution of the survey result is shown in Graph 2.

      Before Song Recommendation                         After Song Recommendation

         

Reflections

The process of understanding CV technology and developing the prototype can be summarized as quite challenging but also fun and exploratory. The reason I chose to study this technology, which I had not experienced before, was to push my limits. Learning, understanding, and applying a new subject has been a very developmental gain for me. Being able to do this in such a short time is proof that I have proven myself. Searching for many resources, reading articles, and benefiting from educational content was like sailing to new horizons. Exploring new technologies, Spotify’s python libraries, examining API documents, and understanding how the system works increased my desire to learn and my passion for this subject. Overcoming the difficulties, I faced with limited resources in limited time and producing solutions improved my problem-solving ability. Gaining new skills with this experience increased my self-confidence.

CONCLUSION

The emotional status changes experienced by the employees reduce their productivity and motivation from time to time. It has been revealed as a result of various studies that music has an effect on the emotional states of people. Based on this, I developed an application that will follow the employees at random times of the day and make them happy with music therapy in an unhappy emotional state. This application aims to get people out of their unhappy moods by suggesting songs with a happy mood among the songs they have listened to before. For the prototype, it was ensured that music suggestions were made according to the relevant mood on the data set, which was trained by using images of people with various facial expressions and mimics.

References

Alake, R. (2021, December 16). A Beginner’s Guide To Computer Vision – Towards Data Science. Medium. Retrieved March 28, 2022, from https://towardsdatascience.com/a-beginners-guide-to-computer-vision-dca81b0e94b4

Jäncke, L. (2008). Music, memory, and emotion. Journal of biology, 7(6), 1-5.

Lobe. (2020, October 26). YouTube. Retrieved March 25, 2022, from https://www.youtube.com/c/Lobe_ai/featured

Mao, N. (2022). The role of music therapy in the emotional regulation and psychological stress relief of employees in the workplace. Journal of Healthcare Engineering, 2022.

Rao, M. A. (2021, December 28). Realtime Face Emotion Recognition using transfer learning in TensorFlow. Medium. Retrieved March 30, 2022, from https://medium.com/analytics-vidhya/realtime-face-emotion-recognition-using-transfer-learning-in-tensorflow-3add4f4f3ff3

Web API | Spotify for Developers. (n.d.). Spotify. Retrieved January 3, 2022, from https://developer.spotify.com/documentation/web-api/

X. Juan, “The practice path and significance of group music therapy in the psychological treatment of orphans and disabled children,” Contemporary music, no. 12, pp. 37–39, 2021.

X. Zhou, “Skillful use of music therapy to relieve the learning pressure of high school students,” Psychological Monthly, no. 10, pp. 15-16, 2018.

Read more

Safe sexting; the functionality of smart contracts within sharing explicit photos

Sophie van Dael                                        1713547                              State of the Art

Premise

Exploring and explaining the functionality of smart contracts in the protection of explicit photos shared through sexting among 18 – 35-years-olds.

Synopsis

As all aspects of modern human life evolve to include online components, it is only common sense that dating transformed as well. As dating shifts to an online environment, new risks arise. A huge risk affecting those who experience (part of) their relationships online, has been the ease with which explicit photographs can be shared to others. This practice is commonly known as ‘revenge porn’.  This occurs when a partner shares intimate photos send to them in private to others. To shift the conversation and provide users with a proactive stance towards this issue rather than a restrictive one, I will be researching the functionalities of smart contracts and to what extent they can aid in combatting this issue.

Revenge porn

Revenge porn can be defined as the distribution of sexually explicit photographs to third parties with the intent to harm, or gain control of the victim (Harper et al, 2021). Over the past few years, countries have undertaken steps to set-up a legal framework to address this offense, but actionable solutions to prevent the crime from happening have yet to be proposed (Ruvalcaba, 2021). Instead, most solutions are aimed at restricting the sender, and often protect the perpetrator through a sense of invisibility (Scott et al., 2018).

Although sending explicit photos bears consequences, it does not stop people from sending explicit pictures to their partner: research reports about a third to one half of adults between the ages of 18 to 26 engage in this behavior (Benotsch et al., 2012). This is most commonly done through online acquaintance based environments (ABE), where users have relationship together (Jorgensen &  Demant, 2021). As this form of intimacy displayed by new generations is obviously desired, safer alternatives to this should be explored where protection of the sender takes center stage.

Smart contracts

Smart contracts are an important beneficial component of blockchain technology. They are self-executing contracts, in which the terms and conditions are set using software codes (Buterin et al., 2014).  The contracts cannot be forged, and data protection is at the center of this technology, they also allow for vigilant identification of users and the right to revoke access to certain data (Zhang et al., 2021).

After initial research into the technical side of smart contacts, the best way to explore and create an entry-level smart contract seems to be the program language Solidity. Solidity is a new language launched to function on the Ethereum blockchain. It bears similarities to languages such as Python, and functioned as a good starting point for me to discover the technical side of smart contracts. As participants were not knowledgeable about programming, I translated the functions of the smart contract in my prototype to a visual design, for users could not provide insight into programming.

Simple smart contract in solidity; protecting data variables.

Prototyping

To research if the introduction of a smart contract into the online exchange of explicit photos is meaningful to users, I asked 4 people who have previously shared explicit photos to be the test audience for my iteration rounds.

Iteration

Iteration 1

Prototype

As none of the participants had any experience with smart contracts, my first iteration focused on explaining the workings in the context of this problem. To ensure the understanding of the participants, I created a low-fidelity prototype through paper and attributes, such as keys, a paper contract, and a Tupperware.

  1. First iteration round

 

Testing

I tested the prototype per duo, and asked them through role-play to imagine being in the scenario of sending a photograph. I used Tupperware to resemble the safe locking of data on the blockchain. Both were given keys, which with they could access the photograph. Through storytelling, I explained the main functions the smart contracts introduces (the security layer, individual ‘keys’, and the right to revoke access by one of the partners).

Results

The very simplistic prototype ensured the participants would not be overwhelmed while clearly explaining the attributes of smart contracts. The storytelling element was essential, as it conveyed to the participants an understanding of the severity of being a victim to this crime, and the lack of self-protection. Benefits of the prototype included the right to revoke access, although the participants were unclear how this would work into real-life online situations.

Insights

  • The risk of sending explicit photos is noted but not feared to a level that results in abstaining.

Solutions focused on restriction are futile.

  • The trust for the partner, combined with a sense of duty, trumps the potential consequences.

The solution needs to be easily integrated into their habits, too much effort will result in non-use.

  • A proactive need for a way to protect the sender is necessary, as the trust alone is not enough long-term.

Protection of the senders should be prioritized.

Iteration 2

Prototype

The main question left unanswered by the participants was how the smart contract would function in real life. To display this, I created a paper prototype displaying the attributes of the smart contract. These are:

  • Both parties can only access the photo with a mutual ‘key’, which they can use.
  • When the relationship has ended, one can revoke access.

2. Attributes of the smart contract, paper prototype

Testing

The second testing phase focused on the selection of the parameters and attributes of the smart contract. Through co-creation, participants were encouraged to critique and add to said smart contract. To implement the smart contract into their daily lives, the prototype showed a texting environment similar to WhatsApp.

  1. Co-creation, suggestion of alternated attribute by participant.

Results

All participants valued the prototype being integrated into a familiar environment. Getting an access ‘key’ to see the photograph stimulated a sense of safety and was important to all participants. One participant remarked the importance of not limiting revoked access to only the end of the relationship, but rather be an option to the sender at all times (in event of a fight, cheating etc.)

Insights

  • The right to revoke access to shared photos should be more flexible, the sender should be able to set the parameters.

Access should be broadened so access to photos can be revoked at any given moment.

  • The participants value the smart contract and its attributes, but they are not likely to seek this out alone.

The smart contract should be employed within a familiar environment.

  • Participants suggested the ‘no screenshot’ function of Netflix as extra protection.

Photos cannot be screenshotted.

Iteration 3

Prototype

4. Attribute of online prototype; the red button allows the sender to restrict access.

To implement the changes into a visual, online design, all feedback was realized into an updated online prototype in Figma, that displays the working of the smart contract.

Testing

Participants were asked to interact with the prototype online and judge the overall concept according to a user-experience questionnaire.

Results

The participants appreciated the implementation of the smart contract into their familiar surroundings. They acknowledged the change in perspective it brought into the consequences of sending explicit photos and the lack of self-protection currently available. They were pleased with all implemented features.

Insights

  • It is important for participants to be able to retract consent to photos at any given time.

A key value of the final prototype is the autonomy of the sender.

  • The prototype shifts the perspective of participants to better protect themselves from their partners distributing their photos without consent.

The prototype offers insightful information to the possibility of protecting the sender which is currently not provided by traditional online channels, such as WhatsApp.

Further research

For future perspectives, it would be beneficial to investigate the technical possibilities and challenges of creating a plug-in smart contract, and if the Ethereum network is a mere option, or if there are other ways of pursuing this (possibly even outside of blockchain).

References

Benotsch E., Snipes D., Martin A., Bull S. (2013). Sexting, substance use, and sexual risk behavior in young adults. J. Adolesc. Health 52, 307–313.

Harper, C., Fido, D., and Petronzi, D. (2021). Delineating non-consensual sexual image offending: Towards an empirical approach. Retrieved from; https://www.sciencedirect.com/science/article/pii/S135917892100001X

Jørgensen, K. E., & Demant, J. (2021). Shame, shaming and economy: A theory of image-based sexual abuse within different online sharing environments. First Monday, 26(4). https://doi.org/10.5210/fm.v26i4.11670

Ruvalcaba, Y., & Eaton, A. A. (2020). Nonconsensual pornography among U.S. adults: A sexual scripts framework on victimization, perpetration, and health correlates for women and men. Psychology of Violence, 10(1), 68–78. Retrieved from; https://doi.org/10.1037/vio0000233

Scott, G., Brodie, Z., Wilson, et al. (2020) Celebrity abuse on Twitter: The impact of tweet valence, volume of abuse, and dark triad personality factors on victim blaming and perceptions of severity. Retrieved from; https://www.sciencedirect.com/science/article/pii/S0747563219303462?casa_token=pebeQGm_miwAAAAA:SxmupLgJCREi3e_eYmMIaOS4NR8uN5Hn-RXQtU8j5EeudV7vro8ywSfqTSxXIJ02QZqyKi4wxvEB#bib40

Zhang et al., (2021). Secure and Efficient Data Storage and Sharing Scheme Based on Double Blockchain. Retrieved from; https://scholarworks.bwise.kr/sch/handle/2021.sw.sch/2215

Read more "Safe sexting; the functionality of smart contracts within sharing explicit photos"

Combining GSR data and music to reduce stress

PREMISE

Exploring ways to create an effective stress management method using GSR data and music

SYNOPSIS

Stress is undoubtedly a pandemic by itself. With so many different stressors that we have no control of the only thing we can do is focus on the things that we have control over – our mind and body. There are many proven techniques to manage stress such as exercise or guided mindfulness practice, however, most of them have their limitations and are not easily accessible to everyone. I wanted to find a way that would not only be highly accessible and fun to use but also be evidence-based. I chose to explore ways how we can use GSR biosensor to measure people’s emotional response to relaxing music and recommend them songs that will further enhance their stress reduction effectiveness.

PREFACE

Nobody is immune to experiencing stress, regardless of age, sex, race etc., however, a survey by the American Psychological Association revealed that Gen Z (15-21-year-olds) and Millennials (22-39-year-olds) in the US reported higher than average stress levels and higher than any other generation (APA, 2018). Similarly, a study in the Netherlands found young people to be the biggest population group (82%) at risk of burnout in the past year, a high increase from the previous year (NCPSB, 2022). While the main reason identified for such spike in numbers was Covid-19 related stressors such as job insecurity, financial problems, school-related issues, affected relationships etc. (Graupensperger, Cadigan, Einberger, & Lee, 2021), these and similar issues have been weighing down young people for years before the pandemic. One of the most mentioned causes of stress in young people in the existing research is academic-related.

When stress is not managed it can wreak havoc on a wide range of areas of an individual’s life. According to Hazen et al. (2011), as a response to stress, young people commonly experience anxiety, moodiness and irritability, as well as cognitive issues such as difficulty concentrating, which has a negative impact on students’ academic performance (OECD, 2017). Over time, unmanaged stress paves the way for more serious mental health issues such as depression (Kessler, 2012). Physical health is also at risk, research has shown that stressed people are less likely to exercise and more likely to overeat (Dallman et al., 1993; Stults-Kolehmainen & Sinha, 2013).

One of the most common ways students attempt to manage stress is by listening to music (54.8%), others watch internet videos, talk to friends and even indulge in substance abuse (Mahani & Panchal, 2019). Although there is no data on how popular stress-management apps are among young people, a study found that roughly a half of them were not evidence-based and likely were not effective anyways (Coulon et al., 2016). Given the current state of stress and the highly detrimental consequences it has on young people, there must be a solution that is easily accessible to everyone, evidence-based and engaging. With the currently available technological affordances such as wearable biosensors and an abundance of stress management research in the past decades, there is an opportunity for creating a data-driven solution for this problem.

PROTOTYPING

Building a song database

To begin building the prototype, I first needed to have a database with a lot of different potentially relaxing songs. To achieve that, I created a Spotify playlist and added 50 songs that I found in other playlists curated by Spotify (i.e. “Calming Classical”, “Deep House Relax”, “lofi beats”, “Stress Relief” etc.). Then, using Spotistats, I was able to export the playlist data with all the song characteristics such as tempo, genre, valence, acousticness, instrumentalness and many more. After some data cleaning with python, this is what it looked like:

Inducing stress

Before I start testing songs on people, I need to make sure they are in a stressed state. After a few unsuccessful attempts to induce stress, I found a method that worked relatively well. It is called the mental arithmetics task. The participant is asked to subtract 7 from 500 and then repeat as fast as possible (500-7, 493-7, 486-7 etc.). Below you can see how well it worked over 30 seconds – the GSR reading went from around 280 to above 300.

 

Iteration 1

Testing the songs and creating a recommender system

 

Arduino Uno board and GSR sensor by Grove

 

Now that I am able to put the participant in a stressed state, I can start measuring stress. To do that, I started by playing a random song while the participant was wearing the GSR electrodes on their fingers. This is an example of the GSR response to one of the songs:

 

As you can see the GSR dropped in the beginning and then started slowly increasing.

 

I repeated the process together with the mental arithmetics task for seven different songs and stored the data in separate JSON files. At this point, I am still not sure what is the best way to make meaning out of all the ups and downs in the GSR data that I collected. Therefore, for the initial version of the recommender system, I simply used the difference between the first (t=0 sec) and last (t=500 sec) GSR values registered, to determine how effective was the song in reducing stress. I added these values ( “before”, “after”, “difference” ) to the data frame for every song I tested on the participant.

Building a recommender system

I now have a data frame with song titles and their features, as well as GSR data indicating how well some of the songs worked on the testing participant. Below you can see that the song called “Hold On” worked the best out of the other few that I tested. It decreased the GSR by 67 units.

Now I want to build a recommender system that will be able to recommend the user songs that are similar to the one that performed the best. This type of recommender system is called content-based filtering. I will be using a cosine similarity machine learning model, which compares the similarity of two feature vectors. The first step is to select the features I want the recommender system to base its recommendation off and transform them into the range between 0 and 1. I did it using MinMaxScaler from sklearn and used the following features: ‘Energy’, ‘Key’, ‘Loudness’, ‘Mode’, ‘Speechiness’, ‘Acousticness’, ‘Instrumentalness’, ‘Liveness’, ‘Valence’, ‘Tempo’. Then using the sklearn’s cosine_similarity library and the code I found and adapted from machinelearninggeek.com, I was able to generate a list of top 10 songs from the database that were the most similar to “Hold On” (the one that performed the best so far).

The recommended songs

 

Iteration 2

Testing the recommendations with an app interface

Then l tested the first couple recommendations to see how they performed. In addition, I created an interface for the music player with the recommender system, which I also tested out with the user. One feature of the interface is seeing your GSR numbers in real-time. I personally like seeing them because it helps me concentrate on relaxing and my breathing, however, I can imagine that for some people it might be stressful to watch them, especially if the numbers are increasing. Unfortunately, I was not able to implement this feature into the design due to my limited time and skills. Therefore, I tested this feature by asking the participant to watch the live numbers on the serial monitor of Arduino IDE on a laptop.

 

The top two recommended songs were “Romeo” and “Mutual Feeling”. The first song performed relatively well with a decrease of 28 units, however, the second only showed a decrease of 7. However, this is not surprising since the dataset is very small. Regarding the interface, the participant prefered not to see the real-time data for the reason that I suspected, it was too distracting.

Iteration 3

 

Although there is evidence that specific genres and types of music will have a universal relaxing impact on one’s body, a number of studies have also shown that when exposed to one’s preferred genre of music, the indicators of relaxation can significantly increase and anxiety may decrease compared to when exposed to other unfamiliar music (Bernardi, 2005; Salamon, 2003). Therefore, for the third iteration, I decided to let the user choose their own relaxing song that they are familiar with. In addition, I changed the interface based on the user feedback and enabled an option to hide real-time data. Also, to improve the user experience and the effectiveness in stress reduction, I added a breathing guide that tells the user when to inhale and exhale. Such diaphragmatic breathing can significantly reduce stress after just a single practice (Arsenio & Loria, 2014). Although I was not able to implement it due to time and skills limitations, I imagine it to send light vibrations via the phone or a wearable device indicating when to inhale or exhale. However, I asked the participant to take deep breaths during the listening session.

 

The participant chose a song that they personally find relaxing and listened to it while taking deep breaths. The song performed rather well, GSR went from 280 to 230 in the first 40 seconds, likely due to the deep breathing too.

Listening to a personally preferred relaxing song while taking deep breaths.

 

REFLECTION AND CONCLUSION

Overall, the process of testing GSR in response to music turned out to be trickier than I thought. It is difficult to ensure the participant is in a stressed state over and over again as they become habituated to the stressor. There are so many other factors that should be accounted for to ensure accuracy in testing, such as room temperature, testing time etc. After doing the experiments, GSR seems to be quite an appropriate indicator of one’s emotional arousal in response to music, however, it would be more powerful if it was combined with other biosensors such as HR monitor. The recommendation system should be tested further with a larger and better organised data set but I think having quite an extensive variety of song characteristics to train it with was very helpful. Lastly, incorporating breathing techniques proved really effective in reducing stress. There is potential to make it even more engaging and data-driven by combining the breathing guide with HR data.

REFERENCES

APA. (2018, October). STRESS IN AMERICA<sup>TM GENERATION Z</i>. Retrieved from https://www.apa.org/news/press/releases/stress/2018/stress-gen-z.pdf

Arsenio, W. F., & Loria, S. (2014). Coping with Negative Emotions: Connections with Adolescents’ Academic Performance and Stress. The Journal of Genetic Psychology, 175(1), 76–90. https://doi.org/10.1080/00221325.2013.806293

Bernardi, L. (2005). Cardiovascular, cerebrovascular, and respiratory changes induced by different types of music in musicians and non-musicians: the importance of silence. Heart, 92(4), 445–452. https://doi.org/10.1136/hrt.2005.064600

Coulon, S. M., Monroe, C. M., & West, D. S. (2016). A Systematic, Multi-domain Review of Mobile Smartphone Apps for EvidenceBased Stress Managementsa. American Journal of Preventive Medicine. https://doi.org/10.1016/j.amepre.2016.01.026

Dallman, M. F., Strack, A. M., Akana, S. F., Bradbury, M. J., Hanson, E. S., Scribner, K. A., & Smith, M. (1993). Feast and Famine: Critical Role of Glucocorticoids with Insulin in Daily Energy Flow. Frontiers in Neuroendocrinology, 14(4), 303–347. https://doi.org/10.1006/frne.1993.1010

Graupensperger, S., Cadigan, J. M., Einberger, C., & Lee, C. M. (2021). Multifaceted COVID-19-Related Stressors and Associations with Indices of Mental Health, Well-being, and Substance Use Among Young Adults. International Journal of Mental Health and Addiction. https://doi.org/10.1007/s11469-021-00604-0

Hazen, E. P., Goldstein, M. C., Goldstein, M. C., & Jellinek, M. S. (2011). Mental Health Disorders in Adolescents. Amsterdam, Netherlands: Amsterdam University Press.

Kessler, R. C. (2012). The Costs of Depression. Psychiatric Clinics of North America, 35(1), 1–14. https://doi.org/10.1016/j.psc.2011.11.005

Mahani, S., & Panchal, P. (2019). Evaluation of Knowledge, Attitude and Practice Regarding Stress Management among Undergraduate Medical Students at Tertiary Care Teaching Hospital. Journal of Clinical and Diagnostic Research. https://doi.org/10.7860/JCDR/2019/41517.13099

NCPSB. (2022). AD: Tachtig procent van jongeren zit door corona tegen burn-out aan. Retrieved from https://nationaalcentrumpreventiestressenburn-out.nl/ad-tachtig-procent-van-jongeren-zit-door-corona-tegen-burn-out-aan/

OECD. (2017). Most teenagers happy with their lives but schoolwork anxiety and bullying an issue. Retrieved from https://www.oecd.org/newsroom/most-teenagers-happy-with-their-lives-but-schoolwork-anxiety-and-bullying-an-issue.htm

Student Development, 51(1), 79–92. https://doi.org/10.1353/csd.0.0108

Stults-Kolehmainen, M. A., & Sinha, R. (2013). The Effects of Stress on Physical Activity and Exercise. Sports Medicine, 44(1), 81–121. https://doi.org/10.1007/s40279-013-0090-5

Read more "Combining GSR data and music to reduce stress"