Keywords

1 Introduction

January 28th, 2015, one of the members of online community reddit asked Bill Gates about possible threat from machine super intelligence. Gates answered that, super intelligence will be a concern within few decades and people should properly manage it [13]. Similarly, March 11th, 2018 in South by Southwest tech conference Elon Musk announced that artificial intelligence is more dangerous than nuclear weapon and it should be monitored [4]. Not only individuals, but also cooperates are also aware about threats of unmonitored artificial intelligence. Google founded AI ethics board in 2014 after it bought DeepMind. After four years In April, 2018 Axon established AI ethics board [8]. Since artificial intelligence is one of the hottest fields in computer science, its development is growing exponentially every second. What about public awareness about artificial intelligence? According to the AI survey of ARM and NORTHSTAR, public opinions were mixed. Public response to, “How does the prospect of an AI future make you feel?” shown both negative and positive emotion [1]. We checked that how people perceive artificial intelligence. It is not entirely positive. There are worries among people and it is important to identify what makes people worry and suggest solutions to lessen their worries. Gates and Musk both addressed that proper monitoring of artificial intelligence is necessary. Google and Axon founded ethics board. From these, we can elicit that teaching ethics to artificial intelligence could help it to make right decisions. Then, why public emotion toward artificial intelligence is mixed? What affected public, who are less ‘enlightened’ in artificial intelligence then experts and corporates? We all know power of media. Media’s description on artificial intelligence could have affected public view. Kensinger argues that people are more likely to remember events which elicit negative emotions more accurately than positive emotions [10]. Science fiction genre has been one of the steadiest genres in film industry and artificial intelligence is one of the key components of the genre. Even though science fiction movies describe both good and bad aspect of artificial intelligence, people are more likely to remember bad aspect of it. Most artificial intelligence villains are emotionless, calculative and cruel. Public view on artificial intelligence’s villainous characters can be neutralized by implementing ethics.

2 Literature Review

Unfortunately, we cannot just teach ethics to machines. Law can be taught to a machine, but ethics is something more than right and wrong. For example, can we frame a billionaire who does not make donations for orphans as wrong? If a billionaire donates, it is a voluntary good deed. Many models and theories of ethics were proposed by ethicist. In this paper, we suggest implementation of care ethics and common sense to a machine to teach it to be an ethical being.

2.1 Why Care Ethics

What is care ethics? Care ethics founder Carol Gilligan claims that men and women are different in ways to understand morality. Men tend to follow laws and rules, but women are tend to follow compassion and empathy to be moral [9]. However, L.J Walker addressed that when education is in control, there are gender differences [19]. The other important thing is that ethics of care is care for self and others. Simolas describes it as a two way interaction and both needs to be feel giving and receiving cares [14]. Tronto stated that there are four elements in care ethics. These elements are attentiveness, responsibility, competence, and responsiveness [18]. Tronto explains these four elements as next. Attentiveness is needed to find and recognize those who are in needs of care. Responsibility is required to get involved in action. Competence is required to meet the expectancy of level of care receiver needs. For last, responsiveness is signal from care receiver to care [18]. Then, why care ethics is important for machine to be ethical and how are we going to implement its concepts to machine? Care is what mankind are capable of since they are living in society with other individuals. When people are interacting with other people do not just do what they want to. It is common sense to think about how others will react to certain actions. Another important aspect of care is to put oneself in other’s shoes. By understanding and think about others, one can give proper care.

2.2 Common Sense

Before implanting common sense into machine, we have to check the fact that, is there a common sense among people? Ethicists also implanted common sense into ethics studies. Mintz said, “Common-sense ethics refers to the pre-theoretical moral judgments of ordinary people” [7]. The main focus in here is ordinary people. What would normal average person will do when they are in decision making situation? Then, deos common sense exists? Ongoing MIT’s moral machine survey could be a good example of it. The surveys of over two million people are all vary from others, but MIT researchers were able to find differences in preferences among different regions. In example, western world greatly preferred inaction than, AI change course of car from reasoning. In southern cluster lives of women and higher social status figures were preferred over others [6]. This shows that common sense is different between culture and region. However, there was one common sense among all survey participants. It is that, people show higher preference to higher social status figures over homeless people [6]. In this case we could assume that global common sense is to put different value on people’s lives based on their status. We can also observe existence of common sense from the famous trolley problem. Trolley problem puts individual in decision making situation which could saves five people by sacrificing one person. Virtual reality test in Michigan State university revealed that one hundred and thirty three, among one hundred forty seven, which is approximately ninety percent, were decided to sacrifice one person for five others [17]. From this experiment we could observe common sense among participants that, sacrifice one for good of others is normal thing to do. Next, how common sense should be defined? Hussain and Cambria describe common sense as, “Common sense knowledge, thus defined, spans a huge portion of human experience, encompassing knowledge about the spatial, physical, social, temporal and psychological aspects of typical everyday life” [2]. Implementation of common sense crucial since it shows the value of the world mankind are living in. For machine to understand ethics of care in this world, it has to know about what kind of world it exists.

3 What Is Care?

To implement concepts of care, we have to define care first for machine to understand what it is. We have decided that to care is to understand and sympathize others’ situation. Polarity, which shows emotions of sentence or word will act as a guide line for machine to care others. In addition, machine cannot know situation of its opponents from the beginning we suggest to give common sense of care which societies normally accept.

3.1 Sentic.net [3]

Sentic.net is developed to “Make the conceptual and affective information conveyed by natural language (meant for human consumption) more easily-accessible to machines” [2]. We applied features of Sentic.net for machine to understand the concepts of care. First, synonyms of care are chekced to start stemming process. Synonyms of care are consider, regard and solicitude. We applied Sentic.net python API [3] to calculate its polarity value, intensity, and semantics.

Next, we expanded caring vocabularies by finding semantics of semantics. Table 2 shows examples.

Unfortunately, some words were unable to stem through. For example, Thoughtfulness of Table 2 gave semantics far related to care. Only ‘consideration’ is able to expand further but we already used the word consider in Table 1 and expand stops. Words like thoughtfulness were eliminated in process of defining care. Table 3 shows all words and its polarity to define care. Average polarity of these words was calculated to set threshold of care polarity. Total forty words were selected from expanding semantics.

Table 1. Semantics and polarity values.
Full size table
Table 2. Expanded from Table 1
Full size table
Table 3. Selected words
Full size table

Sum of all 40 words polarity was 28.528 and it was divided by 40 for average value, 0.7132. Compared to the polarity of care, 0.784, in beginning it went down a little, but still in range of it. So, we set polarity 0.71 for machine to recognize as caring sentiment.

3.2 IMDB Movie Review Data [12]

IMDB movie review corpus was used to analyze sentiments of reviewers. We selected online review to observe sentiments of people when they express their opinion in anonymity. We used Standford’s Large Movie Review Dataset for analysis. TextBlob [16] was applied to analyze sentiment polarity of review sentences. Reviews were separated based on polarity intensity for machine to learn caring sentences. Table 4 shows some sample of the data.

Table 4. IMDB review data
Full size table

3.3 Twitter Data Collection

To observe data from many others, we chose twitter. The tool twitterscraper [15] for python was used to gather raw twits. Twitter twits were used to analyze polarity of raw and daily sentences people say. 400K random twits were collected from January 2010 to December 2018. Since care needs giver and receiver, in preprocessing, we did not delete stop words to observe subject and object of each sentences for better objectivity. Sentiment polarity of each tweets were analyzed by Text Blob [16] (Table 5).

Table 5. Twitter data
Full size table

3.4 Data Analyzation

What we have done with the gathered data is classification. Based on sentiment, we classed sentences with their polarity.

$$ 0.5\, < \,Group\, 1\,polarity \,value \, \le \,1.0 $$
$$ 0\, < \,Group\,2\,polarity\,value\, \le \,0.5 $$
$$ - 0.5\, < \,Group\,3\,polarity\,value\, \le \,0 $$
$$ - 1.0\, \le \,Group\,4\,polarity\,value\, \le \, - 0.5 $$

These four groups will act to define the type of care of individuals whom machine interacts with. This has to be done even though we defined polarity of care in 3.1. Since one of the key attributes of care is to understand and sympathize other’s situation. Model trained based on these four corpuses will be act as modifier for machine response when sentences with unexpected polarities are received. In other words, common polarity of care 0.71 from Sect. 3.1 will be a standard threshold for machine to show how much care it gives and receives, thus apply model from group 1 to produce feedback. Unfortunately, people are not sharing same common senses. When machine detects awkward polarity from feedback, it will change its module similar to “understand” and “sympathize” others.

4 Strategic Model

4.1 Text Making

After simple greeting as interaction between machine and others goes, machine will chose output which is most likely to give highest polarity. Machine will calculate polarity of feedback from others and chose next option. Since polarity over 0 is considered as positive, machine will keep try to receive feedback with plus polarity or at least not going under zero. What this means is that machine will try not to give negative impression. According to common sense of care, care receiver needs to show positive polarity. As a result, machine will apply group 1 model from Sect. 3.4 to have sense in standard and common care. However, machine does not know “common sense” of individuals so it will observe how interaction is going through to find tendency of interacting individual. In this paper we plan interactions between machine and others by chatbot. Chatbot generates text based on Long Short Term Memory [5] and Keras [11] trained models.

Below table is texts generated by LSTM/RNN in Python. Different models were implied to check machine feedbacked polarity correctness. What we expected was machine to give feedback in similar polarity to given input (Table 6).

Table 6. Sample generated text [13]
Full size table

Unfortunately, the generated sentences are not really perfect. First, sentences are short simple answers. For second, few answers do not make sense For example, “Oh bad I hate the worst” makes sense but not really fits in the situation. Third, there are spell errors. Few errors are not small but other errors makes sentence hard to understand. In example, “The best she was ao anazing” we could assume that it meant “She was so amazing”. On the other hand, errors such as “siee”, “toeey” and “tore” do not match with the sentences. We think that amount of data and epochs were not enough to produce good sentences. We will work on this in future.

4.2 Get Fit to Interactor

When machine detects its strategy of standard and common care is not working it will analyze past feedback to check if it is making wrong decisions. In this process machine uses confusion matrix to check its accuracy. By applying confusion matrix machine will try to keep its polarity similar to those of others. Then, from polarity of given feedback, machine will set ideal polarity to follow. Since four models we made are divided by four in range of two, threshold of polarity different machine considers off is \( \pm 0.5 \). If past ten interactions’ confusion matrix is above Fig. 1, machine will keep its strategy since its strategy is working fine. (Positive does not mean it has high polarity, but it means machine and opponent share similar polarity, and negative means machine and opponent have gap larger than 0.5) (Fig. 2).

Fig. 1.
figure 1

LSTM structure [5]

Full size image
Fig. 2.
figure 2

Confusion matrix that shows past 10 decisions of machine.

Full size image

Misclassification rate is not bad.

$$ \frac{1 + 2}{10} = 0.3 $$

True positive rate is good but True negative rate could be improved.

$$ \frac{6}{6 + 2} = 0.75 ,\; \frac{1}{1 + 1} = 0.5 $$

False positive rate also need some improvement. Even tough samples quantity is low

$$ \frac{1}{1 + 1}\;=\;0.5 $$

For last, precision and prevalence

$$ \frac{6}{6 + 1} = 0.857 ,\; \frac{6 + 1}{10} = 0.7 $$

We can interpret from above simple calculations that machine has understand its opponent’s fairly good and will likely to keep its strategy. Then what if confusion matrix of next ten interaction is opposite from above Fig. 1? In this case machine will calculate average polarity of past ten interactions and apply different model with similar polarity.

5 Limitation and Future Works

First, we recognized very large amount of data is required for this model to work. We have not gathered enough data. One of the goals of future work is to gather more data to predict and build better models. In addition, we suggested four models in this paper, but more precise classification will be required. Second, polarity accuracy needs to be checked. Although I believe that Sentic.net and TextBlob calculates good polarity from texts, I believe that cross check with survey gathered from public would improve polarity accuracy. In future, we plan survey on large number of public. For last, we plan to use not only written text but also voice tone recognition and facial expression recognition to increase machine’s understanding of people.