User:DBINiels/sandbox

1. What is NLP?
The main objective of Natural Language Processing (NLP) is to find structures and patterns so a computer can understand, generate and manage a natural language (Helgesson Törnqvist & Stålhandske, 2017). It is an old technology, but due to the recent exponential growth of data, smarter algorithms and faster processing, NLP technology has evolved in such a way that it is now able to

emulate the human brain (Deloitte, 2018). The developments are one of the reasons IBM’s supercomputer, Watson, could win the quiz show Jeopardy (Ferrucci et al., 2010).

NLP is able to understand the sentimental meaning of a sentence, by adding sense relation, part of speech tagging and combining sequences of words rather than analyzing word by word (Cavnar et. al., 1994).NLP is applicable in a lot of different areas in business because most of the data is unstructured and NLP can structure unstructured data.

Common examples to mention are reports, email, social media updates or any other documents containing natural-language text (Russom, 2007). In addition to this, it has certain characteristics, that makes it attractive for businesses to use and help them disrupt different fields.  

Since this wiki is not elaborating on the technology behind NLP, the video below is incorporated in order to gain more understanding about the technology when needed.

2. Short history
During the first phase of the history of NLP, from 1950 to 1965, machine translation received considerable attention (Bates, 1995). “NLP was originally distinct from text information retrieval (IR), which employs highly scalable statistics-based techniques to index and search large volumes of text efficiently” (Bates, 1995). But in that time, translating between one natural language and another was much too complex to be expressible (Bates, 1995). Later on, in the 1960s, the focus turned to question answering and some beneficial and successful Natural Language systems were developed (Chopara, Prashar & Sain, 2013). This first phase was a period of enthusiasm and optimism. It is remarkable how much was done with such poor resources (Jones, 1994). The first NLP workers achieved a lot. Those NLP workers recognized, and attempted to meet the requirements of computational language processing, investigated many aspects of language, addressed the issues of overall system architectures and processing strategies, and began to develop formalisms and tools (Bates, 1995).

The second phase, in the late 1960s to late 1970s, NLP work was Artificial Intelligence (AI) flavoured, with much more emphasis on world knowledge and its role in the construction and manipulation of meaning representations (Jones 1994). The focus was on the kinds of natural language that would be produced by a person interacting with a computer system. Every word in the input had some effect on the meaning that the system produced (Bates, 1995).

Altogether, the third phase (from the late 1970s to late 1980s) can be seen as one of growing confidence and consolidation phases. Partly encouraged by the general enthusiasm, but also well-justified by the ability to build better systems (Jones, 1994). Besides this, people saw the commercial aspect of NLP. Next, researchers began to think that the previous goal of NLP – process every word of the input as deeply as necessary – was difficult to achieve and perhaps unnecessary. Instead, partial understanding began to be seen as a meaningful and useful goal (Bates, 1995). By that time it was discovered that systems which tried that – extract at least some meaning from every input- could succeed better than systems which tried to achieve the old goal (Bates, 1995).

Jones (1994) states that the latest phase: late 1980s and onward, can perhaps be labeled as the “massive data-bashing period”. That is because the focus is into statistical language data processing. Besides this, there was a strong interest in a wide spectrum of tasks that require NLP during this period. “Thus NLP, earlier not found to be sufficiently useful for document retrieval based on abstracts, may contribute effectively to searching full-text files” (Copra et al. 2013).

3. Characteristics
The specific digital characteristics NLP has that distinguishes it from nondigital technology will be discussed with examples.

3.1 Connectivity

Internet of Things (IoT) has great potential in disrupting the world as we know it. However, it has a major challenge in conveying data about the different interconnected devices (sensors and objects) back to the user in a simple human understandable way. In order to realize this, the machine should be able to understand the context, which can be achieved by a chatbot (Kar & Haldar, 2016). A chatbot is a computer application with which users can conduct a dialogue in natural language as if it was conducted with another person (Abu Shawar & Atwell, 2007). For IoT to reach its full potential, the applications should be connected to each other. This can easily be explained by a short example.

Example IoT & NLP about a smart heater (Kar & Haldar, 2016):

User: “Make the living room temperature comfortable.”

Chatbot: “Since the weather outside is 17 degrees Celsius, I am setting the living room temperature to 21.4 degree Celsius.”

In this example, the user’s demand was relatively vague. However, the Chatbot could understand the user’s demand because of NLP, and then the chatbot can use the contextual information from real-time temperature along with knowledge of historical user preferences to perform the specific action. In such a short case the chatbot should be connected to the sensors and actuators of the heater. NLP chatbots could also be used to help the connectivity of firms and customers (Arain et al., 2018). The chatbot in this study was used as a quick way to acquire information without the need to cover the entire website.

The Swedish bank Swedbank also has a chatbot, Nina, which helps customers with easy questions, either by responding with a natural answer or linking them to a web form (Wasankhasit & Nilsson, 2018). As a result, Nina has helped Swedbank improve the customer experience for consumers, with a 78 percent first-contact resolution within the first three months, and it is still improving its resolutions (Hill & Richter, 2016). This is very helpful due to the increasing pressure on contact centers (Deloitte, 2018). Due to the low marginal costs, it enables Swedish bank to decrease its pressure on the customer service. Because more and more customers are using Nina, more data gets generated that improves the chatbot through machine learning. This results in even more customers that will be satisfied by Nina's services and use it more often. The chatbot is getting more valuable as the number of users increases, this is called network effects (Mitomo, 2017).

In the video about the Tay.ai chatbot of Microsoft by Singularity University below you'll find an example of a negative side of this development.

3.2 Reprogrammability

Machine learning (ML) algorithms can generally be divided into two different sub-fields: unsupervised and supervised (Helgesson Törnqvist & Stålhandske, 2017). Unsupervised algorithms, use no prior labeling to recognize different natural clustering in a data set (James et. al., 2013), and in supervised learning, the labels are known before the experiment.

An example of supervised learning is Lexalytics trying to recognize name entities. They train an algorithm to recognize a name of a person, product, place or a company, and then apply it to a new text, where the algorithm should be able to recognize the names (Lexalytics, 2018).

An example of unsupervised learning is clustering. When there is a problem where words should be clustered, K-means is a method that is usually applied due to its simplicity (Jain, 2009). However, K-means gives you only a local optimum (Steinley, 2003) and therefore there is room for improvement, which can be less time consuming because it can be learned unsupervised. So the algorithm can be reprogrammed in two ways: supervised or unsupervised.

In order to see the difference between supervised and unsupervised learning in a real life example, the video on the left is incorporated in order to explain it through the usage of one of the most famous datasets, the Iris dataset.

3.3 Modularity

There are different NLP tools on the market. At the moment there are also modular NLP tools where a module from one tool can be used for another tool. This means that a tool can be modified with these modules in order to tailor it for a specific usage, for instance there are modules for switching to a different language, for example from English to Spanish (Agerri, Bermudez & Rigau, 2014).

3.3.1 Ecosystem of NLP (layered modular architecture)

An ecosystem of technology describes an product platform made up of its key components at different levels ranging from the physical hardware that houses it to the way the software is populated with information.

An example of an ecosystem built around NLP is its use in voice-activated personal assistants such as Alexa and Siri. NLP can be used to interpret and record a lot of data on what phrases and tasks the customer uses the system for. For example, if a user asks Alexa to buy a pair of headphones from Amazon, the system can record and interpret what exactly the user is looking for. These digital traces provide information to Amazon about what phrases that particular customer used to find that certain product. This information is a valuable asset for Amazon since they could take the details and sell it to the firm producing the headphones to help them build their business strategy.

Since NLP is a technology used in so many applications, it has a unique layered modular architecture.

On the device layer, we have any range of technological devices such as an in-home voice-activated assistant (ex. Echo Dot), a smartphone, or a computer/laptop. Any device that can connect to the internet. In 2017 there were an estimated 8.4 billion connected devices and this number is expected to go as high as 20 billion devices by the year 2020 (Knoll, 2018).

On the Network layer, there are wifi/4G connections and internet browsers/internal processes. Internet browsers are a key component when using NLP on a laptop, computer, smartphone since every online service will have a website/application that can be accessed through the internet. An internet connection, whether through wifi or data, is required to send the information to the firm collecting the data.

In the service layer, it could be considered that any application using NLP is considered a service. The more firms or technologies that use NLP helps progress the technology through network effects.

Two pieces of technology that could also be considered as contributing to the service is the use of HMM and Semantic Analysis. These two processes combine to be the backbone of NLP allowing it to fully understand and categorize the data it is receiving from the data sources in the contents layer.

Finally, we have the contents layer, which is primarily made up of the users themselves. Users are the customers that are using the devices mentioned in the device layer. Anyone using a device that uses NLP is constantly adding content to the technology. For example, every time a customer uses their Echo dot to use Alexa, they are contributing useful data that can be interpreted and used by firms.

4. Consequences of Characteristics
Like explained above, the NLP technology has different characteristics. Each of these characteristics brings consequences for other systems, applications and technologies, for ways of working, and usage of data. These consequences are explained in more detail below.

A consequence of connectivity is the interoperability. Interoperability is the ability that a product or a system can work together with other products and systems. From the example of chatbot given earlier, different applications of Internet of Things should be connected to each other in order to make use of the chatbot function. Helmann, Lehmann, Auer & Brümmer (2013) argues that simplifying the interoperability of different NLP tools will facilitate the building of sophisticated NLP applications as well as the combination of tools and might ultimately yield a boost in precision and recall for common NLP tasks. In order to simplify the combination of tools, the management of a firm must improve their interoperability and facilitate the use of Linked Data (Hellmann et al. 2013). This data is used to develop the NLP Interchange Format (NIF), which aims to achieve interoperability between NLP tools, language resources, and annotations on three layers: the structural, conceptual and access. The structural interoperability refers to the way annotations are represented and therefore how these annotations merge consistently from different tools. The interoperability for different NLP applications is therefore important for having connectivity between the NLP tools.

Helmann et al. (2013) state that interoperability of different NLP tools will facilitate the building of sophisticated NLP applications. NLP therefore creates new applications as new products because NLP as a digital product can be edited and reprogrammed by the supplier or autonomously. A consequence is the emergence of new functionalities (Berends, 2018). Formation of differentiation between NLP products, product versioning, and back-and-forward compatibility are examples of emerging functionalities. Another example where things come together regarding NLP is platformization. This is a consequence of modularity discussed earlier. A module is a unit whose elements are powerfully connected among themselves and relatively weakly connected in other units (Berends, 2018). Because of those elements together, a platform can be created around the technology behind NLP or about a NLP system of a particular language.

The technology of NLP depends on the algorithm behind it. The algorithm can translate words into codes which describe something. “Brands track conversations online to understand what customers are saying and glean insight into user behaviour”  (Kiser, 2016). Due to the fact that you can track sentiment as positive, negative or neutral by the tone of a written message on, for example, Facebook or Twitter. Therefore, you don’t need more algorithms to analyze this data because the algorithm can do the same for different kinds of data. If this is compared to analyzing manually, it is not required to have more people analyzing the data. Therefore the marginal costs of analyzing the data will decrease due to the NLP technology. Besides reducing the marginal costs, another consequence of the algorithm technology behind the NLP tool is the promise of continued improvement. The algorithm learns and improves itself to be better (Brownlee, 2017). As a result of the deep learning factor of NLP is based on real results, the improvements appear to be continuing and perhaps speeding up.

Like stated earlier, data can be used in different ways. Therefore the technology of NLP could also lead to a wake of innovation. If data was not meaningful to analyze four years ago, maybe it’s relevant for analyzing right now since technology has progressed and improved so more insights can be gained from the same data set that is now more accessible because of the digitization. Since the algorithm of NLP analyses the data very quickly, you can use the data in different ways. The management of a company can therefore re-use the data by analyzing it again for different purposes.

4.1 Multi-sided platforms

As addressed above, platformization is a consequence of the NLP technology. Like Hagiu & Wright quote in their article (2015): ‘A multi-sided platform is an organization that created value primarily by enabling direct interactions between two (or more) distinct types of affiliated customers’. On the multi-sided platforms discussed below, NLP information is shared for free with different kinds of customers.

There are already existing (multi-sided) platforms for NLP and still, a lot of platforms are emerging. “AllenNLP, a library for applying deep learning methods to NLP research, which addressed these issues with easy-to-use command-line tools, declarative configuration-driven experiments, and modular NLP abstractions”, (Gardner, Grus, Neumann, Talford, Dasigi, Liu & Zettlemoyer, 2018) is an example of such an NLP platform. This platform emerged in 2018, and has already increased the rate of research experimentation and the sharing of NLP components at the Allen Institute for AI. Besides platforms that focus on the methods of NLP research, there are also platforms explaining the models of Natural Language Understanding. An example of such a platform is GLUE. This is a Multi-Task benchmark and analysis platform for Natural Language Understanding (Wang, Singh, Michael, Hill, Levy & Bowman, 2018). This platform is designed to favour and encourage models that share general linguistic knowledge across tasks, including question answering, sentiment analysis, textual entailment, and it is an associated online platform for model evaluating, comparison, and analysis. This variety of tasks can be conducted specifically by research and business firms.

In addition to the NLP platforms focusing on research and methods, there are also platforms with the emphasis on processing documents. Canary is such an NLP platform for clinicians and researchers (Malmasi, Sandor, Hosomura, Goldber, Skentzos & Turchin, 2017). This platform is designed for processing documents to support the extraction of information from natural language text using user-defined grammars and lexicons. The software guides the user through the steps to successfully build a rule-based language model for the identification of a particular concept or a set of concepts (Malmasi et al. 2017).

All these NLP platforms can exist in a specific language. Language Technology Platform  (LTP) is an example of such a platform (Che, Li & Liu, 2010), which focuses on the Chinese Language Technology for NLP. Another platform is ITU: a Turkish NLP web service (Eryiğit, 2014). The Chinese language belongs to one of the central languages. These languages are making continued progress in terms of accuracy and speed. Besides Chinese, this improvement is seen mostly for a handful of languages such as English, Japanese, German, French, and Russian. “These are the languages which consume the most research funding in NLP and for which most NLP applications have been developed” therefore are called ‘central languages’ (Streiter, Scannell & Stuflesser, 2006). The other ‘non-central’ languages have been able to build substantial NLP resources and are now situated close to the center. However, these languages are not supported by consistent investments in NLP technology and often lack money, infrastructure, an academic environment, commercial interest, and suitably trained personnel.

Research centers will work on those central languages to compete for funding and recognition, each center hopes to obtain a relative advantage over its competitors by keeping developed resources inaccessible to others. This same phenomenon occurs with corporations making investments in NLP technology. For these corporations, it is a waste of time and money to invest in non-central languages. More advancements could be realized in this field by central languages making their information open-source to non-central languages, giving them a chance to compete.“Without direct competition, a research center should suffer no disadvantage by making its resources publicly available” (Streiter et al. 2006).

5. Implications for innovation managers
For innovation managers, it is very important to know whether or not a technological innovation is disruptive or sustaining because the implications for their company can be huge. As a consequence of the type of innovation, Adner & Snow (2010) defined three types of strategies that an incumbent firm could choose for this innovation: Defensive, Offensive or Retreat. Which option a company chooses depends on whether they view NLP technology as an opportunity or a threat.

5.1 Disruptive or Sustaining In this section, there are a few examples of how NLP is a disruptive or sustaining technology depending on the industry. When exploring industries that have been affected and disrupted by the introduction of NLP, there are many areas that need to consider. Primarily the following examples can be viewed through the lens of servitization, which means that within these industries, NLP is integrating products and services into complex digital systems. 5.1.1 Marketing Industry

The first example we can explore is the use of NLP in marketing. Marketing firms and departments have seen the ways in which NLP has disrupted the way they conduct and analyze market research. Previously, when marketing groups (whether a firm or department) wanted to learn about how their potential customers feel about one of their products they would gather a handful of individuals in a room and ask them questions regarding their work. This can be an example of overshooting compared to using NLP. The core purpose of a focus group is to listen to people in order to understand and learn from them based on their social processes and methodology (Liamputtong, 2011). The results of this research would influence the changes or adjustments firms would make to their current products or messaging. While as an effective research method, focus groups take a long time to conduct and are very costly. All the participants need to be selected, questioned, and have their answers analyzed and composed into a digestible format. This information can be transmitted and stored at a much lower marginal cost than it would take an entire team to conduct a focus group. Additionally, focus groups can be seen overshooting firms needs and expectations by producing more in-depth results over a much longer period of time compared to NLP. As digital technology continues to make everything progress faster and evolves faster, firms and organizations value getting a snapshot of information quickly and easily rather than conducting months of in-depth research, even if this may lead to less robust results (Sample, 2018).

NLP allows marketing firms to de-couple the physical process into a digital form and to converge different industries, such as market research and analytics, which are characteristics of the disruptive technology. When customers conduct their online activity, for example, leaving a negative Amazon review on a blender or texting their friends about how much they would love to own the new BMW, all of their actions leave digital traces that can be easily analyzed on popular websites, such as the examples seen in Figure 1.

All of this primary research can be interpreted and delivered to a firm as propensity signals indicating a positive or negative attitude towards a product or service (Srihari, 2015).

To summarize, NLP is disrupting the marketing industry because firms now have a more convenient way of getting insights from their customers at a lower price point. In addition to allowing firms to spend less money on market research, NLP enables organizations to see an expedited trajectory through machine learning and crowd-based improvement.

On the other side of this market, we can find that NLP can be seen as sustaining to other industries, such as online marketplaces and social areas. For example, Amazon, which sells over 350 million products in the US alone as of May 2016 (Retail Touch Points, 2016). Amazon has the ability to collect all of the information their customers are saying about items in their marketplace, on mobile, virtual assistant, computers, and could choose to sell the information they collect through NLP to the firms selling items through their marketplace. Thus they would be offering a superior product to the firms that are not utilizing NLP at a higher cost point.

Figure 1. Gemalto. (2017). Platforms that could use NLP to gain insight from users digital traces and online activity. 5.1.2 Security and Intelligence Industry Similar to marketing, the intelligence and security industry is another example of how NLP is disrupting industries on a more critical level through the way it collects and analyzes data.

To put the amount of data that average police officers generate per case into context, in the example of the Criminal Investigations Department (CID) of the Customs Force in a European country generates 3.5 terabytes of data for every case (Banerveld, 2014). This amount of data, which consists primarily of written evidence such as emails, requires investigation officers to review, analyze, and summarize the data in order to gather and finally act on the information, which will extend the amount of time that the potential criminal is walking freely.

Figure 2. Scharr, J. (2013). Example of lolita. The CID handles over 450 cases a year and a result needed to adapt to the oncoming threat with an offensive response to embrace the opportunity to innovate. They introduced NLP technology to help their department analyze and summarize the hundreds of terabytes of data that was coming into their department. The use of this technology has made police investigations more convenient and has even helped prevent crimes since officers were able to respond to inquiries faster with the relevant information gathered from digital traces. Following the results of the study into the CID’s NLP system by Banerveld and his team in 2016, it was concluded that de-coupling the department from a physical review process to an automated digital process, using the NLP system saved the department invaluable time in the following ways. By programming the NLP to target certain areas, the system was able to help optimize the investigation process by finding evidence in large datasets that could have been overlooked if being reviewed manually and did so in less time. These advantages allowed investigators in the departments to respond to more cases more quickly, saving the department money as well by limiting the number of time officers needed to spend reviewing large sets of case data.

A radical example of sustaining innovation can also be seen in the merging of two industries, security and chatbots. In 2013, a chatbot named Negobot was developed in Spain. The chatbot, also called “Virtual Lolita”, was designed to “pose as an emotionally vulnerable teenage girl and [tried] to trick online predators into giving away information that would help authorities track down pedophiles…and even lure them into a physical meeting to be detained by police” (Scharr, 2013), an example of a conversation with Negobot can be seen in Figure 2.

The creators of Negobot could sell this use of NLP to authorities on an international level to help them catch more predators before they have a chance to inflict damage, mentally or physically, to a real child. Through this example, you can see the evolution of the sustaining nature of chatbots. Originally, they were inferior and could only answer FAQ’s and now they have evolved to the point of successfully imitating a human. 5.1.3 Chatbots as a Virtual Assistant Chatbots were already mentioned in section 3.1. This example is about a chatbot that could also disrupt an industry. Chatbots were first used for the purposes mentioned in section 3.1: answering FAQ questions. They were, in the beginning, inferior to a human assistant, but they can now complete more complex tasks. (PWC, n,d).

Currently, there are chatbots that could help executives (Martin, 2018). These chatbots can make expense reporting, prepare documents, manage emails, make travel arrangements and schedule appointments. The advantages of a virtual assistant in comparison to a human one are outlined in figure 3. So, in the beginning, the chatbots, that could answer frequently asked questions, were not attractive for the most attractive customers, the higher management of a company, but now they are superior to a human assistant in a timescale of 7 years (PWC, n,d). These chatbots have the same digital enablers as the earlier mentioned chatbots: low marginal costs and ML.

Figure 3. PWC. (n.d.). The Characteristics of a Well - Designed Chatbot

5.2 Business strategies

The NLP innovation should be interesting for the most attractive customers. In this situation of innovation, a manager at an incumbent firm will most likely be more successful than new entrants with sustaining innovations like NLP, simply because they have a lot more resources to allocate on this technology. You can see this at companies like Apple, Google, and Amazon with their AI assistants like Siri and Alexa that are essentially paving the way in NLP. However, when new entrants start applying NLP into product or services that threaten to disrupt an industry managers need to decide which response will be preferable for their company. Three of these strategies suggested by Adner & Snow (2010) are explored below.

5.2.1 Bold Retreat

The market research field is an example of an industry that has been hi-jacked by NLP. Market research firms are based on a large team spending day, weeks, and maybe even months to collect and analyze information on behalf of their clients who were willing to wait for their insights. With the widespread introduction of NLP in handheld and in-home devices, now clients want surface level insights much faster and more easily available.

An incumbent organization that was faced with this transition away from the in-depth analysis is a company called Drive Research. They recognized the effects that NLP could have on their industry and decided to reformat the way they market their services to clients. They emphasize that while having a lot of data is helpful, knowing what steps to take next is the bigger battle, which they are able to help their clients with in order to improve their profits and customer service journey (Kuhn, 2017). Framing their services in this light provides potential and current customers with a sense of security that goes beyond what NLP can offer.

Other ways that this situation could be approached with a more defensive strategy where firms only try to embrace the use of NLP in their business practices and don’t offer anything beyond their current business plan. While this might help sustain them for a little while and get them ahead of other firms that are not willing to embrace the technology at all, this might help retain customers in the long run as NLP services become less expensive and more easily available to consumers and smaller firms as seen in the S-Curve model (Tuerstcher, 2018).

Or another example is how Google adapted their NLP model to include customer feedback in their Google Translate service. They take the technology of NLP a step further and make it customer facing using direct network externalities. This means that the more people that use the Translation service and provide feedback on the results they get, feed the network through machine learning, which in turn improves future results for more users (Puchala-Ladzinska, 2016).

5.2.2 Offensive Response

The introduction of NLP can be seen as a disruptive technology. Like we already mentioned, NLP creates a new way of working in the police industry and disrupts the criminal activity sector. Using these techniques, NLP has the ability to analyze suspicious activity online, even for specific targets and delivers all of the relevant information to the police in a summarized format.

This disruptive technology can be seen as an opportunity. Clark (2003) mentioned in his article that the net effect of the industry changed by disruption has been total market growth. ‘Disruption can be a powerful avenue for growth through new market discovery for incumbents as well as for upstarts.’ And like Christensen (2016) also explained, the disruptive opportunity can be seen in the context of a new value network, or autonomous business unit, separate from the parent company, which has the freedom to enact its own business model and pursue a disruptive opportunity.

In the article of Cearly, Burke, Searle, and Walker (2017), the top technology trends of 2018 are explained. They state that the way we interact with technology will undergo a radical transformation over the next five to ten years. For example, conversational platforms will provide more natural and immersive interactions with the digital world.

“A conversational platform provides a high-level design model and execution engine in which user and machine interactions occur. As the term "conversational" implies, these interfaces are implemented mainly in the user's spoken or a written natural language” (Cearly et al, 2017). Cearly et al. (2017) mentioned that over the next few years, conversational interfaces based on natural-language interfaces will become the main design goal for user interaction. This is applicable to some industries because in other industries like marketing or security natural language interfaces have already made a large impact. So for some companies, making use of NLP can be a huge opportunity.

5.2.3 Defensive Strategy

A defensive response to the looming threat of NLP that could be taken by a company can be seen in the translation industry. Online services such as Google translate offer free, convenient, and easy to use ways to translate words and phrases. This can be seen as a disruption to the translation industry since it is a cheaper and more convenient option for translation services, however, it is not always a truly accurate translation. Firms can exploit this weakness and take advantage by going to a higher market. For example, if a large organization such as Apple needs to create a version of their website into Dutch, a translation firm would be preferred, even though it would be more costly. Since Apple has a lot of financial resources, they would able to pay the premium price to ensure their website didn't have any translations errors.

6. References
Abu Shawar B., Atwell E. (2007) Chatbots: Are they Really Useful?, LDV-Forum Journal for Computational Linguistics and Language Technology, 22 (1), pp. 29-49.

Adner, R., & Snow, D. (2010). Old technology responses to new technology threats: demand heterogeneity and technology retreats. Industrial and Corporate Change, 19(5), 1655-1675.

Agerri, R., Bermudez, J., & Rigau, G. (2014). IXA pipeline: Efficient and Ready to Use Multilingual NLP tools. In LREC (Vol. 2014, pp. 3823-3828)

Arain, A. A., Manzoor, A., Brohi, K., Haseeb, K., Halepoto, I., & Korejo, I. (2018). Artificial Intelligence Mark-up Language Based Written and Spoken Academic Chatbots using Natural Language Processing. Sindh University Research Journal-SURJ (Science Series), 50(01), 153-158.

Are you sacrificing respondents and data quality on the altar of speed and price? (2018, May 22). Retrieved from https://www.p2sample.com/blog/are-you-sacrificing-respondents-and-data-quality-on-the-altar-of-speed-and-price/

Banerveld, M. V., Le-Khac, N., & Kechadi, M. (2014). Performance Evaluation of a Natural Language Processing Approach Applied in White Collar Crime Investigation. Future Data and Security Engineering Lecture Notes in Computer Science, 29-43. doi:10.1007/978-3-319-12778-1_3

Bates, M. (1995). Models of natural language understanding. Proceedings of the National Academy of Sciences, 92(22), 9977-9982.

Berends, H. (2018). Management of Digital Innovation. Nature of Digital Technologies [Powerpoint slides]. Retrieved from   https://canvas.vu.nl/courses/35851/pages/nature-of-digital-technologies?module_item_id=136761

Brownlee, J. (2017, September 25). Promise of Deep Learning for Natural LanguageProcessing. Retrieved from https://machinelearningmastery.com/promise-deep-learning-natural-language-processing/

Cavnar, W. B., & Trenkle, J. M. (1994). N-gram-based text categorization. Ann arbor mi, 48113(2), 161-175.

Cearly, D. W., Burke, B., Searle, S. & Walker, M. J. (2017). Top 10 Strategic Technology Trends for 2018. The Top, 10.

Che, W., Li, Z., & Liu, T. (2010, August). Ltp: A chinese language technology platform.In Proceedings of the 23rd International Conference on Computational Linguistics: Demonstrations (pp. 13-16). Association for Computational Linguistics.

Chopra, A., Prashar, A., & Sain, C. (2013). Natural language processing. International journal of technology enhancements and emerging engineering research, 1(4), 131-134.

Christensen, C. M., Altman, E. J., McDonald, R., & Palmer, J. (2016). Disruptive innovation: Intellectual history and future paths. Harvard Business School Working Paper 17-057.

Clark, G. (2003). The disruption opportunity. MIT Sloan Management Review, 44(4), 27.

Deloitte. (2018). Chatbots Point of View. Retrieved from https://www2.deloitte.com/content/dam/Deloitte/nl/Documents/deloitte-analytics/deloitte-nl-chatbots-moving-beyond-the-hype.pdf

Eryiğit, G. (2014). ITU Turkish NLP web service. In Proceedings of the Demonstrations at the 14th Conference of the European Chapter of the Association for Computational Linguistics (pp. 1-4).

Ferrucci, D., Brown, E., Chu-Carroll, J., Fan, J., Gondek, D., Kalyanpur, A. A., … Welty, C. (2010). THE AI BEHIND WATSON — THE TECHNICAL ARTICLE. AI Magazine. Retrieved from https://www.aaai.org/Magazine/Watson/watson.php

Gardner, M., Grus, J., Neumann, M., Tafjord, O., Dasigi, P., Liu, N., & Zettlemoyer, L. (2018). Allennlp: A deep semantic natural language processing platform. arXiv preprint arXiv:1803.07640.

Hagiu, A., & Wright, J. (2015). Multi-sided platforms. International Journal of Industrial Organization, 43, 162-174.

Helgesson Törnqvist, M., & Stålhandske, T. (2017). Unsupervised text clustering using survey answers.

Hellmann, S., Lehmann, J., Auer, S., & Brümmer, M. (2013, October). Integrating NLP using linked data. In International semantic web conference (pp. 98-113). Springer, Berlin,Heidelberg.

Hill, E., & Richter, V. (2016). Nina Virtual Assistant from Nuance Brings Human Touch to Swedbank Customer Service. Nuance [Burlington]. Retrieved from https://www.nuance.com/about-us/newsroom/press-releases/nuance-brings-human-touch-to-swedbank.html

How Many Products Does Amazon Carry? (n.d.). Retrieved May, 2016, from https://www.retailtouchpoints.com/resources/type/infographics/how-many-products-does-amazon-carry

Jain, A.K. (2009). Data clustering: 50 years beyond k-means. Pattern Recognition Letters, Volume 31, Issue 8, Pages 651-666.

James, D., Witten, G., Hastie, T. & Tibshirani, T(2013). An introduction to Statistical Learning. New York, Springer.

Jones, K. S. (1994). Natural language processing: a historical review. In Current issues in computational linguistics: in honour of Don Walker (pp. 3-16). Springer, Dordrecht.

Kar, R., & Haldar, R. (2016). Applying Chatbots to the Internet of Things: Opportunities and Architectural Elements. arXiv preprint arXiv:1611.03799.

Kiser, M (2016, August 11). Introduction to Natural Language Processing (NLP). Retrieved from: https://blog.algorithmia.com/introduction-natural-language-processing-nlp/

Knoll, L. (2018, March 16). Developing The Connected World Of 2018 And Beyond. Retrieved from https://www.forbes.com/sites/forbestechcouncil/2018/03/16/developing-the-connected-world-of-2018-and-beyond/

Kuhn, G. (2017, December 01). Why You Need a Closed Loops for Customer Experience (CX) Surveys. Retrieved from https://www.driveresearch.com/single-post/2017/06/07/Why-You-Need-a-Closed-Loop-Process-for-Customer-Experience-CX-Surveys

Lexalytics. (2018). Machine Learning vs. Natural Language Processing - Lexalytics. Retrieved from https://www.lexalytics.com/lexablog/machine-learning-vs-natural-language-processing-part-1

Liamputtong, P. (2011). Focus Group Methodology: Introduction and History. Focus Group Methodology: Principles and Practice. Retrieved from https://www.sagepub.com/sites/default/files/upm-binaries/39360_978_1_84787_909_7.pdf

Malmasi, S., Sandor, N. L., Hosomura, N., Goldberg, M., Skentzos, S., & Turchin, A. (2017). Canary: An NLP platform for clinicians and researchers. Applied clinical informatics, 8(02), 447-453.

Martin, L. M. (2018). Evaluating The Top 22 Virtual Assistant Companies of 2018. Retrieved from https://biz30.timedoctor.com/virtual-assistant-companies/

Mitomo, H. (2017). Data Network Effects: Implications for Data Business.

Puchala-Ladzinska, K. (2016). Zeszytynaukowe Uniwersytetu Rzeszowskiego. Machine Translation: A Threat Or An Opportunity For Human Translators? 89-98. doi: 10.15584/sar.2016.13.9

PWC. (n.d.). Chatbot: The intelligent banking assistant. Retrieved from https://www.pwc.in/consulting/financial-services/fintech/fintech-insights/chatbot-the-intelligent-banking-assistant.html

Russom, P. (2007). BI search and text analytics. TDWI Best Practices Report, 9-11.

Scharr, J. (2013, July 13). Controversial 'Lolita' chatbot catches online predators. Retrieved from https://www.nbcnews.com/technolog/controversial-lolita-chatbot-catches-online-predators-6C10622694

Social media trends 2017: Everything you need to know - Just Ask Gemalto EN. (2017, January 30). Retrieved from https://www.justaskgemalto.com/en/social-media-trends-2017-everything-need-know/

Srihari, R. (2015, March 16). How can NLP technology be used for marketing? Retrieved from https://econsultancy.com/how-can-nlp-technology-be-used-for-marketing/

Steinley, D. (2003). Local optima in K-means clustering: what you don't know may hurt you. Psychological methods, 8(3), 294.

Streiter, O., Scannell, K. P., & Stuflesser, M. (2006). Implementing NLP projects for noncentral languages: instructions for funding bodies, strategies for developers. Machine Translation, 20(4), 267-289.

Tuertscher, P. (2018). Lecture 3 Industry Level Dynamics [PDF]. Retrieved from https://canvas.vu.nl/courses/35851/files/folder/Lecture%20slides?preview=820205.

Wang, A., Singh, A., Michael, J., Hill, F., Levy, O., & Bowman, S. R. (2018). GLUE: A Multi- Task Benchmark and Analysis Platform for Natural Language Understanding. arXiv preprint arXiv:1804.07461.

Wasankhasit, N., & Nilsson, E. (2018). Look before you leap: A take on customer support agents' pre-implementation attitudes on chatbots.