Development And Application Of Chatbots In Finance Using AI And NLP, An Essay.

To a large extent, Artificial Intelligence (ΑΙ) fuses daily life with the development and analysis of intelligent software and hardware, often identified as intelligent agents. The resulting intelligent agents are postulated to carry out various tasks including but not limited to labor work and sophisticated operations such as learning and data analysis. An example of one of the most popular intelligent agents is the chatbot which is a denotation of intelligent Human-Computer Interaction (HCI) (Bansal & Khan, 2018). Primarily, chatbots can emulate among other processes, human conversation as well as entertain users making them applicable in fields like education, information retrieval, business, finance, healthcare, and e-commerce (Abu & Atwell, 2010). According to Adamopoulou & Moussiades (2020), a chatbot is a computer program that generates responses like a smart entity when engaged through text or over voice and understands one or more human languages through the use of Natural Language Processing (NLP).

The popularity of chatbots has been driven by the need to engage users over live chat interfaces which essentially facilitates the provision of real-time customer services across various fields. In practice, users utilize the underlying chat services to gain and access information such as product details or obtain assistance such as solving technical issues in products etcetera (Adam, et al., 2021). Besides, since knowledge in the use of a given chatbot can be transferred to the application of other chatbots, as well as there being limited data requirements, the development of chatbots allows for faster scaling and widespread usage (Klopfenstein, et al., 2017). (Klopfenstein, et al., 2017) argues that, communication reliability, quick and straightforward development iterations, absence of version fragmentation, as well limited design requirements for the agent interface are among the main advantages for developers.

Mission Statement

In this study, we seek to develop a knowledge base chatbot prototype that depending on user requirements provides general information on various aspects of foreign exchange markets. As such, the target audience of the proposed prototype is players in the finance industry. The chatbot will provide general information on issues such as what foreign exchange entails, the history of the foreign exchange market, the market’s structure, parties involved, importance, etcetera.

Overview of Tutorial Improvement

Some of the areas that we have improved on the original tutorial include, adding a scoring method i.e., cosine similarity for evaluating the quality of responses to be given to users, adding user interaction interfaces such as collecting user information such as username, developing a relatively advanced database with questions and expected answers, optimized the chatbot to respond to both keywords and phrases that are input by users. For instance, if a user passes a phrase like what about forex the expected response will be the same as that given to a user who passes the keyword forex.

Literature Review

Moreover, by defining an option of the similarity metric to be used which primarily would be either cosine, cityblock, euclidean, l1, l2, and Manhattan it enables the developers to twerk the chatbot configurations to improve the quality of the responses. 

History of Chatbots

Chatbots have over time, symbolized a shift in how individuals interact with data and services online (Brandtzaeg & Følstad, 2017). Following the proposal of Alan Turing in 1950 that “Machines can think”, the concept of a chatbot was first popularized (Adamopoulou & Moussiades, 2020). Sixteen years later, Eliza, the first publicly known chatbot which used a pattern matching approach (Brandtzaeg & Følstad, 2017) and template-based response method, was developed to function as a psychotherapist providing the user utterances in the form of a question. While Eliza was not comprehensive, it paved way for the development of other chatbots such as Parry in 1972 (Sean, 2020) by Kenneth Colby and Alice in 1995 (Wallace, 2009). The Alice chatbot which won the Turing awards in 2000, 2001, and 2004 was dependent on a simple pattern-matching algorithm integrated with an underlying intelligence built on the Artificial Intelligence Markup Language (AIML) (Gracas, et al., 2013).

Research interest in chatbots as shown in figure 1 below, based on a Scopus search by  (Adamopoulou & Moussiades, 2020), increased significantly in the years following 2016.

Figure 1: Number of documents that explore chatbots between 2000 and 2019, source: https://www.ncbi.nlm.nih.gov

In 2001, the development of the SmarterChild chatbot was a pivotal moment for the evolution of chatbot technology after it was made available on several platforms such as America Online (AOL) and Microsoft (MSN) to facilitate the retrieval of information, marking an important developmental point in machine learning (ML) and human-computer interaction since information could now be accessed based on discussions by chatbots as opposed to human agents (Molnár & Zoltán, 2018).

McTear (2021) argues that the emergence of chatbots is due to the need of addressing various features of conversational computer systems in industrial applications with the recent research interest being a direct response to the “…uptake of so-called virtual assistants by big tech companies” (Dale, 2016), in particular, the adoption of Siri as a constituent of the Apple operating system in 2011, Amazon's proposal of Alexa as from 2014 as well as the conversational turn of both Facebook, Microsoft, and Google since 2016 (Dale, 2016).

Usage of Chatbots

(Brandtzaeg & Følstad, 2017) conduct a study on why people use chatbots. According to Brandtzaeg & Følstad (2017), chatbots are used in several ways including in customer service, social and emotional support, information, and entertainment, besides connecting users to other individuals or machines. This applicability of chatbots has prompted the adaption of chatbots across different fields and applications, ranging from education to finance, enveloping healthcare and entertainment (Caldarini, et al., 2022).

Application Of Chatbots

The upward trend observed in 2016 (see figure 1) was brought about by the evolution of artificial intelligence which changed the way individuals communicate with manufacturers. Moreover, the development of chatbots for social media allowed users to carry out some daily tasks using their social media messengers. By the end of 2016, there were approximately 34,000 bots in usage across fields such as “…Marketing, Supporting Systems, Health Care, Entertainment, Education, and Cultural Heritage” (Adamopoulou & Moussiades, 2020).

Chatbots should be implemented anthropomorphically (that is, adopt a human-like form) and generate a sense of social presence (Zhang, et al., 2012). This is achieved through adopting features of human-human communication. In their study (Adam, et al., 2021) argue that most chatbot designs are generally focused on anthropomorphic design cues and how they will impact human behavior while taking into consideration the perceptions and adoption of the agents.

Pattern Matching

During implementation, rule-based agents match the user input to a specified rule pattern and return a predefined response from a list of responses by using Pattern Matching algorithms (Marietto, et al., 2013). Both Eliza and Alice were implemented using pattern matching. The snip below shows an example of a pattern-matching algorithm.

Parsing

Parsing is used to examine the input provided by users by using natural language processing or computer language. The technique transforms the input into a list of lexicons i.e., lexical parsing to establish the grammatical distribution of the text before generating appropriate responses (Devakunchari, et al., 2019).

Other approaches for building chatbots include ChatScript, SQL, and Relational Database for building knowledge chatbots, Language tricks which uses sentences and other language structures to expand the knowledge of chatbots, and which uses NLP to produce responses by establishing relationships between various documents (Devakunchari, et al., 2019). 

Our goal was to develop a foreign exchange market knowledge base chatbot whose purpose is to interact with users and provide responses to the user in question based on their queries related to concepts of foreign exchange. To achieve this, we used a natural language approach that combines some features from parsing and latent semantic analysis. The whole process was interesting given the necessity of selecting an appropriate source of information, processing the information, adding the information into two documents for relationship matching, and developing the interaction component of the chatbot.

Source of Information

I realized that there isn’t a ready-to-use source of information which posed a challenge of how to collect the data. Most of the information sources that I came across focused on a specific aspect of the foreign exchange market such as the characteristics, functioning, risks, etcetera. Second, having found a relatively comprehensive source of information, the second challenge was to structure the database such that it will not cause issues during data processing which prompted me to store the data in a text format.

Processing Information

During prototype implementation, I noted that developing the chatbot requires about three components i.e., the introduction where the chatbot interacts with a user for the first time, collects information such as their name, offers greetings and basic directions, accepts user input, and generating responses to the user input depending on the most relevant response from among all the possible responses. Therefore, this necessitated generating and defining all possible keywords and phrases that are related to the foreign exchange market which was a challenge since I had to read through the original information to establish all the relevant text manually before adding it to the databases for processing and automation.

Besides, I realized that it is important to have a valid and relatively well-performing document comparison metric. In this case, I experimented with the Cosine, Euclidean, and Manhattan distance metrics.

Different text processing approaches tended to lead to varying responses. I assumed this is because approaches such as vectorization, lemmatization, stemming, etcetera lead to different outputs which are processed differently when generating responses.

Interaction Component

Lastly, I noted that user-friendliness is an important factor, and to make the chatbot convenient for user interaction, an exit criterion needed to be defined. For instance, we allowed the user to terminate the chat session by typing inputs such as Bye or Quit. 

In this section, we discuss an evaluation of the final chatbot prototype including the decision-making process, test plans, testing results, and alterations suggested by test/evaluation results.

Instances that required decision-making included database selection, question and answer definition as well as response generation.

Wikipedia is one of the largest knowledge databases on the internet which, unlike scholarly articles, tends to provide a general overview of a given concept.

Table 1: Overview of the table of components for foreign exchange

Table 1 above provides an overview of the topics explored in the Wikipedia page of Foreign Exchange Market which can be accessed through the Foreign exchange market link. This aligns with the original objective in which we sought to provide a user with general information regarding the foreign exchange market. However, during the process of testing, we noted that the definition of risk provided in the Wikipedia page had a document similarity with the speculation component. We therefore decided to collect information regarding the market’s risk from https://corporatefinanceinstitute.com/resources/knowledge/finance/foreign-exchange-risk/ which improved the performance of the prototype.

As noted earlier, database section was split into two i.e., the possible questions and response databases.

There are many possible questions that a user could supply to the chatbot and since different individuals tend to require different information, it is not feasible to generate all possible questions. As such, we used the titles provided in the database to generate a list of possible questions. Using these questions, the similarity between a new question and those in the database was computed using a pairwise similarity metric and linked to the response that has the best similarity. An example of the information source for the question What is the market size? is given in the snip below.

The response to the questions that were generated corresponded to the content below each of the titles from which the questions were generated. For instance, the response to the What is the market size? Was selected to be as shown in the snip below.

Other decisions such as the username with which to refer to the user and when to terminate the interaction depended on the user input. For instance, if a user, defined their name as John Doe, the program will explicitly refer to them as John Doe until the program is terminated and a new username is supplied. This was achieved through the use of conditional statements such as if-else statements. An example of a user-supplied name is shown in the following snip.

The performance of the chatbot was evaluated on how well it generated responses. Up to three document similarity metrics were used and the resulting responses were explored. This involved using the same 5 questions per metric and evaluating the quality of the responses. The number of times that the chatbot generated the correct responses were computed and used to determine the overall accuracy of the chatbot per metric. During our evaluation, we noted that the cosine similarity metric had 8 correct responses out of 10 compared to that of the Euclidean and Manhattan metrics each of which generated 6 correct responses.

 Prior to generating responses, the optimal response was that which had the closest similarity with responses to questions that are similar to those supplied by the user. To determine the most relevant response, a threshold of a cosine similarity of less than 30 was selected. In this case, if a response to a question had a similarity score greater than 30, it was considered not relevant and a message was displayed. Otherwise, the response with the maximum cosine similarity was selected and returned to the user.

To evaluate the effectiveness of the chatbot. We considered the use of both a keyword and phrase relating to what the foreign exchange market (i.e., Forex market) is. The original question in the question database is: What is forex while the keyword is forex. The following snip provides an overview of the responses generated by the chatbot with different variants of the question. From the snip below, the chatbot generates 6 correct responses out of 7 questions. We however note that for the incorrect response, there is no inclusion of the keyword Forex which might have led to the retrieval of the provided information.  

Conclusion

Most organizations that are technology-oriented are increasingly adopting the use of AI technology to improve the quality of services offered to their clients. With an ever-growing competition, it is only necessary that the respective organizations integrate new methods like intelligent agents such as the chatbot to automate various processes including but not limited to collecting customer orders, providing customer support, information, and entertainment. The objective of the current study was to implement an information-retrieval chatbot whose role was to provide users with various aspects of the foreign exchange market. To this end, we used a natural language processing approach to preprocess the data and genearte patterns that could allow Latent Semantic Analysis for the comparison of information to generate responses.

Since based on past chatbot projects, we noted that sample response data can be stored in json files. We noted that collecting data and processing it as json is bit tedious compared to collecting data and storing in different two seprate text files. This allows for faster scaling of the databases compared to the sytax based data storage requirement by json.

How has academic theory or practical advice from the sources you have used informed and improved your chatbot?

There are different approaches towards generating meaning from text in semantic analysis. Following practical advice, we learnt that while lemmatization improves the quality of data, vectorization I useful in converting input data from its raw format (that is, text) into vectors of real numbers that machine learning algorithms support. In our case, using vectorization helped the cosine similarity model compute the similarity between potential responses to a question based on the content of the new user query.

Were there any aspects of developing your chatbot that involved you working in new ways or ways that you had not anticipated?

The chatbot required that we have two separate databases i.e., questions and expected responses. We had not anticipated that the chatbot could generate insights and corresponding responses from two different databases. Besides, we tried to experiment with different similarity metrics and data processing techniques for the chatbot.

While using the count and Term Frequency — Inverse Document Frequency (tfidf) vectorizers, we noted that the model generally returned more relevant responses when we used the count vectorizer which was unexpected since we had assumed the tfidf vectorizer generates better insights from the text data.

An Honest Appraisal of Your Performance and The Produced Chatbot

The chatbot performs relatively well when supplied with input that has almost similar composition as the questions used to build the chatbot. However, since the approach is based on supervised learning, the chatbot when supplied with completely new input with a different composition, it sometimes generates irrelevant output. In such cases, we would ordinarily expect that it will terminate but the bot terminates on few occasions when the similarity score is very low. The chatbot does not preprocess data to remove noise such as punctuations and numbers which might influence the performance of the chatbot especially if it is adapted for machine learning.  

References

Abu, S. B. & Atwell, E., 2010. Chatbots: are they really useful?. J. Lang. Technol. Comput. Linguist., Volume 22, pp. 29-49.

Adam, M., Wessel, M. & Benlian, A., 2021. AI-based chatbots in customer service and their effects on user compliance. Electron Markets, 31(2021), pp. 427-445.

Adamopoulou, E. & Moussiades, L., 2020. An Overview of Chatbot Technology. Neos Marmaras, Greece, AIAI, pp. 373-383.

Bansal, H. & Khan, R., 2018. A review paper on human computer interaction. Int. J. Adv. Res. Comput. Sci. Softw. Eng., 8(2018), p. 53.

Brandtzaeg, P. B. & Følstad, A., 2017. Why people use chatbots. In: e. a. Kompatsiaris I, ed. Internet Science. Cham: Springer.

Caldarini, G., Jaf, S. & McGarry, K., 2022. A Literature Survey of Recent Advances in Chatbots. Information, 13(41).

Dale, R., 2016. The return of the chatbots. Nat Lang Eng, 22(5), pp. 811-817.

Devakunchari, R., Agarwal, R. & Agarwa, E., 2019. A Survey of Chatbot Design Techniques. International Journal of Engineering and Advanced Technology (IJEAT), 8(2S2), pp. 35-39.

Gracas, B. M. M. d. et al., 2013. Artificial Intelligence Markup Language: A Brief Tutorial. International Journal of Computer Science & Engineering Survey, 4(3), p. 1–20.

Klopfenstein, L., Delpriori, S., Malatini, S. & Bogliolo, A., 2017. The rise of bots: a survey of conversational interfaces, patterns, and paradigms. s.l., Association for Computing Machinery, pp. 555-565.

Marietto, M. et al., 2013. Artificial intelligence markup language: A brief tutorial. International Journal of Computer Science and Engineering Survey, 04(2013).

McTear, M., 2021. Conversational AI: Dialogue Systems, Conversational Agents, and Chatbots. Williston: Morgan & Claypool.

Molnár, G. & Zoltán, S., 2018. The role of chatbots in formal education. s.l., IEEE.

Sean, B., 2020. Chatbot History: The Parry Chatbot, s.l.: Yakbots.

Wallace, R., 2009. The anatomy of A.L.I.C.E. In: R. Epstein, G. Roberts & G. Beber, eds. Parsing the Turing Test: Philosophical and Methodological Issues in the Quest for the Thinking Computer. Cham: Springer, pp. 181-210.

Zhang, H. et al., 2012. Mood and social presence on consumer purchase behaviour in C2C E-commerce in Chinese culture.. Electronic Markets, 22(3), pp. 143-154.