Generative vs. Retrieval Chatbots for Higher Education

Generative vs. Retrieval Chatbots for Higher Education

?Let's delve deeper into NLP AI chatbots for student communications in colleges and explore the difference between generative and retrieval-based models.

?NLP (Natural Language Processing) AI Chatbots:

?Definition: Chatbots like the Comm100 AI NLP Chatbot utilize natural language processing and can understand, interpret, and produce human language in a way that's meaningful. They don't rely on pre-defined rules or exact keywords; instead, they process the language to discern intent and context. This category includes both NLP AI chatbots and Generative AI chatbots.

?Application in Colleges: All of these chatbots can be used across the student lifecycle for recruitment/admissions, student advising/services, IT support, alumni/advancement and more.

?Generative vs. Retrieval-based Models:?

?Generative AI Bot:

?Definition: Generative AI chatbots generate new responses from scratch based on their training data. The bots leverage Large Language Models (LLMs). The most popular LLM in 2023 is ChatGPT however there are a number of different LLMs and many approaches and these systems will evolve rapidly in the coming months. Chatbots use LLMs to understand user inputs, to generate responses, to ensure conversational flow across multiple exchanges, to create content, to adapt/learn and much more.

?Pros:

1. More Flexible: Can potentially respond to a vast array of questions, even if they've never encountered them before.
2. More Conversational: They can provide customized answers based on the user's query.
3. Faster Training: A Generative bot can be given access to a website, content like FAQs, etc. and be trained in only a few hours vs a few months for a retrieval bot.

?Cons:

1. Unpredictable: Can occasionally produce irrelevant or incorrect answers. Chat GPT 3.5 for example is about 80% accurate vs GPT 4 which is about 90% accurate. That means however that many answers will be incomplete or incorrect.
2. Resource Intensive: Generally, need more computational power.
3. Variable Costs: Any calls to ChatGPT incur a cost. This cost varies between Chat GPT 3.5 ($0.002 per 1K tokens) to ($0.03-$0.12 per 1K tokens)
4. Inability to Integrate: They cannot retrieve data from (for example) major systems like the SIS (Ellucian, Oracle, Workday, etc.), CRM (Slate, Salesforce, Oracle, etc.), LMS (Instructure, Moodle, D2L, etc.) so it can handle a lower number of queries.

?Retrieval-Based Bot:

?Definition: These chatbots select the most appropriate response based on their learning data. Retrieval based chatbots use predefined responses, input processing, intent recognition, response selection, context management and fallback answers among other things. They also leverage systems like Natural Language Understanding (NLU) to create human like responses.

?Pros:

1. Predictable: 100% consistent responses, as they select from a predetermined set of data and answers.
2. Less Risky: No likelihood of generating inappropriate or nonsensical answers.
3. Integrated: Can connect to SIS, CRM and LMS systems as well as many others out-of-the-box so it can handle a higher number of queries.
4. Fixed Cost: No variable fees like tokens to consider.

?Cons:

1. Limited Flexibility: Restricted to their predefined responses.
2. Potential for Repetition: May appear less "intelligent" if repeatedly providing the same responses to slightly varied questions.
3. Longer Training Time: The bot requires data, definitions and intents to be able to answer and provide data, links, videos, etc.

?Applications for Student Communications:

• Admission Queries: NLP chatbots can answer questions about admission requirements, financial aid, application deadlines, tuition fees, enrollment, onboarding, orientation, and more.
• Advising/Academic Information: Provide details about student services, counseling, career advising, course enrollment, grade inquiries, faculty, academic schedules, course syllabi and more.
• IT Support: Help students navigate IT Issues like passwords, account creation, support tickets, etc.
• Alumni/Advancement: Handle queries related to events, sponsorships, day of giving donations and more.
• Feedback Collection: Solicit feedback on courses, instructors, or facilities.

?Implementation Considerations:

• Training Data: For a chatbot to be effective, it requires quality training data and/or intents relevant to the college and its students.
• Systems Integrations: Decisions on what data needs to be integrated with a retrieval bot and existing college databases and systems (ex: LMS, SIS and CRM).
• Updates & Maintenance: All chatbots need regular updates, especially retrieval-based ones, to add new information and perfect/train more accurate responses.
• User Experience: A friendly and intuitive user interface and even a cute name can make the difference in the chatbot's adoption rate and overall effectiveness.