FEATURE
Virtual Readers’ Advisory Using bANTerbot Code
by Kelsey Brown and Danielle Kane
This innocent question—“I really like the Game of Thrones books. Do you have another series like that?”—jump-started a multiyear project at the University of California–Irvine (UCI) campus library (UCI libraries), culminating in adding book recommending code to the library’s chatbot. UCI is a public university in southern California, notable for being a Hispanic-serving institution with a significant international student population. UCI libraries also functions as a public library for some patrons and has responded to this usage by subscribing to leisure reading titles via OverDrive. Reflecting on experiences of providing reference at the desk and via chat, a pattern emerged. Patrons were approaching the desk with readers’ advisory requests that were difficult to fill on-the-fly. Conversations with the student assistants staffing the desk confirmed that requests such as “I’m trying to read more Black authors—who do you recommend?” were not possible to answer with simple subject or keyword searches. The consideration about using bANTerbot code began in fall 2019 as a class project for Kelsey Brown’s M.L.I.S. course on emerging technologies. She wanted to leverage an existing library technology, ANTswers, to address patrons’ desire for readers’ advisory and ease the workload for her student assistants and colleagues. After developing desired outcomes, Brown met with Danielle Kane. This initial meeting focused on determining the feasibility of the idea within the scope of the class project. Kane modified the existing ANTswers code, tested it with a sample book list, and determined the project was possible to attempt on a larger scale.
Adding bANTerbot Code to ANTswers
Libraries have been slow to adopt AI. While some have implemented chatbots or are using machine learning for analysis projects, it has not reached full adoption. There was even less research or information on creating a library chatbot back in 2013 when the ANTswers project was first started. ANTswers provides 24/7 unrestricted access to answer simple and directional questions about library resources and services. It is an AIML chatbot using Program-O, which is an open source AIML interpreter written in PHP. ANTswers logs are monitored to continue to improve and refine the service. From March 2014 to December 2022, ANTswers has been asked 16,678 questions, of which 10,636 were library related. No personal information beyond the chat logs is kept by the system unless provided by the chatter. The chatbot was created by a librarian with reference, instruction, and collections experience. ANTswers uses natural language processing and pattern matching to return a response that is the best match. Currently, ANTswers is not set up to learn from library patrons but is instead modified by the botmaster.
The bANTerbot code is a subset of code belonging to the ANTswers chatbot. The bANTerbot code consists of 29 AIML files organized by subject along with a pickup line file, which contains the sentences that library patrons can use to ask for book recommendations. While the bANTerbot code could be a fully functional standalone bot, it was decided to integrate the code into the existing ANTswers chatbot. This required careful consideration of the pickup lines for book recommendations and how the language needed to be distinct from the language used as pickups for book searches.
Project Workflow
As the reference department student supervisor, Brown organized a team of five student assistants trained to serve on the desk who were skilled in catalog searching, familiar with UCI patrons, and confident in locating information outside of their area of expertise. She then divided the bANTerbot project into two distinct phases. The first was a spreadsheet of popular books matched to recommendations for similar titles. These recommendations were primarily young adult titles—a reflection of the most common requests. While the overall goal of adding the bANTerbot code was to recommend items available in the collection, this specific section was broadened to name books that patrons could request through interlibrary loan. This “If you like” list was fully curated by a student assistant with a passion for young adult literature and the expertise to search for titles and assess their similarity. The result was a list of 16 popular titles with a minimum of three alternative recommendations for each.
The second phase was a spreadsheet of titles for requests related to genre and author demographics. Locating and adding titles needed to be a standardized task suitable for student assistants. Brown established the initial genre tags, created a controlled vocabulary of genre tags and appropriate synonyms, and approved suggestions based on relevant titles and anticipated usage. Within the genre or tag list, there were two distinct workflows. If a genre was easily discoverable through a subject search, students would review the list of results, pull out titles that appeared interesting to a leisure-reading audience, and search for additional information to fill in tags related to author demographics before filling in the spreadsheet. As noted through reference desk and chat interactions, tags were not discoverable via a simple subject or keyword search. A few examples include books with movie adaptations and books by an author of color. In these cases, students searched book recommendation websites for appropriate titles and then searched the title in the UCI libraries catalog and OverDrive collection. If the review process required a judgment call that students were uncomfortable making, they flagged the title for review by their supervisor. The added bANTerbot module launched after 32 titles were coded.
Keeping Recommendations Up-to-Date
The original 450 titles in the Excel spreadsheet were instrumental in creating the beta code for the bANTerbot module quickly and efficiently. The students spent approximately 1 year adding titles to the Excel spreadsheet, along with their other work. The layout/design of the spreadsheet made it an ideal project to work on at the end of a shift and while waiting for patrons at a service desk. When COVID occurred in March 2020, finding titles for the recommendation bot was an easy task for students transitioning to an online, asynchronous work environment. The project was also supported by a library school intern who worked on it from January to June 2020. The intern developed a set of initial code files that could be tested against the ANTswers code and that were the basis for the fully completed bANTerbot code.
Once the bANTerbot code was integrated into ANTswers and was working correctly, a research guide was created to highlight how library patrons can ask for book recommendations. The bANTerbot code updates were based on both the recommendations library patrons asked for that might not have been coded and also the discretion of the botmaster. New code was added for UCI authors and tabletop games, and a concentrated effort was placed on increasing the number of diversity, equity, inclusion, and accessibility (DEIA) titles.
Moving from the original project to maintaining and updating the code, the Excel spreadsheet has been retired, and titles are now added directly to the AIML files. New titles are found through popular book lists, lists of award winners, and book collections made by librarians. New categories are added based on requested book recommendations by the campus community. Some requests are not coded due to the academic library not having that type of collection (for example, Christian fiction recommendations would be more appropriate for a public than an academic library). In these cases, code is added that no book recommendations are available for that topic.
Replicating Our Success
The bANTerbot code can be used in any system that accepts AIML code. The titles and URLs can be updated to work with other libraries’ OPACs and systems. If employing chatbot software that does not accept AIML code, the spreadsheet can be used to adapt the book recommendations. The bANTerbot code can be downloaded and used on its own or with other chatbot code. To successfully utilize both the spreadsheet and the bANTerbot code, begin by reviewing the types of book recommendations requested and use the spreadsheet to develop a title list, including URLs. Prior to coding, think about the chatbot back end to ensure the correct programming language is used and consider if it will be a full chatbot or a chatbot solely focused on book recommendations.
Chatbot Limitations
Through the planning and implementation phases, there were several limitations to consider. First, the current back end of ANTswers is at its end-of-life and is no longer updated. Program-O has been replaced with the Lemur engine, another AIML chatbot back end. At the time of this writing, ANTswers has not been updated to work with a new back end. Second, ANTswers and bANTerbot are not set up to learn on their own or from the patrons using it and require a botmaster to review and update them. Although this takes more staff time, it also removes the worry that patrons may teach the chatbot inappropriate responses.
During the initial book list creation, granting students a high level of autonomy resulted in recommendations skewed to their interests or niche book lists they found online. This led to interesting categories in which book recommendations have yet to be requested. Also, as the title and tag list grew, students needed to revisit previously coded titles to apply new tags. Another limitation of the bANTerbot code is that it can only pull recommendations from a single category; if patrons add multiple keywords, then a book recommendation cannot be supplied.
Marketing
When rolling out a unique technology such as bANTerbot code, targeted marketing is key to reach users. Unfortunately, COVID-19 disrupted initial marketing opportunities for bANTerbot. Original plans for outreach included creating fun swag, attaching bANTerbot to information-literacy programming, and piquing student interest via cryptic marketing across campus.
Undergraduate students love kitschy swag, especially buttons. Marketing efforts for the bANTerbot code take advantage of the many depictions of the university’s mascot, Peter the Anteater, to highlight the fun experience of the chatbot. An image of Peter saying, “Looking for your next beach read? Chat with the interface to explore books available through UCI libraries.” can be featured on buttons, stickers, bookmarks, and informational postcards. These materials can be distributed during tabling events and workshops in conjunction with a live demonstration and used as passive marketing at the reference and circulation desks. A recent series of UCI libraries videos featuring a puppet mascot using library resources and services has been incredibly popular among students. A short video of the mascot engaging with the chatbot would deliver a demonstration in an engaging package.
One advantage of using bANTerbot code is that it is easily incorporated into a LibWizard tutorial or quiz, making it an ideal question for a library scavenger hunt. These questions can be open-ended and low-stakes, emphasizing fun and discoverability by offering students several genres to explore. Depending on intended usage, questions could describe reading a catalog record, saving a record to a library account, and locating permalinks.
In addition to traditional marketing methods, there is also an opportunity to connect directly with students across campus by posting fliers and stickers with the QR code and a caption about finding a new favorite book. This could help students connect the interaction to leisure pursuits, rather than continuously associating it with the library as an academic space.
In Conclusion
With the addition of the bANTerbot code to ANTswers, we have seen an increase in patrons requesting book recommendations. A guide was created with information about requesting recommendations and the categories available. Utilizing student assistants to create the initial book recommendation list was incredibly helpful, especially early in the pandemic when we needed online work for the students. After the book recommendation support was localized to one person, the process was streamlined by adding the code directly to the appropriate file.
Currently, there are 1,711 individual monograph titles across more than 68 genres. Recommendations have been added for 144 books with movie adaptations and 137 tabletop games. With the increasing popularity of AI, libraries and academia will need to decide on an approach to utilizing it in the future. This will require evaluating AI back-end systems and working with library IT to implement a solution, server space, and programming support.
|
Kelsey Brown [L] is a student success librarian and liaison to the women’s, gender, and sexuality studies program at the University of Connecticut. Brown’s work is student-centered, focusing on information-literacy programming, outreach events, research support, and collaborating with campus partners. Previously, she was a library assistant at the University of California–Irvine’s libraries, where she taught workshops, provided in-person and online reference assistance, created student assistant training, and curated displays.