Large-Language Models Are... Models

Large-language models are just that... models. They are representations of a reality, and us librarians have been creating models of a similar ilk for centuries...

Models abound. A person stands in the middle of room while artists draw the person. In this case the person is a model of the human form. Are the drawings of the person "correct"? I assert, "No, because the given question makes no sense; 'correct' is not a valid way characterize a model." Instead, words/phrases like "accurate", "useful", "expressive", or "technically competent" are more apropos.

Here in Library Land we have been modeling text for centuries. Examples include acquisitions lists, shelf lists, bibliographies (in this form, that form, and the other form), indexes, abstracts, etc. More modern models of text include relational databases, sets of MARC records, library catalogs and discovery systems, or Zotero libraries. In all of these cases there is a collection of texts and us librarians have modeled it in a number ways. None of the models are "correct", nor do they embody any values "good", "bad", "right", or "wrong". Instead, they are tools intended to be used in particular ways.

Large-language models are very similar to bibliographies, indexes, abstracts, etc. They are representations of collections of text. They are not "right" nor "wrong", "good" nor "bad". Instead, they are tools and as such they can be used in the manner intended or not.

I employ many different modeling techniques in my close and distant reading: counts & tabulations of extracted features (ngrams, named-entities, parts-of-speech, keywords), clustering (topic modeling, KMean, etc.), graphing networks, full-text indexing, semantic indexing, concordancing, etc.

And now I have added a new tool to my toolbox, large-language models. Large-language models do two things which are very helpful. One, they can summarize texts very well. Second, given a set of sentences and a question, large-language models can address the questions very well. I assert the result, while not necessarily true, is plausible. More over, both of these examples "save the time of the reader", which is a goal of library practice.

The acceptance of new technologies always comes with trepidation. Just think how the librarians who literally wrote library cards must have screamed and hollered with the invention of typewriters. Just think how librarians wrung their hands with the implementation of relational databases and/or the full text indexes where zero controlled vocabulary terms were required.

Large-language models are simply an evolution in the process of modeling text. They are well worth our time and effort to learn how to integrate them in to library practice.