Digital Humanities, with David Bamman

Guest: David Bamman
Hosts: Waleed Ammar, Matt Gardner

In this episode, we invite David Bamman to give an overview of computational humanities. We discuss examples of questions studied in computational humanities (e.g., characterizing fictionality, assessing novelty, measuring the attention given to male vs. female characters in the literature). We talk about the role NLP plays in addressing these questions and how the accuracy and biases of NLP models can influence the results. We also discuss understudied NLP tasks which can help us answer more questions in this domain such as literary scene coreference resolution and constructing a map of literature geography. David Bamman's homepage: http://people.ischool.berkeley.edu/~dbamman/ LitBank dataset: https://github.com/dbamman/litbank

Matt Gardner
00:00

Hello and welcome to the NLP highlights podcast where we talk about interesting work in natural language processing.

Waleed Ammar
00:06

This is Matt Gardner and Waleed Ammar, we are research scientists at the Alan Institute for artificial intelligence.

Matt Gardner
00:12

Okay. Today our guest is David Bamman, who is an assistant professor at the information school at UC Berkeley. I met David when we both started at the same time at Carnegie Mellon University many years ago, I guess not that many years ago, but it feels like it was a long time.

David Bamman
00:27

It was.

Matt Gardner
00:28

David, welcome to the program.

David Bamman
00:29

Thank you guys. Thank you for having me.

Matt Gardner
00:30

So I wanted to talk to David because there’s this application area of NLP that David has done a lot of work that I think is really interesting and not a lot of people know about. And so, David, what we’re going to talk about is digital humanities. Can you tell us what this is?

David Bamman
00:49

Yeah, sure. I’m happy to. So digital humanities, I think you can think of that as being more or less a community of practice. That’s not too dissimilar from other comunities of practice that have been emerging over the past 30 years. So other examples of this would be computational social science, computational journalism. Digital humanities is something that’s very nebulous. It really encompasses a lot of different kinds of activities that people undertake that is in this intersection of digital anything with humanities anything, right? It could include things as diverse as creating a digital edition of a text, building a public facing website for some work that you’re doing. In GIS, you know, where you’re plotting locations on the map and doing something of service to humanities inquiry with that. I tend to prefer the term computational humanities for the kind of work that we do that really intersects with NLP in particular with these different areas and these different disciplines of humanities like literature, art, film, music, theater, where it’s not just that we have this general desire to bring together digital things with humanities things, but really explore what kind of affordances we have in bringing computation in particular and empirically in- particular to the analysis of objects of inquiry like literature.

Matt Gardner
02:10

So when you say computational humanities then are you focused mostly on literature or are there other applications that you have in mind?

David Bamman
02:16

Okay. So the kind of work that I do is very much focused on literature, but I think the other kind of work that is in this general term would include the use of computation for arts, for film, for music. You know, these other areas that have different disciplines within the humanities that people have been doing a lot of work on, not at an inter-sectional NLP with the humanities, but rather at the intersection of computer vision or of music information tree.

Matt Gardner
02:43

Okay. And literature, I know you spent some time with the Perseus project. Maybe you’re the better person to, to describe this. This is classical literature, right?

David Bamman
02:52

Well. So the Perseus project in particular is classical literature, but, classics in a very narrow sense of, right, so not classics like The Canon, like where you’d see Charles Dickens be a part of the classics, but rather focus very explicitly on literature and new text just in those two languages. But yeah, my own background, I started my academic career as an undergraduate majoring in English and then switched to classics. So my undergraduate degree is in Greek and Latin. Went from there to a Master’s program in applied linguistics and then spent five years working for the Perseus project, which is doing you know, the kind of NLP for Greek and Latin text before going out to a PhD program in computer science as you know.

Matt Gardner
03:34

And so then when you say literature again, just trying to make the scope of what we’re talking about here clear, you could include things like the Perseus project, Ancient Greek and Latin literature, but also more modern novels or what?

David Bamman
03:50

Yes. I mean, I think anything that, anything that falls under the sphere of fiction would fall within literature. And even there’s some areas that are beyond that or exist as border cases. Biographies would also be used as an example that are not quite reality, but also we’ve got quite fictional items. But yeah, so the kind of work that I tend to focus on is more narrowly focused on fictional works. And that’s what I mean by literature.

Matt Gardner
04:12

And so let’s say I’m some scientist in the humanities, what kinds of questions might I be asking about literature that the computational humanities might help me answer?

David Bamman
04:23

I’m not sure if humanists would bristle at being called scientists in the humanities.

Matt Gardner
04:28

Sorry.

David Bamman
04:28

I think many of the might, but I think that the kind of people who do work at this intersection who are in humanities departments right now give you a sample of the kinds works that people have investigated in this space. Let’s see, Angie Piper has done some work on exploring fictionality, which is the idea that, you know, it goes back to work by John Searle and Derrida Bock in fact, who said very famously that there’s some works, there’s no textual property that would identify a text as being a work of fiction. Right. So that if you take some random story, right. So a story like John went to the store, he bought some eggs, he paid for it at the cashier, and he walked outside. There’s nothing within the inherent text itself that gives you a signal that it’s fictional or not fiction.

David Bamman
05:16

It’s either about true events or events that have been imagined, and Searle claimed that there’s nothing within this text that gives you an indicator of it being fiction or not. It’s only within the intentionality of the speaker or in the way that the reader is receiving that intentionality that the fiction, the work fiction itself is defined. And what Piper did in this work was to try to explore whether or not there are distinguishing properties between texts that had been marked as fiction and those which have been, marked as not fiction. And of course finding that with over 90 or 95% reliability, you could classify work as being fiction or not. Not to say that a given piece of text could not also be seen as the opposite class. Right, as being presented in a work of being imagined or being true, but still gives you some sense in which you can use these kind of empirical methods to tease apart this question of what makes a text actually fictional is actual use, right empirically [inaudible]. Laura McGrath has done some interesting work on measuring literary novelty in text specifically in early, modernist literature. So early modernist literature had a very strong bias toward, toward new things, right as Ezra Pound said that you should make it new. So what you looked at in this work was how much a text ends up repeating itself within the body of the text. How much intro, textual similarity. There is uh, again, looking at question of how much you can see this overall desire of an entire movement being embodied within the specific text. Again being measured by computational devices and what counts as being similar. The work of my own that I am probably most proud of is in collaboration with Ted Underwood and Sabrina Lee at the University of Illinois. And what we looked at there was in trying to measure how much attention is being given to characters in fiction as a function of the gender of the characters and the gender of the author.

David Bamman
07:16

So in that work we found that in looking at about a hundred thousand novels, English language novels, over about 150 year period, we saw that women as authors ended up giving about equal attention to male and female characters while men as authors ended up giving three times more attention to male characters than the female characters. I think the ways in which you can see NLP interacted with this space, is in providing these kind of computational and empirical methods that really enable a kind of measurement that can be used for either for testing some hypothesis on its own or for being some piece of evidence that’s being marshaled in the service of a larger argument.

Matt Gardner
07:55

Yeah, I’m trying to imagine answering some of those questions 10 or 20 years ago before we had really good tools, just being really laborious and how NLP could really help this.

David Bamman
08:09

Yeah. Well actually this is one of the things I think is so interesting about this entire area of research that we think about all of these different communities of practice, like the digital humanities and computational social science and competition in journalism as really, you know, reaching there, not hated, but really having a great surge of energy and activity just over the past five or 10 years. But all of these really have really long traditions of empirical work that go back, I would think at least to the 19th century. I mean a lot of classic German philology was just people sitting around counting things to make arguments about how groups behave. In fact, when some of the earliest work I think in this space was that by Thomas Mendenhall back in the 1880s that looked at ways in which they could discriminate works of Francis Bacon, Shakespeare and Christopher Marlowe using the distributions of word link. Again evolving just these simple acts of counting, but in a way that would be a lot more laborious than we have today.

Matt Gardner
09:08

Right. I remembering an example even earlier, wasn’t there some treaty that was supposedly written by a pope in ancient Greek that was like, you might remember this better than I do. I only took just a few classes in the classics, but like one of the first philological studies was disproving, or there was a document that was supposedly written way earlier than it was actually written and you could look at word usage, knowing ancient Greek in order to show that this document was much later than it actually was purported to have been written.

David Bamman
09:37

Oh, interesting. Yeah. No, I don’t know that specific case, but yes, I’m sure there are a million examples of things like that,

Matt Gardner
09:42

But, again, yes, that you could view NLP as a means of speeding up this kind of analysis. Not that it hasn’t been done before, but now it’s accessible to a lot more people and with a lot less overhead. Yeah. So you can answer more of these questions and answer them more quickly.

Waleed Ammar
09:57

Could you elaborate a little bit on how this could be used as a measurement device for literature. So, you mentioned for example, I didn’t find gender, what other problems you perceive NLP community helping with? Whether it’s already something that’s already like we have good models for or not.

David Bamman
10:14

Yeah. Oh, for sure. The core areas that we see as being part of the NLP pipeline end up resulting in some kind of measurable quantity, right? So parsing, taking named entity recognition, syntactic parsing, uh co-reference. I mean all of these, you end up being translatable as formal qualities, narratologically for a lot of questions that you could answer about literature. So an example of this would be so the work that we did in trying to measure the amount of attention being given to character. It’s not gender there. That’s the measurement. The atomic unit that we’re trying to measure is what character is and how much attention you give to a character. And to get to this notion about what a character is, you need to be able to have methods that are good for recognizing what the names are, first of all, that are being mentioned in a text and give you a sense of what the character’s names are. And more importantly, you also need to have some sense of co-reference of those names, and so the co-reference of the names and co-reference of all the pronouns that are co referring to the same thing as those things. What I say that we’re doing in NLP is developing these kinds of algorithmic measuring devices that can be used for other fields. That part is really what I mean in this particular context, that it’s NER in particular in conjunction with coreference resolution that gave us the ability to point to a character in a text and then think about ways in which we can decide what attention to that character needs. And for us, you know, we treated attention as being the amount of things that that character did, right? The number of activities that they were the agent of the things that they had done to them as being the patient. We also looked at the amount of dialogue that they have. And that again is a structural quality. It’s a structural thing that is, it involves being able to parse the text blown up and assign different regions of the dialogue to the person who’s speaking it and then resolving that mention of the person to the entity to which they’re correct. So all these revolve around being able to get these kinds of structural properties right in a way that can enable this atomic measurement of the character. But even the low level things, you know, beyond this high level structure, the character, the low level things were also important. So Matt Wilkins at Notre Dame also had some interesting work a couple of years ago and looking at the, how much attention is being given to different cities in the United States in, I can’t remember if it was fiction or nonfiction, but before and after the civil war.

David Bamman
12:35

So in that case you, you need to have a named entity recognition system that can recognize place names well enough in text to then treat the place name as being this atomic unit that you want to develop this measurement around right, to say how much attention is being given to Boston or to Berkeley or to Seattle before and after the civil war. So all of these low level things that we work on in NLP, in the service of some higher level system are often useful on their own. We’re building some argument around it in this space.

Waleed Ammar
13:03

I See, I imagine the accuracy of the various tools will be, will vary, depending on what year the story was written or, maybe different dialects. What are the tools that you can use in order to account for these differences?

David Bamman
13:18

Yeah, so this is a great question. This is one of the things that keeps me up at night and it really drives a lot of the things I work on right now. That you know, one of the dirty secrets that we have in NLP is that a lot of our systems that we have state of the art performance for, work really well, not just on a newswire but on the 1989 Wall Street Journal, right? So if you take a model, that has been trained on the entry bank for parsing or for any of these other data sets that have been built on top of the entry bank for all these different layers of annotations. If you take a model that’s been trained on that and then just run it out of the box on Charles Dickens or Jane Austin, it’s going to blow up in very predictable ways. Right. So we see time and time again for a lot of different problems we parse speech taking names and recognition, parsing the accuracy for all of these methods that they’ve been trained on. Newswire tends to plummet by like 20 points in absolute terms for across a lot of these different tasks and they’re often embarrassing errors too. And one of my favorite examples of this is just a very simple sentence. “Yes, comma” said Jane, right, where every human would be able to say that Jane is a syntactic subject of the verb say, but because you very rarely see the subject of this other verb showing up after the verb in news texts Jane is often tagged by these out of the box taggers as the object of say, right there like completely ridiculous. And assuming that there is just a nulll subject at all. So yeah, so it’s a real problem. The fact that you can’t just assume that these methods that have been trained on newswires weren’t to work well for texts that are either historical, you know, that had been written a century before 1989 or simply in a different domain. It could even be contemporary with 1989, but just in this kind of fictional context that has much more complex sentence structure, much longer sentences, and just this level of figurative language used that you don’t see in news. The reason we can go about accommodating this this disparity in performance or trying to rectify it. I mean, I found that, I mean, using contextual word representations like ELMO and BERT really help, you know, go some way for, for mitigating this error for just adapting the representations to the fact that we are in a fictional sentence and not in a new sentence does get you some gain.

David Bamman
15:41

All the methods that we have available to us in NLP for domain adaptation, you know, are potentially of good use here. For us, we know that we want to be able to measure how well any of these systems are going to perform on literature overall. So in order for us to do that, we have to have data that we had annotated where we know what the true labels are. So the work that we presented at NACCL a couple of weeks ago and we’ll also present the ACL in Florence is a new data set that we are creating for a hundred different English language novels where we’ve annotated all of the entities in them, uh nested entities. So complex structure including uh people, places, organizations of a couple of different types and also all of the event triggers as well. Because we know that we have this data to at least say how poorly existing models are doing.

David Bamman
16:29

We can then also build more data to train those systems natively on this domain. And what we found here is that when we do that, when we train a system natively on literary texts, we end up performing much better than if we were to doing it to train that system on something that, a domain that is not used; a domain that is not literature.

Waleed Ammar
16:47

Yea, I think the first data set that addresses this problem so directly.

David Bamman
16:51

Yep. This is the object of center. We know that we want to develop a lot of this NLP to work well for literature. And if you want to do that, we really just need to have data that’s in that domain to help make sure that these systems are good and even if they’re not good, just to know exactly how bad they are. I just, it’d be, it’d have some number to help drive or work in this space to help shape it going forward is important.

Matt Gardner
17:11

So you’ve created and released a data set or two recently. Are there others that are in this general space that you might point people to?

David Bamman
17:20

Yeah, so I know there have been a lot of other techniques that people have developed for trying to make these different aspects of the pipeline better for literature, in particular. Juilan Brooke, Tim Baldwin had done some work on this, looking at trying to essentially induce brown clusters and then giving tight category cluster level labels to Brown clusters to say whether the cluster is about all people or all locations. And that has been pretty successful in identifying a lot of these entities for techs. There tends to be a long tail of resources. So a number of resources for specific authors or literary texts, that tend to be relatively small in size, but really optimized for making these methods work well in those domains. But this I think is probably the biggest resource out there for this kind.

Matt Gardner
18:06

Coming back to an issue we touched on briefly a little bit ago, you worry about the domain shift and accuracy if you’re using NLP as a measurement tool in this particular area. And will lead us to question about general accuracy here. I wonder particularly about bias and say gender bias or any of these things. Like if you want to actually ask questions about gender and literature and we know that our systems are biased, how can you be confident in using these as a measurement tool?

David Bamman
18:36

That’s a great question. I think that in all these cases it’s important to have some kind of validation that that lets you quantify exactly what kind of biases are going to creep into this work. And one of the things that we found just in creating this data set was that if you took a model that’s been trained on the H2005 data set for a method used for newswire and use that model trained on newswire to try to identify people in literary texts that you end up having a very strong disparity in performance as a function of the gender of the entity that you’re recognizing. So that if you look at the recall between how well you’re able to identify true mentions of women versus men that a model trained on H2005 ends up recognizing men much better than woman. Right, I mean something on the order of like 10 absolute points of recall better for men. If you look at the H2005 news data you can see why. I mean most of the articles are about men. You simply just don’t have a lot of mentions of women in these articles at the same proportion as you do of men. So if you’re taking them all and applying it to a different data set where you have more parody in dimensions between women, you would expect to see that kind of reports, and we do.

Matt Gardner
19:50

That’s really interesting and problematic. We need to do better.

Waleed Ammar
19:53

That reminds me of a talk that ABC by Claudia Wagner, she was talking about how the users that read the content is being used in a lot of ways to kind of like as a replacement of standard more traditional ways of surveying people’s opinions or understanding people in general. Then the same problem happens there where the kind of content that’s available online where people contribute is biased in various ways or choose basically advocating that we should be using the same methodology that’s been traditionally used in survey research in order to account for the various kinds of errors we were making this measurements.

David Bamman
20:31

Yeah, absolutely.

Matt Gardner
20:31

So I think we’ve talked a little bit about how NLP helps in answering questions in the humanities. Do you think the contribution can go the other way in some cases where this, this kind of humanity or literature helps NLP in general?

David Bamman
20:49

Oh yeah. I mean absolutely. I mean the one hand literary analysis on its own is really challenging domain. Alright. So making sure that the tools that we’re developing within NLP work well, not just for news, for Twitter, for product reviews on Yelp. I mean having another example of a, of a challenging domain that’s very difficult. Those I think is, is useful for us just in terms of making our tools more robust to make sure that even if you’re not talking about optimizing the literature, you can see it as being an example of a domain that is like the others and be an example of something that you know you want to do well on overall, right broadly across them all. I think that a lot of the work where in metaphor in particular and figurative language I think has a lot to bring in, from this entire area of literature that is steeped in this kind of theory. So I think we’ve seen even a lot of work and metaphor recognition, metaphor identification that’s drawn on a lot of this theoretical work from in philosophy and English. And I think a lot of these questions about fairness and bias I think really also are informed by fundamentally humanistic lines of inquiry.

Matt Gardner
21:53

So are there particular tools that you wish you had better access to or that you wish performed better? I guess to state this a little better, you think NLP in literature analysis is helpful as a measurement tool. What kind of measurement tools do you wish you had that NLP doesn’t give you?

David Bamman
22:10

That’s a great question. So I think that’s generally a lot of the low level components of the NLP pipeline could be better for literature. I think that the hard cases come in the fact that novels in particular are really long, right? So the average length of a novel or something like 150,000 words and the kind of methods that we have developed for application areas like co-reference resolution just don’t scale well to, to documents that are that long. Or we can end up having co-reference work well in a really local context, right? It’s, it’s a lot easier to link a pronoun with its nearest antecedent than it is to state that you know a given car in a book that you see at the very beginning of it is the same car that you see the very end, Right. That’s almost an impossible problem that I think we just need to have a lot more different sort of scope in how we think about this problem to really think about that we wouldn’t necessarily see from just working primarily on news documents.

David Bamman
23:10

But beyond this question about what specific kinds of tools you think would be useful in this space of NLP for as literature in particular. I think that one of the real advantages in this space, one of the real, what were the real space and opportunity is in literature is in thinking of what these new problems are in the first place for literature I think this space of innovation here really is pretty tremendous. So, if I was to encourage somebody who wants to get into this space, if they should choose to work on making some particular tool better or in using what they understand about the state of the art for NLP Right now in 2019 I would encourage them actually to really think about what these new problems would look like in the space of literature. That we could apply all of these different methodologies and tools and methods that we have been developing for the past 30 years to explore what a new question looks like in this space.

David Bamman
24:08

I mean, just to give you a couple of examples of things that my group is taking about are literary scene code records, right? It’s something that doesn’t really exist for news texts. A literally scene co-reference would be the problem of trying to identify which events are taking place in the same physical location in a novel. So where you know, which actions in which scenes are taking place in the same room or in the same palace or in the same village or whatever. Going beyond that, trying to think of ways that we can use all of these different fundamental tools to create a map of a literary geography. I think that’s a really fascinating question that doesn’t really show up in any other texts that is grounded in the real world, right? So this is another way in which literature is very different from news.

David Bamman
24:51

News has descriptions of places that are anchored in reality, right? They were, every place has at least some latitude, longitude or some boundary box that tells you where on earth it is. With books sometimes this is true, right? Sometimes you have mentions of London or Seattle that you know you can anchor on the earth, but other books you don’t. The Lord of the Rings is a great example of this where it’s an entire imagined universe. And try to think about ways in which we could decide where, what location fits with respect to another. It’s something that we have the capacity to do. We have the capacity to think about how we could apply these methods in NLP to figure out this problem, but it’s one that really doesn’t exist anywhere outside of this entire universe of imagined realities. Lots of examples of these kind of problems like that where we focus not specifically on making specific tools better, but about thinking about what these new problems are. That literature really opens up for us. Now, we could approach with these methods in NLP. I think that is where there’s this real opportunity is.

Matt Gardner
25:56

Yeah, that’s really interesting. I’ve been talking recently with a bunch of folks about what it means to read and understand a passage of text and one of the things we’ve been toying around with is can you reconstruct a scene from like a paragraph description of, say a room, and this gets very similar to this [idea of] recovering a map from the geographic descriptions in a novel though at this particular case it wouldn’t necessarily be restricted to just fiction. It’s an analogous problem. If you think a little bit broader, it’s in a lot of places. Yeah. It’s interesting.

David Bamman
26:30

Absolutely. I think that’s a good example of a thing that you see definitely existing in this universe of literature that could then also be expanded to other domains as well. But you brought up this question about how readers react to a new piece of text, I think that is also a huge opportunity here. We’re thinking about how people react to literature as they’re actually in the course of reading it. Because a lot of this work in this space really assumes that you, I mean computationally, you have access to all of the texts in a book instantaneously, right? So if you’re trying to, you know, to tackle authorship attribution, right? To figure out who the author is of a given text or genre connection, right? To figure out if a book is a drama or a crime novel or a science fiction novel. A lot of those methods assume that you can see the entirety of the texts from the very first word to the very last word and use all the information in making your judgments about what class you know is appropriate here.

David Bamman
27:27

But there’s an entirely different area of research that you can imagine here that asks, what does the reader know at the moment of each word in the novel? Right? So if you’re starting from the very beginning of the book and read to the very end, what is the state of your understanding of that fictional world on page 10. How does that change from page 11? How is it changed from page 100? It’s interesting because when you take this approach of how you model what a reader knows about what they have read a lot of really interesting questions open up like foreshadowing, like what’s the relationship between the world that’s being built up for the reader, you know, by page 10 and what is predicted to happen, you know, a hundred pages later. Who ends up being the bad guy at the very end of the book. Can you predict that; at what point are you able to predict that from the reader’s point of view? So I think the more ways that we can think about treating this text as being not just a static entity but rather a thing that a real person is reading and has impressions of in the course of their reading; these temporal dynamics. I think there’s a whole range of interesting questions that could be asked there as well.

Matt Gardner
28:35

Yeah, that’s really interesting. Maybe I’ll spring a question on you that I’ve asked a whole bunch of other people. I’ve been trying to categorize reading comprehension; what it means to read and you can imagine at a very basic level, this is like understanding the the linguistic structure of every sentence in the text. Typical. You might judge that with syntactic analysis if you really want to or whatever. Then discourse analysis, there’s discourse parsers that try to get at this. There’s co- reference entity, event, co-reference, this kind of stuff. Then implications of the paragraph that you see. Like if I see that Bill loves Mary and that Mary was diagnosed with cancer, what does that tell me about how Bill must be feeling right now? That’s one level of understanding. The implications of what I read and then being able to recover a world model, being able to understand the communicative intent of the author who wrote this and how well they succeeded in their intent. I’m curious what your perspective is from literary analysis. Like what other things should I add to this categorization of what it means to read?

David Bamman
29:37

Let’s see. I mean, I don’t know if I would add anything to the perspective of literary analysis. I think if anything, trying to analyze what the author’s intent is I think is a very controversial thing in the space of literary analysis where you know, there’s a whole set of theorizing about how it’s the reader who’s the one who makes the meaning of the text by the act of reading and whatever intent the author had baked into the text is not what the reader is going to end up pulling out. It’s the readers who construct the meeting from the words that they see given on the page. But I think generally that’s an interesting question about how you go about trying to learn what these implications might be. One of the things I’ve always wanted to work on was to, to, to take pairs of people who are talking to each other and infer together what their own shared common knowledge must be for that conversation to make sense. What things is personal A presuming person B knows by saying a certain thing. It’s very much into this framework of rational speech, but for perspective of reading comprehension, I think you’re thinking the same thing, right? That if you can infer what a plausible state of mind would’ve been of a person who’s authoring the question or authoring the paragraph that I think would get at what some true component of reading comprehension must be in the end. But it’s a hard one.

Matt Gardner
30:52

Okay. To wrap up, my last question is about what do you think are the most interesting open research questions? You mentioned one of them about like just trying to think about what are the interesting problems. Is there anything else you would, you would give as an interesting open research problem for digital humanities?

David Bamman
31:08

Yeah, I mean, absolutely. There are a whole slew of these open research questions and people are going to find them interesting to different degrees. For me in particular, I think that the ways of having an operational definition of plot is really, really interesting and really something that we, there’s a lot of theoretical work on and even some computational work, but for the most part that work has been really focused on trying to represent plots or narratological structure as really being just the distribution of sentence over the temporal domain of it all. Right, to say how happy or sad is the tone at a given moment in time. It goes back to a very famous video that Kurt Vonnegut did essentially defining what typical story arcs look like from their perspective of the sentiment. We know that plot has to be something more complex or we know that plot has to involve people, it has to involve places, it has to involve times that actions are occurring.

David Bamman
32:00

In many cases the main theories also involves important objects. And we have a way of at least focusing our attention on getting all of those individual components right from the perspective of NLP or be able to recognize people, places, you know, some at least early work in ordering those events and along some timeline. But how that goes about combining together to create some operational definition of plot or, or a storyline I think is an open question that is almost within our grasp. We have the building blocks to get there, but we’re, we’re not quite at that state of being able to put them together. So I think there is really amazing work to be done in doing just that. And I think even in this larger question of how we use these kinds of NLP methods to get at different narratological primitives, I think is really compelling.

Matt Gardner
32:55

It’s interesting. That’s not the direction I thought you were going to go when you brought up plot. I was expecting something about like prototypical narrative like The Hero’s Journey. Like, can you classify different novels into these categories?

David Bamman
33:07

Oh, no. So I mean, I think there is work that’s been done on an aspect of that. A lot of that dates back to Russian formalism back in the 1920s. I don’t know. I don’t particularly, I think there’s a lot of other interesting questions that could be asked about how individual works are, that don’t fall into this notion of an archetype or a stereotype is really what I think what they ended up being reduced to. That these works of literature I think are, are often much more complex than being able to be reduced down to these classical stories. Okay. I would disagree.

Matt Gardner
33:39

Yeah, that’s fine.

David Bamman
33:41

Cool.

Matt Gardner
33:41

This was a really interesting conversation. Is there anything that we missed that you want to talk about or any final thoughts before we conclude?

David Bamman
33:47

I don’t think so. I think we covered a lot of important stuff.

Matt Gardner
33:50

Great. Thanks for talking with us. I’m looking forward to seeing more work on literary analysis in digital humanities.

David Bamman
33:55

Okay. Thanks for having me.