Affect as a proxy for literary mood

arXiv:2304.02894 · cs.CL · Submitted 2023-04-06 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Affect as a proxy for literary mood".

Jane: The gist The study proposes using affect as a proxy for mood in literary texts by exploring differences between computationally detecting tone versus detecting mood through affective word embeddings and…

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we're looking at this paper today called "Affect as a proxy for literary mood." Basically, the whole point of the study is to see if we can use how words make us *feel*—affect—as a way to figure out the *mood* of a piece of literature.

Jane: That sounds really interesting because mood and tone are often confused, right? The paper proposes using affect as a stand-in for mood in literary texts, and they test this by comparing two different ways to detect things: detecting tone versus detecting mood.

Lu: They use affective word embeddings to look at the distribution of emotion words in different segments of text, which lets them pinpoint where the mood is created and how that connects to affect #pg2. If you can map that out, it should give you a clearer picture of the emotional effect on the text than just looking at tone detection alone.

Meng: But what does this actually mean for us on a day-to-day level? Like, if we're building an AI system, how does using affect as a proxy translate to something we can deploy? Is it just about getting some kind of score?

Lalam: Lalam connects that idea to the real world; they also present this method for enhancing emotion lexicons. They make sure this enhancement takes into account both semantic shift and the specific domain of the text, which is supposed to produce results that match what people actually say in qualitative analyses #pg1.

Tom: Yeah, so they aren't just looking at words in isolation; they're trying to create a more robust tool for understanding emotion by making sure it understands how meaning changes and where those words come from. They aim for real-world congruence with existing human interpretations of mood.

Jane: That makes sense, because the authors mention that sentiment recognition tools, like some earlier work on CLS approaches, are often not useful to literary scholars studying affect in a traditional sense #pg3. This paper is trying to bridge that gap between computational methods and literary analysis by focusing on things like tone and mood.

Lu: They specifically test this hypothesis quantitatively first on different segmentations of the texts before they move into the qualitative testing against expert close-readings #pg1. That’s a solid way to build up the argument for what works best in practice.

Paper summary: Meng: I wonder about their method for segmenting the text, though. They mention using two specific sections as targets: first, the first three paragraphs of each book, and second, the first two hundred tokens from each chapter #pg1. Does that segmentation actually capture enough of a typical literary mood?

Tom: That’s a good practical question for Meng. The finding they report is that using these two different text sections approaches a distribution of emotions in the lexicon when you use this specific segmentation #pg1. That suggests focusing on the opening paragraphs might be key to understanding the mood of a text.

Jane: So, what do we get from that? The qualitative analysis showed negative emotions were depicted and triggered, things like feelings of deception, fear, anxiety, disgust and hatred #pg1. It’s not just abstract emotion; it’s specific emotional content they are finding in those opening paragraphs.

Lalam: Lalam finds that the results for the texts show that the valency and intensity of emotions in those opening paragraphs approach a distribution of emotions in the lexicon when using this segmentation #pg1. It's showing a strong link between where you look in the text and how much emotion is present there.

Tom: And they suggest that because focusing on those opening paragraphs gives you a fairly accurate idea of the mood, it’s more stable and less likely to get confused with tone #discussion section]. This leads into what we need to consider next about why this specific starting point matters so much for literary analysis.

Jane: So, if you're a reader or an analyst looking at a text, this suggests that the mood of a book is most reliably found in those initial sections, rather than having to track the entire thing #discussion section]. But what does this actually change for someone who only listens to the show? It makes you think about where to focus your energy when you’re trying to grasp an author's emotional intent.

Lu: It shifts the computational focus toward segmentations that are more likely to be stable indicators of mood, which is a big idea for representation learning #pg2. If we can get the model to reliably find that initial affective distribution, then we have a better anchor for mood detection.

Paper summary: Meng: From an engineering side, if you want to deploy something based on this, you’re looking at how much compute it needs for those embeddings and lexicon enhancements #discussion section]. And they did mention trying to fine-tune Finnish BERT but hitting vocabulary limitations because the data was too different from the original training data for Finnish BERT.

Tom: That’s a real challenge when working with smaller, domain-specific languages. It shows that even with powerful models, you have to handle the mismatch between general training and specialized literary text #discussion section]. The authors settled on using the Turku Neural Parser because it gave them the best results in an easy-to-use pipeline for their specific data.

Jane: So, to summarize this paper "Affect as a proxy for literary mood," they propose using affect as a way to find mood, testing it by looking at how emotion words are distributed across different text segments and improving emotion lexicons to match real human interpretations #pg1. The main conclusion they draw is that the opening paragraphs of a text are the most stable place to look for an accurate idea of the mood #discussion section].

Lu: And their final point is that while evoking emotional effects in literature involves more than just emotion words, it also includes things like narrative point of view and style #pg1. They hope to add established emotion categories from literary affect studies into the lexicon to make it even better for this domain #discussion section].

Meng: So, the paper suggests that for larger collections of text where word choice matters a lot, using purely lexicon-based methods and affect as a proxy for mood can be supported by big data results #discussion section]. But they also flag a limitation: they’re aiming to incorporate intentionality detection in future work to better separate tone from mood #discussion section].

Tom: Exactly. The preliminary studies show that the "big data" results support using purely lexicon-based methods and affect as a proxy for mood when dealing with larger collections of text where word choice is an important factor #discussion section]. It’s not a perfect solution, but it gives us a strong foundation for how we can use affective signals to study literature.

Jane: So, the big idea here is that we can use computational tools to get more consistent readings of mood by focusing on stable parts of the text and making our emotion dictionaries smarter about context #discussion section]. That’s what they’re saying about the implications for understanding literature computationally.

Conclusion: Tom: So we've been looking at how this paper "Affect as a proxy for literary mood" uses word emotions to figure out the feeling of a story and what they found is that focusing on the first few paragraphs is a pretty solid way to do it

Jane: Right, so it takes words that make us feel—affect—and turns them into an idea of mood in literature, but the big win they show is that you don't have to read the whole book for this kind of analysis

Lu: Exactly. They found that by looking at those initial segments of text—the opening paragraphs and some tokens from each chapter—the emotional distribution really lines up with what people say is the mood

Meng: And that means if you want to build a system to gauge a piece of writing, you don't need to analyze every single sentence; focusing on that stable beginning section gives you a reliable starting point

Lalam: I'm pretty sure I can say this with high confidence. The results show that using this segmentation—those opening parts—gets the emotional intensity and valency right in line with the emotion lexicon they built

Tom: So, when you look at the authors and the title, "Affect as a proxy for literary mood," what's really going on there? Is it just a fancy way of saying they're using feelings to guess how a book feels

Jane: It’s more than that. It suggests that we can use computational tools to get these consistent readings of mood by focusing on stable parts of the text and making our emotion dictionaries smarter about context

Lu: They also pointed out that while this lexicon approach works well for large collections, it doesn't capture everything about literature, like the narrative point of view or the author's style

Meng: So they’re not suggesting this is a perfect solution for every literary question, but it’s a strong foundation for using big data to get an initial sense of emotional direction in text

Lalam: And they actually say they plan to add more established emotion categories from literary studies into that lexicon later on to make it even better for this domain

Tom: So the big implication is that we can use purely lexicon-based methods and affect as a proxy for mood when dealing with large collections where word choice is really important

Jane: And they're hinting at future work, specifically wanting to incorporate intentionality detection down the road to better separate tone from mood

Lu: That separation of tone and mood is what they want to tackle next, because that’s the key distinction they think will make their tool much more powerful in the long run

Emily Öhman, Ωikka Rossi

Waseda University · University of Helsinki

cs.CL

Submitted: 2023-04-06

Updated: 2023-05-23

Comments: Submitted to Journal of Data Mining and Digital Humanities

Journal ref: Journal of Data Mining & Digital Humanities, NLP4DH (August 13, 2023) jdmdh:11164

DOI: 10.46298/jdmdh.11164

Code: https://github.com/esohman/FinLit-corpus5https:

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 81/100

The gist: The gist The study proposes using affect as a proxy for mood in literary texts by exploring differences between computationally detecting tone versus detecting mood through affective word embeddings

Key concepts

Affect as a Proxy for Mood: This research explores using the emotional tone or 'affect' found in words within literature as a stand-in for determining the overall mood of the text. The goal is to see if analyzing how emotions are expressed computationally can accurately reflect how readers perceive the mood of a story.
Affective Word Embeddings: These are computational tools that map words into numerical spaces based on their emotional properties. The researchers used these embeddings to look at the distribution of emotion within different parts of a text, helping them compare affective patterns between sections.
Lexicon Enhancement: This method involves improving an existing list of emotion-related words (the lexicon). They created a new way to enhance this lexicon by considering both how words change meaning over time (semantic shift) and the specific context or topic of the text to ensure more accurate results.
Tone vs. Mood Detection: The study specifically compared two computational goals: detecting 'tone' (the author's attitude) versus detecting 'mood' (the feeling evoked in the reader). They found that focusing on opening paragraphs is a better way to capture mood than trying to detect tone across the entire text.

Terminology

Embeddings — 1× in this paper · explained in 2 episodes

Summary

The gist The study proposes using affect as a proxy for mood in literary texts by exploring differences between computationally detecting tone versus detecting mood through affective word embeddings and enhancing emotion lexicons to produce real-world congruent results.

How it works

The researchers propose using affect as a proxy for mood in literary texts to explore the differences between computationally detecting tone versus detecting mood through affective word embeddings and enhancing emotion lexicons to take both semantic shift and the domain of the text into account producing real-world congruent results closely matching both contemporary and modern qualitative analyses. They utilize affective word embeddings to look at the affective distribution in different text segments. Furthermore, they present a simple yet efficient and effective method of enhancing emotion lexicons to take both semantic shift and the domain of the text into account producing real-world congruent results closely matching both contemporary and modern qualitative analyses.

Data Collection

The data collection involved downloading nearly 1000 literary works published in Finnish around the year 1900. The filtering criteria for the downloaded books included (1) the language was Finnish, and (2) the text was in utf-8 plain text format. The final data consists of 975 books instead of 1000 due to encoding issues. The final dataset comprises 2,938,032 sentences and 41,417,116 tokens.

Methodology

The preprocessing steps involved lemmatizing the texts and splitting them into paragraphs. They attempted to fine-tune Finnish BERT but found the vocabulary was not improved sufficiently to work with data so different from the original training data for Finnish BERT. The researchers settled for the Turku Neural Parser as it produced the best results in an easy-to-use pipeline. For mood detection, they used two different text sections as targets: the first three paragraphs of each book, and the first 200 tokens from each chapter in each book.

Lexicon Enhancement

The study utilized the Finnish Emotion Intensity Lexicon (FEIL) as a starting point for detecting affective terms. This approach showed promise and was very accurate within the small sample size in both recognizing semantically similar words and in mitigating issues with semantic shift.

Findings

The results for the texts showed that the valency and intensity of emotions in the opening paragraphs of the book and the opening paragraphs of all chapters approach a distribution of emotions in the lexicon when using this segmentation. The qualitative analysis showed that negative emotions were depicted and triggered, such as feelings of deception, fear, anxiety, disgust and hatred. The study recommends that the quest for mood should begin with the opening paragraphs of a text.

Discussion

The computationally derived results correspond highly with qualitative evaluations of the same target texts in terms of established interpretations of mood specifically when compared to the emotion word distribution of the first three paragraphs. The study suggests that focusing on the opening paragraphs provides a fairly accurate idea of the mood because it is more stable and less prone to confusion with tone. The qualitative analysis demonstrated that evoking emotional effects in literature is not restricted to emotion-associated words but also includes facets like narrative point of view and style. The results support using purely lexicon-based methods and affect as a proxy for mood when dealing with larger collections where word choice is an important factor. The authors hope to add established emotion categories from literary affect studies to the lexicon to further improve its usability in the literary domain. The study aims to incorporate intentionality detection in future work to further separate tone from mood. The preliminary studies show that the big data results support qualitative analyses and further justify the use of purely lexicon-based methods and affect as a proxy for mood when dealing with larger collections of text where word choice is an important factor. The authors hope to add established emotion categories from literary affect studies to the lexicon as a measure to further improve its usability in the literary domain<ref:2304.028971,we hope to add established emotion categories from literary affect studies (see e.g.

Improvements for AI systems

textbfEnsure Lexicon Robustness via Semantic Embedding Lookup for Additions/Modifications: The method developed to objectively add words to the lexicon created word embeddings of our texts and used them to look up words in the lexicon with high cosine similarity to mitigate semantic shift, which can be crucial when incorporating new emotion categories or addressing noun/verb distinctions in languages like Finnish. This allows for an objectively add[ing] words to the lexicon based on contextual similarity rather than subjective interpretation of intensity scores, reducing new biases into the lexicon based our own interpretations of emotion intensities.
textbfIntegrate Segmented Analysis for Mood Detection: The methodology suggests that mood is better detected by focusing on specific text segments, noting that "the tone of a literary text tends to shift much more within and between chapters, and therefore by focusing on the first paragraphs of each chapter, or even the opening paragraphs of the first chapter only, we can get a fairly accurate idea of the mood. This enables an AI system to generate more nuanced mood assessments by comparing distributions across different segmentations (first three paragraphs versus the first 200 tokens from each chapter").
textbf Differentiate Tone and Mood Detection: The paper proposes using emotion detection/sentiment analysis as a proxy for mood while explicitly aiming to separate tone from mood by exploring approaches like the use of intentionality detection, which is suggested for future work. This allows the system to distinguish between how the reader feels about the text [mood] and how the implied author feels about it and uses words to convey their own attitude towards a topic or subject [tone].
textbf Optimize Lexicon Curation Iteratively: The process requires that lexicon maintenance is a iterative and continuous cycle, stating that these two steps are iterative and continuous in that they should be repeated whenever the lexicon or lemmatization is altered. This mandates an automated feedback loop where changes to the lexicon necessitate re-verification of word associations and intensities using cross-checking mechanisms like inter-annotator agreement scores.

Abstract

We propose to use affect as a proxy for mood in literary texts. In this study, we explore the differences in computationally detecting tone versus detecting mood. Methodologically we utilize affective word embeddings to look at the affective distribution in different text segments. We also present a simple yet efficient and effective method of enhancing emotion lexicons to take both semantic shift and the domain of the text into account producing real-world congruent results closely matching both contemporary and modern qualitative analyses.