Combining Qualitative and Computational Approaches for Literary Analysis of Finnish Novels

arXiv:2406.01021 · cs.CL · Submitted 2024-06-03 · Read on arXiv

Listen

Transcript

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

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "Combining Qualitative and Computational Approaches for Literary Analysis of Finnish Novels".

Jane: The gist The study examines how computational methods of sentiment analysis can be used in conjunction with a qualitative or more ‘traditional’ approach to literature and affect to learn what…

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

Title and authors: Jane: So, beyond what they did in this specific study, what are the suggested improvements or future directions that this paper points toward for how we use these tools in this field?

Lu: They suggest that the system can perform a fine-grained type of sentiment analysis by combining emotion detection with affective qualitative analyses of Finnish literature conducted by domain experts.

Tom: That means going beyond simple positive or negative valency and actually measuring the emotional intensity and valence of each individual emotion-associated word in a text, which is a significant step up from basic sentiment scoring.

Jane: And they also suggest using quantitative output to determine traditional descriptions of an author’s style, showing how emotion arcs reflect those styles through specific affective vocabulary rather than just general topic modeling.

Meng: That would be useful if we could use that to compare authors based on their specific emotional tone, not just whatever general themes they write about. That moves it toward a more detailed kind of profiling.

Lalam: They also suggest designing the system to address lexicon limitations by incorporating methods that account for context, because lexicon-based analysis can be criticized for not taking context into account as much as machine learning models can.

Tom: So they're trying to mitigate the issue where a semantically negative sentence might read positively depending on the reader’s perspective by making the model more context aware.

Jane: Plus, there's an idea about genre and period adaptation, where the system could be specifically tuned using a carefully curated emotion lexicon adapted to Finnish turn-of-the-century literary texts.

Lu: That helps prevent misclassification by addressing cultural connotative differences noted in previous lexicon revisions, ensuring the model’s associations are culturally relevant and as objective as possible.

Meng: From a production view, tuning it for specific cultural contexts is necessary if we want it to be reliable on real traffic, not just a general-purpose tool. That adds a layer of specificity that might be hard to achieve otherwise.

Lalam: Finally, they mention the need for narrative significance filtering, which would allow the system to differentiate between narrative importance and lexical intensity in scenes.

Tom: That addresses the finding that the most intensive scene isn't necessarily the most significant in terms of plot structure. It suggests a future development where we can distinguish between main characters or side episodes based on their emotional impact.

Jane: So it’s about moving from just counting words to actually understanding narrative construction through emotion, which is a deeper level of analysis.

The paper's summary: Tom: So to wrap up the paper "Combining Qualitative and Computational Approaches for Literary Analysis of Finnish Novels," the main point is that a quantitative study on the lexicon can provide supporting evidence on how emotions are evoked in literary texts.

Jane: It shows that it can confirm impressions, like the melancholic background mood in Aho’s Rautatie or support evaluations of positive moods in Jotuni’s Everyday life.

Lu: And they found that quantitative representation confirms the impression of Aho’s work by connecting higher intensity words like hajota, hätä, and uskaltaa to his style.

Meng: While they also found evidence for the strong negative emotions previously attached to Sillanpää’s Sacred Misery through intensive negative emotions in his work.

Lalam: The paper ends with the comparison of qualitative and quantitative analysis indicating that existing sentiment analysis methods still have many challenges and need further development.

Tom: The study demonstrates that computational analysis is largely in line with a qualitative reading of the lexicon intensity, but a nuanced view to the interpretation of the narrative still requires human analysis.

Jane: It’s about fine-tuning our methodology to detect mood more accurately so we can understand how mood is constructed in literature.

Lu: I think this paper sets up a framework where you use computational methods as a support tool for close-reading based analyses rather than replacing the traditional reading entirely.

Tom: So that's everything on "Combining Qualitative and Computational Approaches for Literary Analysis of Finnish Novels." The core idea is using computational methods to back up qualitative readings of emotional intensity.

Jane: It shows how these tools can confirm impressions about mood in specific Finnish novels, giving us concrete data to support our understanding.

Lu: It’s about using the work as a framework where you use computational methods as a support tool for close-reading based analyses instead of replacing the traditional reading entirely.

Meng: It’s useful for seeing how emotion is distributed across large bodies of texts, which helps us see patterns that might be missed in a smaller set.

Lalam: The paper shows we still need to develop these methods to handle context better and narrative significance, which is where the real next steps are.

The paper's improvements: Tom: So, we've seen how they used their computational approach to back up qualitative readings of emotional intensity in Finnish novels, and now they're talking about what that means going forward.

Jane: Right, so this paper points toward a few ways to make this kind of analysis better. They suggest improving the system by combining emotion detection with those deep qualitative analyses done by people who know Finnish literature really well.

Lu: The first improvement is moving past simple positive or negative scores, which means they want a fine-grained type of sentiment analysis that measures the emotional intensity and valence of each individual word associated with an emotion.

Tom: So it's not just saying "this sentence is sad," it’s measuring *how* sad—is it deep sadness or just mild disappointment? That’s a big shift in what the model has to track.

Jane: And they suggest another area, style-based author profiling. They think the quantitative output from these emotion arcs can be used to describe a traditional author's style based on their specific emotional vocabulary and tone.

Lu: That means we could compare authors not just on what themes they write about, but how their specific emotional language differs across books.

Tom: I like that because it moves the focus from general topic modeling to something much more granular, like a fingerprint of an author's feeling.

Jane: Then there's a way to deal with the limitations of just using a fixed lexicon. They propose incorporating methods that account for context, which helps when one word can have two different feelings depending on how it’s used in the sentence.

Lu: They are trying to solve that problem where lexicon-based analysis can’t take context as well as a full machine learning model could, especially with things like metaphorical language or shifting perspectives.

Tom: That addresses the issue of those tricky situations where a semantically negative sentence might actually read positive depending on what you're looking at.

Jane: And they are also thinking about cultural relevance. They suggest tuning the lexicon specifically for Finnish turn-of-the-century texts to avoid misclassifications caused by older, less precise word associations.

Lu: That’s smart because it makes sure the model's emotional associations are culturally accurate, rather than just pulling from a general English or global lexicon.

Tom: So they want the system to be hyper-tuned for that specific historical and cultural context of Finnish literature.

Jane: Finally, they suggest adding narrative significance filtering. This helps the system distinguish between scenes that are emotionally intense and scenes that are actually important to the story's plot structure.

Lu: That means if a scene has a lot of high-intensity words but isn't pivotal to the main events, we can separate it from the truly meaningful moments in terms of narrative construction.

Tom: So they’re trying to build a model that understands both the emotional language and its actual role within the story, not just which words are strongest.

Jane: It sounds like they're moving toward a system that understands both the feeling *and* the structure of what's being felt.

Lu: And it sets up a path for us to explore how mood is constructed in literature more accurately, but we still need to fine-tune the methodology for that nuance.

Conclusion: Tom: So we’ve covered how combining quantitative lexicon analysis with qualitative deep reading shows that computational tools can support traditional studies on Finnish literature affect.

Jane: It really highlights that these models aren't just for scoring positive or negative; they can give us evidence about the specific emotional tone of a text, which is what we need to understand literature better.

Lu: The paper focuses on how they developed a robust computational approach using word embeddings and emotion lexicons adapted for Finnish turn-of-the-century novels, specifically looking at works by authors like Aho and Jotuni.

Meng: From an engineering standpoint, the way they handled that lexicon development is interesting—they used word vectors to fill in missing words based on semantic space similarity, which sounds like a solid way to handle vocabulary expansion.

Lalam: I’m sure of what this paper says; it shows that a quantitative representation of Aho’s work confirms the melancholic background mood they already felt qualitatively.

Jane: And for Jotuni’s work, the computational data provides supporting evidence for the positive mood they described in their everyday life stories.

Tom: The caveat is that while it matches qualitative readings, they admit existing sentiment analysis methods still have a lot of challenges and require more development before they can fully capture narrative nuance.

Lu: They are pointing out that even with this computational approach, a human analysis is still necessary to get the full picture of how mood actually gets constructed in literature.

Meng: That means we can't just deploy these models for final judgments; they need to work alongside human experts for interpretation, which adds complexity to the pipeline.

Lalam: I think this study on "Combining Qualitative and Computational Approaches for Literary Analysis of Finnish Novels" proves that the combination works as a powerful support tool rather than a complete replacement.

Jane: It’s a solid piece of work that shows how we can use these new computational tools to gather concrete, measurable evidence for older theories about literary affect.

Tom: That’s what this is all about—using the numbers to back up the readings, not replacing the human insight entirely.

Lu: Next time we look at papers like this, we should really watch how they tackle those context issues and narrative significance filtering that they flagged as future directions.

Emily Öhman, ıikka Rossi

Waseda University · University of Helsinki

cs.CL

Submitted: 2024-06-03

Updated: 2024-06-03

Comments: Accepted in Scandinavian Studies Journal, issue 97.3 (2025)

DOI: 10.3368/sca.97.3.27

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

Importance score: 79/100

The gist: The gist The study examines how computational methods of sentiment analysis can be used in conjunction with a qualitative or more ‘traditional’ approach to literature and affect to learn what can

Key concepts

Computational Emotion Analysis: This involves using computer methods like sentiment analysis and word vectors to automatically measure and map the emotional valence present in large bodies of text, such as Finnish novels. It allows researchers to quantify emotional distribution across entire texts.
Emotion Lexicon Adaptation: The researchers created a specific vocabulary (lexicon) tailored for turn-of-the-century Finnish literature, based on Plutchik's wheel of emotions. This lexicon was enhanced using word embeddings to assign emotion values to words not explicitly listed, ensuring the analysis is contextually relevant.
Qualitative vs. Quantitative Approach: The study contrasts traditional, expert close-reading with computational methods that provide large-scale data. The goal is not replacement but support: using quantitative results to validate or provide evidence for interpretations made by literary experts.

Terminology

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

Summary

The gist The study examines how computational methods of sentiment analysis can be used in conjunction with a qualitative or more ‘traditional’ approach to literature and affect to learn what can be learned from the classics of Finnish literature by using computational emotion analysis.

Combining Qualitative and Computational Approaches

This article tries to answer this question by examining how computational methods of sentiment analysis can be used in the study of literary works in conjunction with a qualitative or more ‘traditional’ approach to literature and affect This article tries to answer this question by examining how computational methods of sentiment analysis can be used in the study of literary works in conjunction with a qualitative or more ‘traditional’ approach to literature and affect. We present and develop a simple but robust computational approach of affect analysis that uses a carefully curated emotion lexicon adapted to Finnish turn-of-the-century literary texts combined with word embeddings to map out the semantic emotional spaces of seminal works of Finnish literature. We focus our qualitative analysis on selected case studies: four works by Juhani Aho, Minna Canth, Maria Jotuni and F. E. Sillanpää, but provide emotion arcs for a total of 975 Finnish novels. We argue that a computational analysis of a text’s lexicon can be valuable in evaluating the large distribution of the emotional valence in a text and provide guidelines to help other researchers replicate our findings. We show that computational approaches have a place in traditional studies on affect in literature as a support tool for close-reading-based analyses, but also allowing for large-scale comparison between for example, genres or national canons.

Methodology and Data Preparation

The experimental setup includes qualitative analyses of key novels that are used as a baseline to which we compare the results from our computational approach to prove hypothesis (1) and show the usefulness of the 975-book corpus with emotion arc annotations. Our data (the literary works and their metadata) were collected from Project Gutenberg and preprocessed with a Finnish version of chapterize both to remove the legal texts in the beginning and end of the text files and later to examine chapter specific affective language use. The size of our total corpus is 975 comprising the first 1000 Finnish language books available on Project Gutenberg that are actually encoded in utf-8 (25 were discarded due to encoding issues). After experimenting with different lemmatizers and text normalization (“translating” nonstandard words to standard language) we chose to lemmatize the texts using the Turku Neural Parser. The final data consists of 2,938,032 sentences and 41,417,116 tokens.

Lexicon Development and Analysis

We use the Finnish Emotion Intensity Lexicon (FEIL) as a starting point in our work. FEIL is based on the NRC Emotion Intensity Lexicon (Mohammad & Kiritchenko 2018). The lexicon uses Plutchik’s wheel of emotions as a basis with the emotions anger, anticipation, disgust, fear, joy, sadness, and trust with the words linked to one or more of these emotions with a value between 0 and 1 representing the intensity of the emotion for that specific word. To ensure domain- and period-suitability we extracted two lists from the original results. The first one contained the most common words in the texts and the second one contained the most common emotion associated words from the texts. We used a word vector approach (word2vec, Mikolov et al., 2017) where we created a vector space model based on the words in context in all 975 books, and then with the help of cosine similarity look up the words with the closest semantic space match and assigned that word’s emotion association and intensity values to the word that was missing from the lexicon. The final lexicon had a fairly balanced distribution of emotion words with fear, trust, joy, and anger between 1176-1560 each, and anticipation and disgust at 832 and 951 respectively.

Results on Case Studies

The results show that there is a sweet spot in the amount of data lexicon-based methods work best for. Hurskas Kurjuus clearly contains the most and the most intense emotion words particularly negative emotion words (although anticipation is usually considered a positive emotion, words that relate to fear, tend to also be associated with anticipation so anticipation on its own can be either positive or negative). Jotuni’s novel is known for its positive outlook on life and has been referred to as peaceful and optimistic. The computational analysis provides further evidence on the positive mood. In contrast, the high emotional intensity in these scenes includes a high proportion of joy-related words in these scenes. Sillanpää’s novel is rich in accurate depictions of poverty and misery focus on filth and contamination in primitive living conditions. The computational analysis also suggests a peak of emotions in this chapter (the chunks 88-94).

Conclusion and Future Directions

In conclusion, our study demonstrates that a quantitative study on the lexicon can provide supporting evidence on the evocation on emotions in literary texts. The analysis of Aho’s Rautatie and Jotuni’s Arkielämää points in this direction. The quantitative representation of Aho’s work confirms the impression of the melancholic background mood of the novella. In the same way, the quantitative analysis provides supporting evidence for the qualitative evaluation of the positive mood in Jotuni’s Everyday life. An unexpected finding relates to religious vocabulary. The quantitative analysis provides evidence also for the strong negative emotions that have previously been attached to Sillanpää’s Sacred Misery and point to intensive negative emotions. At the same time, the comparison of qualitative and quantitative analysis indicates that existing sentiment analysis methods still include many challenges and require further development. While our study demonstrates the computational analysis is largely in line with a qualitative reading of the lexicon intensity, a nuanced view to the interpretation of the narrative requires a human analysis. We hope to further explore the connection between affect and mood and fine-tune our methodology to detect the mood of a literary text even more accurately and thus contribute to the understanding of how mood is constructed in literature.

How it works

The study combines quantitative tools for literary analysis with qualitative analyses conducted by domain experts. The core idea is to use computational methods like emotion detection or emotion recognition to support a qualitative analysis and use a priori qualitative interpretations of select texts to compare the quantitative results against. This approach aims to showcase simple methods that can be used with a qualitative analysis to gather supporting evidence for traditional theoretical frameworks.

Data Processing Steps

Data collection and preprocessing involved collecting data from Project Gutenberg and preprocessed it with a Finnish version of chapterize both to remove the legal texts in the beginning and end of the text files and later to examine chapter specific affective language use.
Text normalization was performed using different lemmatizers and text normalization (“translating” nonstandard words to standard language) before choosing to lemmatize the texts using the Turku Neural Parser. The final data consists of 2,938,032 sentences and 41,417,116 tokens.
The Finnish Emotion Intensity Lexicon (FEIL) was used as a starting point in our work. FEIL is based on the NRC Emotion Intensity Lexicon (Mohammad & Kiritchenko 2018). The lexicon uses Plutchik’s wheel of emotions as a basis with the emotions anger, anticipation, disgust, fear, joy, sadness, and trust with the words linked to one or more of these emotions with a value between 0 and 1 representing the intensity of the emotion for that specific word.
Word vector approach was used to add missing words to the lexicon by creating a vector space model based on the words in context in all 975 books, and then with the help of cosine similarity look up the words with the closest semantic space match and assigned that word’s emotion association and intensity values to the word that was missing from the lexicon.

Evaluation Process

The paper compares computational results with qualitative analyses of key novels to prove hypothesis (1) and show the usefulness of the 975-book corpus with emotion arc annotations. For instance, in Aho’s Rautatie, while sadness is not highlighted in our quantitative results of emotion intensities in Aho’s Rautatie, the emotion of fear is more explicit in the quantitative analysis of Aho’s novel. The computational analysis matches this increase of intensity by connecting higher intensity words such as hajota (disintegrate), hätä (emergency), uskaltaa (to have courage). For Jotuni’s Arkielämää, the measured intensity of Jotuni’s novel is lower than the emotional intensities of Rautatie, Hurskas kurjuus and Kauppa-Lopo.

Improvements for AI systems

Fine-Grained Affective Mood Detection: The improved system can perform a fine-grained type of sentiment analysis by combining emotion detection with affective qualitative analyses of Finnish literature conducted by domain experts. This allows the AI to go beyond simple valency (positive/negative) and capture nuanced emotional intensity, specifically measuring the emotional intensity and valence of each emotion-associated word in a text.
Style-Based Author Profiling: The system can utilize quantitative output to determine traditional descriptions of an author’s style, as the paper shows that quantitative results reflect the authors' styles through emotion arcs. This enables the AI to compare authors based on their specific affective vocabulary and emotional tone, rather than just general topic modeling.
Contextual Valence Shifter Mitigation: The system can be designed to address lexicon limitations by incorporating methods that account for context, as suggested by the critique that lexicon-based analysis can be criticized for not taking context into account to the same extent as machine learning models. This would improve accuracy when dealing with phenomena like phenomenological perspective, where a semantically negative sentence can appear positive to the reader.
Genre and Period Adaptation: The system can be specifically tuned using a carefully curated emotion lexicon adapted to Finnish turn-of-the-century literary texts to ensure cultural relevance. This prevents misclassification by addressing cultural connotative differences noted in previous lexicon revisions, ensuring the model's associations are culturally relevant and as objective as possible.
Narrative Significance Filtering: The system can be enhanced with mechanisms to differentiate between narrative importance and lexical intensity. This addresses the finding that the most intensive scene is not necessarily the most significant in the narrative, suggesting future development of methods to distinguish between main and side characters, the main and the side episodes in the plot.

Abstract

What can we learn from the classics of Finnish literature by using computational emotion analysis? This article tries to answer this question by examining how computational methods of sentiment analysis can be used in the study of literary works in conjunction with a qualitative or more 'traditional' approach to literature and affect. We present and develop a simple but robust computational approach of affect analysis that uses a carefully curated emotion lexicon adapted to Finnish turn-of-the-century literary texts combined with word embeddings to map out the semantic emotional spaces of seminal works of Finnish literature. We focus our qualitative analysis on selected case studies: four works by Juhani Aho, Minna Canth, Maria Jotuni, and F. E. Sillanpää, but provide emotion arcs for a total of 975 Finnish novels. We argue that a computational analysis of a text's lexicon can be valuable in evaluating the large distribution of the emotional valence in a text and provide guidelines to help other researchers replicate our findings. We show that computational approaches have a place in traditional studies on affect in literature as a support tool for close-reading-based analyses, but also allowing for large-scale comparison between, for example, genres or national canons.