Affective computing applied to literary studies Photo: TUCHONG

Affective computing—using computational devices and algorithms to automatically recognize, interpret, and process features related to human emotions—has become a research focus in contemporary cognitive science. By integrating multimodal data such as images, audio, video, and text, computers can fuse multiple emotional features, thereby recognizing human emotions with greater accuracy and enabling high-quality human–machine interaction.

In recent years, some scholars have applied affective computing to the field of literature, conducting quantitative emotional analysis on the texts of various works of literature and thereby opening a new frontier in quantitative literary studies. Although quantitative literary research has a long history, literary computing has long relied heavily on linguistic models, often neglecting emotion as a crucial dimension. In traditional interpretive criticism research, however, emotional analysis has always been an important avenue of research. In narratology in particular, emotion is regarded as a factor closely tied to narrative structure, leading to the development of the interdisciplinary theory of “affective narratology.” The advent of affective computing, therefore, can be seen as filling a key gap in quantitative literary studies, making it possible to build a higher-level theory of computational criticism that goes beyond traditional linguistic models.

Literary affective computing provides a quantitative analytical tool that allows researchers to explore the emotional patterns embedded in literary works in unprecedented ways. This extends literary computing beyond word frequency and stylistic analysis to the realm of emotional narrative, revealing and summarizing complex patterns of emotional structure in texts and clarifying how writers construct emotional experiences through language. Combining quantitative approaches with traditional qualitative research offers a richer, stereoscopic perspective for literary criticism, deepening our understanding of the characteristics of literature in specific periods, genres, or regions. By examining emotional tendencies expressed in works from different periods or regions, we can trace shifts and distributions of social attitudes and values, and in turn investigate the cultural, social, and economic roots behind these changes. From the perspective of cultural dissemination, the findings of literary affective computing can also support the publishing industry as well as film and television adaptation by informing content creation and market positioning. By accurately predicting which emotional elements are most likely to resonate with readers or audiences, producers can refine and optimize narrative strategies while remaining faithful to the original spirit of a work, thereby enhancing both its appeal and its cultural impact.

Method: Why rely on sentiment lexicons

Quantitative sentiment analysis of texts generally takes two forms: sentiment polarity analysis and emotion analysis. The first divides the sentiment of a text into positive and negative categories, using computational methods to determine the overall polarity and its strength. The second employs a more fine-grained set of emotional categories, locating the text within a multidimensional vector array of emotions.

The technical approaches to affective computing fall into two broad categories. One consists of relatively straightforward methods such as sentiment lexicons or bag-of-words models. The other uses algorithms such as machine learning or deep learning to build predictive models, which typically require supervised training and large, pre-annotated corpus data.

Unlike computational linguistics, researchers in the field of literary computing do not aim to produce universal algorithms. Instead, they generally focus on specific works or corpora, seeking to address particular questions related to them. Because no large-scale, publicly available corpus of literary texts annotated for emotion yet exists, most current studies rely on sentiment dictionaries. Compared with methods that depend on annotated corpora, lexicon-based approaches are not only more convenient, feasible, and flexible, but their associated algorithms are also more transparent and interpretable, leaving necessary space for literary interpretation and criticism.

The process of lexicon-based sentiment calculation is clear and straightforward: The system identifies and tallies emotion words in a text, matches them to the lexicon, assigns sentiment polarity, intensity, or category, and then performs algebraic operations on the corresponding values. It is precisely because of the simplicity of this mechanism that the use of sentiment analysis in literary research has expanded so rapidly, becoming an emerging area of inquiry.

Different directions of literary affective computing

Depending on research aims and methods, affective computing applied to literature can branch in several directions. While all roughly share a core process of assigning sentiment values to the text, some studies seek to classify the overall emotional patterns of works; some focus on generating “sentiment arcs” for analyzing narrative structures; some concentrate on the emotions of individual characters and their interactions; others compare emotional patterns across different genres and schools; still others examine the relationship between textual emotions and particular times and places.

Current difficulties and possible solutions

Like most digital humanities projects, the sentiment analysis of literature remains at a relatively early stage of development. The challenges it faces can be summarized as follows.

In terms of technical obstacles, the first is the absence of a dedicated sentiment lexicon for literary computation. Most research relies on general-purpose sentiment lexicons, which assign sentiment values to words based on their usage in various corpora such as news, academic works, and social media, and do not necessarily reflect the emotional nuance of literary usage. Second, many emotion words are polysemous, and less common emotional intentions may be omitted from lexicons, leading to computational errors. Even when multiple intentions are listed, selecting the appropriate one in context remains a persistent challenge. Third, in many machine learning projects, manual annotation of sentiment values has a high threshold of expertise and cannot easily be outsourced, which further reinforces reliance on lexicons. Fourth, attributing detected emotions to characters requires clearer rules and improved accuracy. This problem also affects other areas of literary computing, such as extracting character dialogue networks, but emotion attribution is even more complex than dialogue attribution, as rules are insufficiently complex and often lack universality.

The above obstacles hinder the development of affective computing in literary works towards greater precision and detail. However, these technical challenges are not the whole story. The diversity of styles and the complexity of narratives in literary works also present many additional difficulties.

First, affective computing struggles with plain description, things left to the imagination, metaphor, or implicit narration. Sometimes, rather than describing a scene directly, authors may suggestively hint at events or evoke a certain atmosphere indirectly.

Second, most current affective computing analyzes isolated short text units such as words or sentences, assigning quantified emotion values at the local level. As a result, broader contextual factors—the broader background of the story, build-up from preceding text, or character development—are easily overlooked, producing deviations.

Third, the surface polarity or conventional emotional category of a word does not always apply in a specific literary context. In Chinese poetry, for instance, emotional contrast and counterpoint are common rhetorical devices, which remain especially difficult for affective computing to identify.

To overcome these challenges—both technical and literary—and advance the field of literary affective computing, several promising solutions may be considered.

First, it is necessary to construct a dedicated sentiment lexicon based on a large corpus of literary works, particularly those rich in emotional expression and stylistic diversity. Emotion words should be extracted using natural language processing, then assigned values through a combination of semantic computing and expert annotation.

Second, to move beyond reliance on local text units, context-aware affective analysis models can be introduced. Such models not only consider isolated words or sentences, but also capture broader elements such as story background, buildup from preceding text, and character development. Deep learning architectures like Long Short-Term Memory (LSTM) networks or Transformer models can be used to capture long-range dependencies, helping the system better interpret the true meaning of affective words in specific contexts, reducing bias.

Finally, to identify metaphors, symbols, and emotional contrasts in literary works more effectively, the latest advances in natural language processing—such as pre-trained language models like BERT and its variants—can be employed, since they are capable of grasping deep semantics within texts. Cross-media analytical approaches, drawing on methods from visual arts or music, could also be explored to enrich interpretation of indirect emotional expressions in literary texts.

Of course, these approaches are far more technically demanding than lexicon-based methods, and often exceed the capacity of individual humanities scholars or graduate students. Interdisciplinary collaboration between literary researchers and experts in computer science and psychology should therefore be encouraged, so that human emotion and its expression can be studied from multiple perspectives. In this way, a more comprehensive and refined framework for affective computing can be built, enhancing the analysis of emotion in literature. Looking ahead, as technology advances and research deepens, literary affective computing will very likely play an increasingly valuable role in serving literary and sociological research, humanistic education, and cultural production.

Liu Yang is a research fellow from the Department of Chinese Language and Literature at Chongqing University.