We claim that psychoanalytic concepts and approaches can be highly instrumental in Computer Science applications that heavily involve the human factor, such as social media and other interactive digital platforms, systems, and tools. Indeed, in such text-based systems, psychoanalytic-based approaches have the potential to extract valuable, usually elusive information by providing insights into the underlying dynamics, motivations, and meanings of texts. They go beyond surface-level analysis and delve into the unconscious, possibly hidden aspects of communication. They can be applied to various types of texts and communication contexts, providing a deeper understanding of the message being conveyed and the mechanisms that shape the content. Such insight can significantly enhance the effectiveness and performance of diverse digital applications and systems that are inherently content-centric; such systems include software tools for detecting fake news on digital platforms, digital health applications for the early diagnosis of mental conditions and diseases, and software tools for deciding whether a given text is written by Artificial Intelligence (AI) programs or by humans.

Sigmund Freud envisioned the future inclusion of quantitative, interdisciplinary concepts and methods (such as those from Physics) into psychoanalysis. In a similar spirit, Jacques Lacan expressed and visualized psychoanalytic concepts through analogs from combinatorial mathematics (sets and networks) and mathematical logic (logical formulae); in particular, he thought of the language of the unconscious as a chain of quasi-mathematical inscriptions, similar to a computer language as well as a cryptographic code. In this line, it is reasonable to assume that Computer Science methods can be quite helpful in establishing solid quantitative contributions to Psychoanalysis. In particular, the use of Machine Learning and more recently of the Large Language Models (LLMs) is high relevant, given its ability, when properly used, to identify complex or hidden patterns, emotional cues, and underlying themes in human speech or writing.

Regarding the first direction mentioned above, we have already conducted promising research on automated detection of fake news. In Gadalla et al. (2023, p. 2), we investigated the incorporation of human factors and user perception in text-based interactive media, focusing on how the reliability of user texts is influenced by behavioral and emotional dimensions. In particular, we designed a Psychoanalytic approach that uses Lacanian discourse types to capture and understand the underlying characteristics of texts (news headlines) and their inherent relation to real and fake news. The approach first identifies Lacanian discourses in a text (e.g., news headlines) and then uses an algorithm to predict whether the text is real or fake based on the type of Lacanian discourses present in it. The performance evaluation demonstrated high effectiveness and accuracy for this Psychoanalysis-based prediction compared with standard methods. As far as we know, this was the first time computational methods were systematically combined with Psychoanalysis.

Following up on this significant first step, the research discussed here addresses the “inverse direction,” applying Computer Science methodologies to Psychoanalysis. In particular, the purpose is to empirically investigate (and then theoretically validate) possible fundamental relations between emotions and Lacanian discourses. J. Lacan himself highlighted the significance of a few fundamental emotions (such as anxiety and anguish, Soler, 2016, p. 18). Recent research Bucci et al. (2022, p. 165) has investigated the important role of some emotions (affects) in Lacanian theory (Lacan, 2004d,h). However, there is a need to produce a study that systematically relates emotions with corresponding Lacanian discourses. This study aims to address this gap by empirically (statistically) investigating this relation in a systematic, fine-grained manner. Such a solid understanding of the relation between emotions and Lacanian discourses is important because it will make the identification of Lacanian discourses in texts easier and more robust, since it will be based on the (easier to grasp) presence of emotions in texts.

The method includes the following key steps: (a) the emotions in a given text are identified using a well-known, fine-grained set of 30 emotions; (b) the Lacanian discourses in the text are also identified; (c) a statistical investigation is performed to identify the potentially inherent relationship between emotions and Lacanian discourses: for each discourse, which are the most characteristic emotions it admits? Which emotions exhibit the highest differentiation power in terms of corresponding discourses? (d) finally, these statistical findings are theoretically (psychoanalytically) validated.

The main contribution of this research is a psychoanalytic one per se: the establishment of a systematic relationship among emotions and Lacanian discourses, and the introduction of the concept of Lacanian Discourse Discovery (LDD), which systematizes the discovery of the Lacanian discourses. Notably, this method can be automated to a great extent, since current computer-based methods (primarily employing AI systems and tools) can effectively detect emotions in texts. Thus, there is great potential to develop effective, real-world applications based on the automated identification of emotions and their corresponding discourse. It is important to emphasize that at this stage, this study does not have any diagnostic purpose.

After this Introduction, in Section 2.1, a theoretical framework is presented that includes how emotions appear in S. Freud's and J. Lacan's works, a review of the five Lacanian discourses, a description of the Lacanian Discourse Analysis (LDA), and a review of the classification of emotions as proposed by different authors. In Section 2.2, a brief presentation of related studies is provided. In Section 3, the adopted methodology is presented in detail. The results and the corresponding discussion are presented in Section 4. Section 5 concludes the paper, indicates future directions, and discusses potential applications of the achievements of this study.

The objectives of this section are to review how emotions are treated in the research of S. Freud and J. Lacan, to review the five Lacanian discourses, to describe Lacanian Discourse Analysis (LDA), and to review the major proposed classification of emotions.

While S. Freud is universally recognized as the “father” of Psychoanalysis, it is worth briefly providing some information about J. Lacan's place in Psychoanalysis and how psychoanalysts and psychologists assess him.

J. Lacan (1901–1981) is considered one of the most influential and controversial figures in psychoanalytic theory. J. Lacan's study reinterpreted Freudian theory through the lenses of structural linguistics, topology, and philosophy, particularly drawing on Saussure, Levi-Strauss, and Hegel (Fink, 1996). J. Lacan's emphasis on language and the symbolic order marked a significant departure from the ego psychology dominant in mid-20th-century psychoanalysis (Evans, 1996, p. 162).

J. Lacan's position within psychoanalysis is both central and controversial. In France, Lacanian psychoanalysis has had and continues to have a significant institutional and clinical presence, influencing numerous schools of thought (Roudinesco, 1990, p. 25). However, in Anglo-American psychology, J. Lacan's study has often been marginalized for its abstract style and complex terminology (Leader, 2012, p. 81). His ideas have been influential in critical theory, feminist psychoanalysis, and cultural studies (Zizek, 2006).

Some psychoanalysts view J. Lacan's contributions as revitalizing the Freudian legacy and providing a richer conceptual framework for understanding subjectivity and language (Nobus, 1999, p. 102). Others criticize J. Lacan for a lack of clinical clarity or empirical support, raising concerns about the scientific rigor of Lacanian psychoanalysis (Crews, 1998, p. xxix). J. Lacan remains a major figure in psychoanalytic theory, yet his reception among psychologists and empirical researchers remains ambivalent.

In recent years, quantitative discourse analysis in psychology has evolved into a robust interdisciplinary field, integrating methods from linguistics, computer science, and cognitive science. Researchers have increasingly turned to computational and statistical approaches to analyze how emotions are expressed and structured in discourse, using tools such as sentiment analysis, topic modeling, and latent semantic analysis (Tausczik and Pennebaker, 2009; Suleman and Korkontzelos, 2021; Lai et al., 2023).

Notably, research that focuses on language use as a psychological marker, with emotional language serving as a window into affective and mental states. For example, the Linguistic Inquiry and Word Count (LIWC) framework has been widely used to associate patterns of emotional word use with personality traits, well-being, and social dynamics (Pennebaker et al., 2003; Boyd and Pennebaker, 2017). These methods enable large-scale, systematic investigations of emotional discourse across diverse contexts – from social media and therapy sessions to autobiographical narratives (Kahn et al., 2007).

Furthermore, emotion classification in text using machine learning and NLP has enabled researchers to identify nuanced affective patterns and their correlations with psychological variables (Mohammad, 2021). Studies have examined how emotional valence, arousal, and discrete emotions such as anger and sadness are reflected in linguistic structures, providing quantitative insights into affective processes (Buechel and Hahn, 2017).

Although these approaches differ substantially from the structural and symbolic perspective of Lacanian psychoanalysis, they offer valuable frameworks for empirically grounding discourse-emotion relationships.

In this section, a few studies are briefly reviewed because of: (i) their background information, which helps to understand the context in which this study has been developed; and (ii) their ideas, which could be used to fuel future research.

In Bucci et al. (2022), the authors present two independent study cases.

The first is an Emotional Text Analysis (ETA). They show that emotions expressed in language are not individual phenomena but organizers of historically determined social relations. They conducted focus groups with young graduates who were about to start their first job in multinational corporations, as well as with their corporate mentors. Using ETA, they identified two distinct clusters of emotions. One was associated with young graduates' expectations, and the other with the mentors' accounts. In this case study, there is no mention of LDA.

The second case study is an example of LDA carried out within a therapeutic community (TC) for children and teenagers diagnosed with psychosis and autism spectrum disorders. Conversations between the therapeutic group and the family group were transcribed verbatim and reviewed by a panel of researchers trained in Lacanian psychoanalysis. The family group adopts a Hysteric Discourse and demands that the therapeutic group assume the Master Discourse and provide answers and solutions to the problem at hand. The leader of the therapeutic group wisely does not fall into the trap and adopts the Analyst Discourse, opening a path for further clarification of the problem.

This case study is closely related to this research and shows the possibility of applying LDA to daily dialogues and how the Agent changes roles as the dialog evolves.

In Vanheule (2016), the author provides a thorough revision of the four classic Lacanian Discourses that goes beyond the explanations given by J. Lacan himself at the time of his writings and by other researchers who repeated what J. Lacan said in simpler terms. For example, in the case of the Master Discourse, in addition to reinforcing that it is associated with authority and power, he includes:

Next, the author delves into the Capitalist Discourse. The Capitalist Discourse is prevalent today and cannot be omitted from any study dealing with LDA and LDD. It underpins the rupture of social bonds and is the bedrock of many addictions that plague society. It has clinical importance because it must be managed differently depending on whether the patient is psychotic or neurotic.

In an important study focusing on the role of signifiers, the authors wrote in the Introduction of Olyff and Bazan (2023, p. 1):

The aforementioned work, published in 2023, is very recent and provides empirical evidence for the mental effectiveness of the signifier. It validates the empirical approach adopted by our own research and highlights the importance and possibility of identifying signifiers and the corresponding signifying chain that underpins LDA and LDD.

Robert “Rob” Haskell (1938–2010) served as a professor of psychology and department chair at the University of New England. His research areas include transfer of learning, small-group leadership, language and communication, unconscious cognition (Haskell, 2008), and analogical reasoning. He developed a novel logico-mathematical, structural methodology for the analysis and validation of sub-literal (SubLit) language and cognition (Haskell, 2003, p. 347–400).

His research, supported by several examples and a solid theoretical basis, demonstrates the presence of a hidden Truth within a discourse. The algorithm he proposed in Haskell (2003) has never been implemented and may be a valid dimension to include in future research to automatically identify signifiers and the associated signifying chain.

In this section, solid arguments are presented for choosing a particular set of emotions that can be combined and used, in light of the discovery of the Five Lacanian Discourses, and for why they are aligned with the fundamental aspects of the LDA and LDD.

In the paper “Classifying Emotion: A Developmental Account,” (Zinck and Newen 2006) propose a systematic classification of emotions that accounts for their complexity and developmental stages. They distinguish between four developmental stages of emotions:

Secondary cognitive emotions are particularly important because they depend on cultural information and personal experience, making them relevant for a nuanced analysis of emotional expression in texts. Additionally, this category is correlated with the Lacanian Symbolic order, which must be considered when approaching a text within the field of LDA (Frosh, 2013). This category also includes the emotions in the GoEmotions dataset, which was selected for this study.

The GoEmotions dataset, developed in Demszky et al. (2020), is the largest manually annotated dataset, comprising 58,009 English Reddit comments labeled for 27 emotions and Neutral. Created by researchers at Google Research, including Alan Cowen, a pioneer in emotion research, it offers a fine-grained typology adaptable to multiple downstream tasks, such as building empathetic chatbots or detecting harmful online behavior. The high quality of the annotations is demonstrated via Principal Preserved Component Analysis (PPCA), which shows reliable dissociation among the 27 emotion categories (Demszky et al., 2020).

Aligned with the classification proposed by Zinck and Newen (2006), the emotions in the GoEmotions dataset can be further categorized as follows:

The GoEmotions dataset is statistically valid, widely accepted, and suitable for a range of applications, including NLP. This extensive dataset is curated to provide a comprehensive and nuanced classification of emotions, aligning well with the developmental stages of emotions.

The robustness of the emotions in the dataset is demonstrated through various statistical analyses. For example, the dataset's annotations were found to be highly reliable, with 94% of examples having at least two raters agreeing on one label, and 31% having three or more raters in agreement. The high quality of these annotations is further validated by Principal Preserved Component Analysis (PPCA), which shows a strong dissociation among the 27 emotions. This ensures that the emotions captured in the dataset are both distinct and representative of a wide range of human emotional experiences.

Furthermore, the GoEmotions dataset has proven useful across various NLP applications. It provides a solid baseline for emotion classification models, particularly when fine-tuning models like BERT, yielding strong performance. This makes it well-suited for tasks such as building empathetic chatbots, analyzing customer feedback, and detecting harmful online behavior. The dataset's adaptability to multiple downstream tasks underscores its practical value and broad applicability.

In summary, the selection of this schema aligns perfectly with the two objectives stated in Section 2.1.5. First, the GoEmotions dataset meets the requirement that the schema be statistically and scientifically valid, not merely empirical and observational. The dataset's high reliability, demonstrated through rigorous statistical validation, ensures that it is grounded in sound human judgment and widely accepted in the community, fulfilling the first objective. Second, the selected schema is also appropriate and aligned with the concepts of LDA and LDD, as it captures the nuanced and complex emotional expressions necessary for meaningful discourse analysis in line with Lacanian principles. This ensures that this research is both methodologically sound and theoretically aligned, facilitating a deeper understanding of emotions within the framework of LDA and LDD.

To complete the set of emotions for this study, two additional emotions, “anguish” and “anxiety,” were added to the dataset because they are frequently cited in the psychoanalytic literature (see Sections 2.1.2, 2.1.1).

The current body of research lacks a comprehensive methodology for systematically identifying traces of Lacanian discourses across various modalities. While Parker (2010)'s study on LDA in interview texts provides valuable insights into identifying these discourses, it has not yet evolved into a fully developed framework for the systematic and efficient discovery of Lacanian discourses across different data types.

Due to the emerging nature of research in this field, this study focuses exclusively on textual data rather than incorporating modalities such as voice, sound, or video. Although relying solely on text may limit the exploration and identification of certain nuances inherent in Lacanian discourses, it offers a solid foundation for developing a methodological approach. Textual analysis provides a clear and structured starting point that can be refined and expanded in future research to include additional modalities. By beginning with text, the intent is to establish a robust analytical framework that can serve as a basis for more comprehensive studies in the future, introducing additional multimodalities as well.

As described in Section 2.1.3, Lacanian discourses consist of two integral components: the sender and the receiver of the discourse. To facilitate the exploration of potential emotions associated with these discourses, and given the limited prior research in this area, it is clear that textual dialogues are ideal for such examination. Dialogues reveal attributes and interpretive nuances that standalone or multi-paragraph texts do not provide.

Dialogues inherently reflect the foundational structure of Lacanian discourses, which involve communication between at least two parties conveying a message. As dialogues unfold, they reveal the latent information and intentions of the interacting parties. This dynamic flow of information makes it easier to identify the elements that constitute Lacanian discourses, i.e., the components S1, S2, $, and a.

Considering the aforementioned factors, the open-source dataset DAILY DIALOG (Li et al., 2017) has been selected, comprising everyday dialogues between two speakers. While other open-source datasets, such as Friends, could be employed, they have been written for specific purposes (e.g., screenplays intended for comedy or sarcasm) and might introduce bias and skew the study's results. Although these datasets may be useful in future research, as has been demonstrated by Joshi et al. (2016) and Poria et al. (2019), they do not align with the current research objectives.

The choice of this dataset was guided by the need for authentic, natural conversations that closely mirror real-life interactions. This authenticity is crucial for accurately analyzing the emotional and interpretive aspects of Lacanian discourses. Focusing on everyday dialogues aims to ensure that the discovered patterns and attributes are representative of genuine communication rather than artificially constructed scenarios.

To establish a theoretical correlation between Lacanian discourses and emotions, discourse and emotion annotations of dialogues were conducted using a custom-built platform. On this platform, each user was presented with a dialogue from the dataset and tasked with assigning discourses and emotions to each part of the dialogue. Along with these assignments, annotators provided the following metrics for each discourse or emotion:

Confidence score of discourse: This metric quantifies the level of certainty with which a discourse is assigned to a segment of the dialogue.

Weight of discourse: This value, ranging from 0 to 1, represents the strength or potency of the discourse within the dialogue. A higher weight indicates a stronger presence and influence of the discourse.

Confidence score of emotion: Similar to the confidence score of discourse, this metric measures the certainty of assigning a particular emotion to a segment of the dialogue.

The voting process on the platform is designed to rigorously annotate dialogues with Lacanian discourses and emotions, supported by confidence and weight metrics. This method enables a systematic analysis of how discourses and emotions are intertwined, providing a foundation for understanding their theoretical correlations. By using a structured voting system, the annotations reflect both the presence and influence of discourses and emotions, contributing to a comprehensive exploration of Lacanian discourse theory in the context of emotional analysis. This annotation process lays the groundwork for statistical modeling and in-depth analysis to uncover potential correlations and patterns.

In Appendix, Figure 1 screenshot of the platform where voters annotated the dialogues is shown.

To derive a valid mathematical formula for evaluating the results of the annotation process regarding the relationship among discourses and emotions, the following methodology was adopted.

Let S be the set of all sentences annotated. The conditional probability of occurrence of a set of n discourses d₁, d₂, …, d_n (where 1 ≤ n ≤ 5) given a set of l emotions e₁, e₂, …, e_l (where 1 ≤ l ≤ 30) in a sentence of the dataset is defined by Equation 4:

where:

The numerator in the formula is a summation over a subset S_{d1, …, dn, e1, …, el}, which includes all sentences s ∈ S that contain the discourses d₁, …, d_n and the emotions e₁, …, e_l. Within this sum, the product of the confidence scores c_s(d_i) for each discourse d_i and the confidence scores c_s(e_j) for each emotion e_j is included. This represents the aggregated confidence for instances where the discourses d₁, d₂, …, d_n and the emotions e₁, e₂, …, e_l co-occur within the dataset.

The denominator sums over a subset S_{e1, …, el}, which includes all sentences s∈S that contain the emotions e₁, …, e_l, without any requirement on co-occurring discourses in the sentences. The denominator involves a product of confidence scores, but here the set of discourses d_i may vary, as indicated by the product ∏i=1ks; where k_s is the number of discourses in the sentence s. This part of the formula represents the total aggregated confidence for all occurrences of the emotions e₁, e₂, …, e_l, regardless of the number of discourses they appear with, i.e, not only when the examined discourses d₁…d_n occur. The denominator, therefore, captures the overall likelihood of the emotions occurring within the dataset, independently of the specific discourses associated with them.

Theorem 3.1 (The probability estimator defined in Equation 4 is unbiased.). Proof. We exemplify in the case of Pr{d₁∣e₁}; the proof can be easily extended to all possible combinations of discourses and emotions.

Let p be the true, unknown probability

of occurrence of (only) discourse d₁ exclusively with emotion e₁. According to Equation 4, our probability estimator for p is:

where the numerator A sums up the total confidence on the exclusive co-occurrence of discourse d₁ with emotion e₁, while the denominator Π sums up the total confidence on the occurrence of emotion e₁; since we condition on the occurrence of emotion e₁, then Π can be considered constant (the “new” probability space created by conditioning), and the conditional probability p^ depends on the fraction of sentences when e₁ exclusively co-occurs with d₁.

We note that when (d₁, e₁) co-occur exclusively, then

because in such a case the right-hand side of Equation 6 clearly becomes equal to the left-hand side.

Such exclusive co-occurrence of (d₁, e₁) happens with probability Pr{d₁|e₁} = p. Let N = |S_e1| the number of sentences where emotion e₁ occurs and let i (1 ≤ i ≤ N) a given sentence in S_e1. Then the confidence on the desired d₁, e₁ exclusive co-occurrence is:

so

where 𝔼[·] symbolizes expectation.

The total confidence in d₁, e₁ exclusive co-occurrence across all sentences is:

so, by the linearity of expectation property, it is:

because, as explained above, Π is constant. Thus

so the estimator p^ is unbiased.

This general formula can be adapted to calculate the probability for any number of discourses n (1 ≤ n ≤ 5) given any number of emotions l (1 ≤ l ≤ 30). By extending the summation and product operations accordingly, the method remains applicable whether one is evaluating the appearance of a single discourse, a set of multiple discourses, or a combination of several emotions. The framework is flexible, allowing for the inclusion of more complex discourse-emotion relationships within the dataset, ensuring comprehensive analysis across different scenarios.

For example, the general formula of Equation 4 can be applied to any specific case involving different combinations of discourses and emotions, as shown below:

A justification for the above complex probability formulation is needed. Indeed, a simpler formulation could have been adopted, where the probability would be the fraction of co-occurrences of the examined discourse-emotion combinations over the occurrences of the considered emotions. However, such an approach could overlook crucial factors, including annotation confidence for discourse, discourse strength, and confidence in the associated emotions. An additional important factor that needs to be addressed is the “discourse weight” w_di, which captures the “strength” of the occurrence of the discourse in a sentence, i.e., how prominently the discourse occurs in the sentence. The inclusion of these three factors provides additional insight into the co-occurrence of the discourse with the emotions.

Table 1 presents the confidence scores for discourse d₁, the weight of discourse d₁, and the confidence scores for e₁ in a hypothetical dataset, where d₁ and e₁ appear together in only four sentences.

For the data in Table 1, even when confidence levels are low, the probability remains high due to the limited annotated diversity in the small dataset, leading to an almost exclusive correlation. Specifically, as e₁ only appears with d₁ the conditional probability Prob(d₁|e₁) is as follows:

In other words, even this sophisticated probability definition falls short of adequately modeling the relationship between emotions and discourses. Therefore, it is necessary to consider the strength of the discourse's presence. This is achieved by incorporating the weights of the discourses as described in Definition 3.1.

Definition 3.1 (Weight level). The weight level (W) of the co-occurrence of the discourses d₁, d₂, …, d_n with emotions e₁, e₂, …, e_l is defined as follows:

The weight level is then given by:

□

The relation R among discourses d₁, d₂, …d_n and emotions e₁, e₂, …, e_l is given in Definition 3.2.

Definition 3.2 (Relation of co-occurrence). The relation of the co-occurrence of the discourses d₁, d₂, …, d_n with emotions e₁, e₂, …, e_l is defined by Equation 8, as follows:

□

Using the data provided in Table 1, the weight level and the relation between d₁ and e₁ are calculated as

and

Interestingly, the discourse-emotion relationship now admits fine-grained values, possibly covering a very broad spectrum regardless of the dataset size, because all factors of the discourse-emotion relationship are now directly taken into account.

This result is then normalized, ensuring that the final value lies within a standard range of 0 to 1 where the maximum value is taken from all relations where the number of discourses is n and the number of emotions l. This final relation is named the relation intensity (RI) and described by Definition 3.3, for the co-occurrence of a certain combination of discourses d₁, d₂, …, d_n with certain emotions e₁, e₂, …, e_l and given by Equation 9:

Definition 3.3 (Normalized relation intensity). The normalized relation intensity among discourses d₁, d₂, …, d_n and emotions e₁, e₂, …, e_l is defined as follows:

The experimental procedure is carried out entirely on the custom-built platform. The platform ensures that voters are aware of the meaning and purpose of the discourses' confidence score and weight, as well as the emotions' confidence score. Furthermore, the platform requires that emotions be selected from the chosen dataset and that voting be possible in combinations of emotions and discourses.

The emotions and discourses assignment is essentially a subjective process. As such, it leads to disagreements and discrepancies in assignments that we prefer to label as diversity. In some cases, diversity simply means a disagreement that does not contribute to identifying any pattern. In other cases, diversity is a desirable feature of the process because it may be informative and enrich the perception of what is hidden behind the manifest narratives.

To quantitatively evaluate inter-rater agreement, we calculated agreement scores for all combinations of raters across the five Lacanian discourses using Krippendorff (2010)'s alpha coefficient. Krippendorff's alpha coefficient is a statistical measure used to quantify the agreement among multiple observers when coding or rating a set of items. When considering all three raters, the agreement scores were 0.70 for the Analyst discourse, 0.60 for the University discourse, 0.55 for both the Master and Capitalist discourses, and 0.52 for the Hysteric discourse. Examining pairwise agreement revealed more variability. The agreement between Voter 1 and Voter 2 was generally higher, reaching substantial levels for the Analyst (0.80) and University (0.71) discourses. Similarly, Voter 1 and Voter 3 showed strong agreement, particularly for the Capitalist (0.89), Analyst (0.72), and Master (0.67) discourses. The agreement between Voter 2 and Voter 3 was more moderate, with the highest score being 0.58 for the Analyst discourse.

The pseudocode for the algorithm to derive the final discourses for the sentences is Appendix Algorithm 2, and its simplified version is shown in Table 2.

Let's take, for example, the algorithm's Rule 11:

In this case, d₃ is a discrepancy that does contribute to the understanding of the enunciation.

Let us take, for example, the algorithm's Rule 18:

If the criterion were to keep only the discourses that achieved consensus, none of the discourses would have been selected. However, in this case, any discourse is selected by more than 1 voter and is informative, which makes the combination of discourses (d₁, d₂, d₃, d₄) to be taken into account.

Let us take, for example, the algorithm's Rule 22:

In this case, d₃ is a discrepancy that does contribute to the understanding of the enunciation. If the criterion were to keep only the discourses that achieved consensus, d₂ would have been discarded. However, d₂ is informative and makes the combination of discourses (d₁, d₂) to be taken into account.

It may seem that the weights assigned by individual voters are not important because they are not considered when deriving the weight assigned by the “common user.” The weights assigned by individual voters help voters reflect their perceptions. In future research, a comparison will be made between the relationships obtained by individual voters, and in this case, these weights are essential.

The algorithm for deriving the discourses for the “common user's” vote considers only up to 4 discourses, as this was the maximum number a voter assigned to a sentence. Also, the cases that led to a discourse of “none” were discarded from the final analysis of the data.

Hereafter, an example of dialog processing is shown to help with understanding the “common user's” vote evaluation.

Dialog ID: FroLatteAd_10

I can't believe Mr. Fro didn't buy it. Who does that guy think he is anyway? Bill Gates?

He had a lot of nerve telling us our ads sucked.

Time to order. Balista, today I want a skinny triple latte.

When did you start worrying about your weight?

I'm not. I just don't feel like drinking whole milk today. Why? Do you think I'm fat?

No, Jess, chill out!

The example shows that:

The algorithm to derive the emotions for the “common user's” vote is shown in the Algorithm 2.

Equation 4 is applied to the “common user's” votes, and a complete table of all conditional probabilities is generated. Table 3 shows an extract from the complete table of conditional probabilities. In any given row, the first column is a set of up to three simultaneous emotions, and the remaining columns show the conditional probabilities for the set of discourses. It may be seen that the sum of the probabilities in each row is 1. In the complete table, the sum of the probabilities in each column is also 1.

Next, Equation 9 is applied to the conditional probabilities, generating a complete table of the relation intensity between a set of emotions and a set of discourses. In any given row, the first column lists up to three simultaneous emotions, and the remaining columns show the relation intensity for all 13 discourse combinations considered.

For clarity, we symbolized the relation intensity for particular combinations of discourses with emotions

in the following way:

making it easier the identification of the involved emotions and discourses.

In general terms, this relation intensity can be represented as:

where 0 ≤ {i, j, k} ≤ 30. The value 0 is used to represent “no emotion considered” and the other values point to a specific emotion of the dataset.

{a, b, c} represent one of the five Lacanian discourses or none of them to take into account a single or a combination of discourses.

1 ≤ m ≤ 13 represents one of the 13 possible discourses combinations and 0 ≤ r_m ≤ 1. As already shown by Equation 9, r_m is obtained by the multiplication of a probability and a weight, both of which are less or equal to 1, and a further normalization. So, it is not possible to have a high value of the relation intensity if the corresponding conditional probability is low. For each set of emotions {e_i, e_j, e_k} the 13 values of r_m can be sorted in descending order, indicating the most prevalent discourses associated with that set of emotions.

The relation intensity results have been summarized in a heat map shown in Figure 2, presented as a 2 × 2 grid of subfigures. Each subfigure highlights one quadrant of the heat map, focusing on the top 5 relation intensity values for each combination of emotions. While the complete relation intensity table has 198 rows, the table limited to the top 5 values includes only 67 rows.

These heat maps show the entire dataset and are not limited to just the top 5.

It is important to point out that the findings and corresponding explanations are preliminary due to the limitations of the experimental procedure and decisions taken for the adopted methodology. Nevertheless we are confident that the qualitative findings are of significance and the quantitative ones deserve to be refined in future work.

Before showing the results, it is necessary to clarify the relevance and consequences of working with a combination of emotions instead of a single emotion for each sentence.

We used the term sentence to refer to a complete enunciation by a speaker in each part of the dialogue. We did not break each enunciation into its constituent parts. Let us hypothetically assume that a speaker provides the following enunciation:

Each voter was asked to assign emotions to the entire enunciation, which clearly expresses at least two different emotions. For the first part—I love my son above all other persons.—the emotion “love” is present. However, in the second part of the enunciation—On the other hand, I despise my ex-wife with the whole strength of my soul.—it is difficult to identify the emotion. It could be “anger”, “disapproval”, or “disappointment.”

The voters made the assignments independently of each other, and in this specific case, it is reasonable for them to cast different votes. The Common User algorithm harmonizes the discrepancies.

As the objective of this study is to establish a relationship, if one exists, between emotions and the Lacanian discourses, the combination of emotions is more representative of the Lacanian discourses adopted by the speaker and provides greater differentiation than single emotions. It is not an objective of this study to provide any diagnosis. In future studies, we will evaluate the possibility of using the Lacanian discourses as an auxiliary tool to be incorporated into diagnostic procedures.

The heat map shown in Figure 2 must be read row-wise, meaning that comparisons of values in the same row are meaningful. However, comparisons of values in different rows have qualitative meaning but not quantitative meaning. In the sequence, some examples will clarify these comments.

Out of the 67 rows of the heat map only 7 have more than one value different from zero. This result indicates that the identification of emotions has a very strong ability to identify the prevalent discourse in a sentence. The rows that have more than one relation intensity value different from zero are the following:

and,

and,

The only Master discourse is uniquely identified by a very large set of emotion combinations, 10 in total. The relation intensity is in the range:

and the emotions show a prevalence of judgement (approval, disapproval), self-image (pride), jouissance (joy), and expectation (anxiety, nervousness). This finding is consistent with S1 in the position of Agent / Semblance, in which a role of power and authority may produce statements charged with judgement and expectation. On the other hand, S1 driven by $ speaks about himself in terms of the image he/she wants to display.

The top four emotion combinations in relation to the Master discourse are:

Several other emotion combinations exhibit the same relation intensity of 0.25.

The only Analyst discourse is uniquely identified by a set of five emotion combinations. The relation intensity is in the range:

and the emotions show a prevalence of caring associated with curiosity and neutrality. This finding is consistent with a in the position of Agent/Semblance driven by S2 in a role that shows empathy and a neutral curiosity to try to access the other's Jouissance. It is also worth noting that the numerical values of RI are consistently lower than in all other cases. This reflects the fact that in daily dialogues, people seldom adopt the neutral speech of an Analyst.

The top five emotion combinations in relation to the Analyst discourse are:

The only Hysteric discourse is uniquely identified by a very large set of emotion combinations, 10 in total. The relation intensity is in the range:

and the emotions span a wide spectrum, with excitement being prevalent. This finding is consistent with $ in the position of Agent/Semblance, driven by a in a role that looks forward to gaining access to knowledge while expressing the power of personal feelings. It is worth noting that the numerical values of RI are consistently higher than in all other cases. This reflects how often, in daily life, people eagerly seek information and understanding, and also how this position is easier to be identified.

The top five emotion combinations in relation to the Hysteric discourse are:

The only University discourse is uniquely identified by a set of 8 emotions combinations. The relation intensity is in the range:

and the emotions are characterized by neutrality, realization, and judgment. This finding is consistent with S2 in the position of Agent/Semblance, driven by S1 in a role that appears restricted to dealing with objective facts while simultaneously seeking the fulfillment of institutional objectives (S2 is driven by S1) and employing judgment to achieve such results. It is worth noting that the numerical values of RI are quite similar to those obtained for the Master discourse.

The top 5 emotions combinations in relation to the University discourse are:

The only Capitalist discourse is uniquely identified by a single combination of emotions, namely (approval, desire, joy). The relation intensity is 0.20. It is likely that if the experiment is applied to a larger set of dialogues, other combinations of emotions can be found. Nevertheless, this set of emotions well represents the fact that $ in the position of the Agent/Semblance attempts to satisfy his/her internal needs not by addressing the other but by directly accessing an “artifact” that represents his/her desire, would produce joy, would have his/her approval, or would generate approval for him/herself, for example, in terms of “likes” on social media.

Many statements present a mix of Lacanian discourses. For example, the combination {Master, Hysteric} has appeared in 8 instances, and the relation intensity falls within the following range:

and the top five emotion combinations in relation to this mix of discourses are:

The combination {Hysteric, University} has appeared in 5 instances, and the relation intensity is in the range:

and the emotions combinations in relation to this mix of discourses are:

The combination {Master, University} has appeared in three instances, and the relation intensity is in the range:

and the emotional combinations in relation to this mix of discourses are:

It is also worth noting that, for example, Anger—a Basic Emotion in the GoEmotions dataset—has not appeared. At this stage, a possible explanation is that the set of dialogues used in the experiment does not include this emotion, but it may appear in a larger experiment. Another possibility is that Anger may have been annotated as Annoyance or Disapproval. Another case is Grief—a Secondary Cognitive Emotion in the GoEmotions dataset—which has not appeared as well. In this case, a possible explanation is that Grief has been annotated as Sadness.

It is left to the interested reader to examine the heat map and verify that the set of emotions associated with the combination of discourses is consistent with Lacan's theory.

As detailed in Section 4.1, our experimental procedure considers a total of 264 sentences across 40 dialogues. We will now rigorously estimate the necessary sample size (i.e., the number of sentences) to guarantee a high level of precision for the examined probability of co-occurrence of combinations of discourses with emotions.

As noted, we estimate the proportion of sentences that satisfy a given property; for example, we examine the case in which discourse d₁ occurs together with emotion e₁. The necessary sample size n to guarantee a certain precision is given by

where Z the z-score for the desired confidence level, p the expected (unknown) proportion and E the desired error margin.

For a worst-case sample size estimate, we assume the unknown probability p equals 1/2, so p(1 − p) is maximized. At a 90% confidence level (which we consider quite adequate for this challenging research setting, which inherently includes high subjectivity), Z = 1.645. For an error margin E = 0.05, we have:

Thus, our sample size of 264 sentences is very close to the required size of 270 sentences, making it adequate for the purpose of our analysis. In fact, for 264 sentences, the margin of error is E = 0.0506.

For additional experimental evidence in Section 24, we present the detailed findings of a meticulous stability analysis via simulation.

To complement the above analysis, we performed a very detailed experimental procedure.