Amuleto by Roberto Bolaño, a comparative stylistic analysis with computational tools.

Amuleto by Roberto Bolaño, a comparative stylistic analysis with computational tools. Stefano Rinaudo Language Technologies and Digital Humanities, Università di Torino, 1166089 Abstract This paper employs computational tools to substantiate qualitative interpretations of Roberto Bolaño’s Amuleto, comparing it to Nocturno de Chile by the same author and Relato de un náufrago by Gabriel García Márquez. Firstly, the efficacy of displaying the 100 Most Frequent Words (MfWs) in a WordCloud for capturing the main themes of the novels is investigated. Findings suggest that WordClouds are effective for gaining a general sense of a book's content when the author’s style is more standard. Secondly, the study applies Type-Token Ratio (TTR) and Standardized TTR (STTR) metrics to determine if the stylistic redundancy highlighted in Amuleto adversely affects lexical variety. Results contradict the hypothesis, prompting a discussion on the utility of STTR in literary analysis and proposing further studies on repetition distribution in the novel. Thirdly, the paper presents data that highlight a specific stylistic feature in the two Bolaño novels: the frequent repetition of proper nouns related to people and places. An analysis of Named Entity (NE) frequency and distribution reveals that tokens labeled as persons (PER) and locations (LOC) are significantly more frequent in Amuleto and Nocturno de Chile compared to Relato de un náufrago, marking a distinctive narrative style. Keywords Stylometric Analysis, Computational Literature, Roberto Bolaño. Introduction The use of computational tools for the quantitative analysis of a text is a research practice that many academics in literary studies still view with suspicion. The general feeling is that computers, even when empowered by AI, struggle to embrace the subjective and multifaceted shades of Literature. Therefore, in a literature textbook it’s very rare to find an in-depth study on the data-driven analysis of a novel, whereas text interpretations are almost always based on the classical principle of authority. Nevertheless, I am convinced that computational tools can represent an innovative analytical resource for literary studies, and humanists cannot afford to ignore innovation. Shutting ourselves in an ivory tower invoking the unreachable deity of Literature risks leaving us unharmed of conceptual frameworks to humanistically deal with new and ever-growing technologies. Thus, in this paper I will follow a methodological approach aimed at creating a dialogue between computational tools and classical literary criticism. In fact, the goal of this essay is to assess whether it’s possible to provide quantitative support to qualitative interpretations of the novel Amuleto by Roberto Bolaño (1999). The choice of this book stems from a specific feature that characterizes the text: the autodiégétique narrator (Genette, 1972) in the person of Auxilio Lacouture. In fact, the style of the delusional narrating voice strays from a plain use of the Spanish language making it difficult to describe even for experienced human readers. The complexity of the text interpretation is a challenge both for critics and machines and it thus represents the ideal opportunity for hermeneutic cooperation. In particular, in this essay I’ll try to provide quantitative data to support two features of the novel’s style described in some main papers on Amuleto (Amaro Castro, 2010; Alvarez, 2012; O’Bryen, 2011; Vallejos, 2022; Villalobos-Guízar, 2022): the repetition of the delusional narrating voice and its constant need to remember places, dates and people calling them by name. This research will be conducted using computational tools to perform a textual analysis of the novel, mostly at a lexical level, through 3 different tasks. I’ll analyze the Most Frequent Words (MFW) and I’ll display them in a WordCloud to see if these tools can be useful to describe the general meaning of the text. Then, computing Type Token Ratio (TTR) and Standardized TTR (STTR), I’ll try to assess if the repetition of the narrating voice perceived by the critics can be retraced in this measure of lexical variety. Finally, I’ll analyze the results of Name Entity Recognition (NER) task performed by NameTag 1 tool on Amuleto to check if the hypothesis of Auxilio’s repetition of places, dates and people can be supported by this data. Since I need a benchmark to compare, understand and discuss this data, I’ll perform the same tests on two other short novels narrated at the first person by the protagonist of the story: Nocturno de Chile by Roberto Bolaño (1999) and Relato de un náufrago by Gabriel García Márquez (1955). The first book was chosen because of its proximity to Amuleto in the style and in the content: in fact, as highlighted by the critics, in both novels « the writer deploys a discourse on power, memory and history in Latin America, proposing as his point of view and discursive strategy the marginal and hallucinated stance of his narrators» (Amaro Castro, 2010. p. 147). Marquez’s short novel was chosen mostly because of its similar length, testimonial content and narrative structure and, on the other hand, because of its less hallucinated and more standard style that makes it a significant benchmark for individuating some peculiar stylistic features of Bolaño’s books. Data In this section, I’ll briefly summarize the content of the three novels. Then, I’ll provide a table with the data obtained by processing the texts in their original Spanish version with Profiling UD 2 . Amuleto is narrated by Auxilio Lacouture, a Chilean poet who finds herself trapped in a women's bathroom at the Universidad Nacional Autónoma de México (UNAM) during the political turmoil in 1968. As she hides, the story unfolds through her redundant and contradictory memories and reflections on her past and future experiences, her relationships with other poets and intellectuals, and the impact of political oppression on art and life. To describe the novel with the author’s words, Amuleto is «Un juego formal y una despedida a los jóvenes que nacieron en los cincuenta y que, como yo, quisieron cambiar el mundo y perdieron» (Rodriguez, 2000, p. 58). Nocturno de Chile is the deathbed memories and confessions of the narrator, Father Urrutia. Persistently hallucinatory and defensive, the story ranges from Opus Dei and intellectual milieu description to falconry and private lessons on Marxism for Pinochet and his generals directed at the unspecified reproaches of a joven envejecido (wizened youth), probably representing a morally intact version of the narrator. Relato de un náufrago recounts the real-life experience of Luis Alejandro Velasco, a Colombian naval officer who survives a shipwreck and is stranded on a deserted Caribbean Island for ten days. Velasco describes the events leading up to the shipwreck, the physical and psychological challenges he faces while stranded, and his struggle for survival. As he grapples with isolation, hunger, and desperation, Velasco reflects on his life, human resilience, and the nature of existence. The following table contains the number of sentences and tokens of the three short novels together with the average number of tokens per sentence. I provide this data to quantitatively understand the dimensions and some features of the corpora. 1 Link to NameTag: https://lindat.mff.cuni.cz/services/nametag/ 2 Link to UDPipe online demo: http://www.italianlp.it/demo/profiling-ud/ To get more information about the tool see Brunato et alii (2020). Table 1. Corpora dimensions Number of sentences Number of tokens Tokens per sentence Amuleto 1720 45433 26.41453488372093 Nocturno de Chile 4523 44988 9.94649568870219 Relato de un náufrago 4161 35021 8.416486421533286 As we can see, the dimensions of the first two short novels are very similar in terms of total number of tokens, whereas Relato de un náufrago is 22,92% shorter than Amuleto and 22,15% shorter than Nocturno de Chile. Another important difference between the three books is found in the average length of the sentences that are significantly longer in Amuleto. This figure can be interpreted as an indicator of the less written and more dialogical and free style of the writer that sometimes results in a long and non-fragmented stream of consciousness of Auxilio 3 . Experiments 1. Getting a general understanding: Most Frequent Words (MFW) and Word-Cloud Representation Before delving into stylometric analysis, I tried to process the three novels by extracting the 100 most frequent words of each, after having lowercased the texts and removed grammatical and semantically poor words 4 . To make their visualization clearer, I chose to represent the MFWs of each novel in a Word-Cloud. Figure 1. Word-Cloud representation of the 100 MFWs Amuleto Nocturno de Chile Relato de un náufrago If the content of Relato de un náufrago is immediately understandable by looking at the biggest words of the WordCloud (balsa, agua, mar, dias), for Bolaño’s novels this kind of representation is way less talkative. In both pictures, the names of some characters and places are particularly embossed (México, Remedios, Lilian, Auxilio, Arturito, Ernesto, Facultad in Amuleto; Chile, Santiago, Sordello, Farewell, Neruda in Nocturno de Chile), letting us hypothesize a particular style feature of the writer that we’ll analyze in the third chapter of the Experiments section of this essay. In Amuleto we can find in little the word madre, particularly interesting for the understanding of the novel 5 , whereas in Nocturno de Chile the word padre provides us information 3 Even if these figures resonate with the reader’s perception, we must remember that the average is a measure of central tendency very sensitive to outliers. Therefore, some very long sentences in Amuleto could excessively influence these results. 4 Link to the Python script I wrote to get the 100 MFWs and the WordClouds: https://colab.research.google.com/drive/1Xklp_WFoUnLJxhb4p-snwYt8pruu97LG?usp=sharing 5 The narrator defines herself as the «madre de la poesía mexicana» multiple times in the novel and the word «madre» appears 35 times. about social status of Father Urrutia, the narrator. But these and other considerations can be done just after having read the two novels: in fact, this kind of visualizations remain poorly informative about the content of Bolaño’s prose, suggesting that his particularly complex style makes the description of his novels difficult if we limit our analysis to the MFWs visualization. 2. Repetitive style and lexical variety: Type Token Ratio (TTR) and Standardized Type Token Ratio (STTR) Since the critics agreed on a certain tendence to repetition and the redundancy in the speech of Auxilio Lacouture 6 , I hypothesized that this particular stylistic feature could result in a negative impact on the lexical variety of the novel. The heterogeneity of the words in the text can be computed through the Type Token Ratio (TTR) metric which refers to the ratio between the number of lexical types and the number of tokens and ranges from 0 (absence of lexical variety) to 1 (no type repetition in the text). I therefore decided to test the null hypothesis stating that there is no significant difference among the TTR figures of the three novels of the corpus. Firstly, I looked at the TTR figures calculated by ProfilingUD that I report in table 2. This online tool computes the metric «for text samples of equivalent length: TTR for both the first 100 and 200 tokens of a text» (Brunato et alii, 2020, p. 7147). Table 2. ProfilingUD TTR TTR[100] TTR[200] Amuleto 0.56 0.535 Nocturno de Chile 0.67 0.595 Relato de un náufrago 0.69 0.65 From these preliminary results, we notice that Amuleto’s style is less lexically varied than both Nocturno de Chile and Relato de un náufago. Although these figures seem to confirm our hypothesis, they are limited to the incipits of the novels, accounting on the lexical variety of 0,22% of the book in the worst case (TTR[100] of Amuleto, the longest text) and of 0,57% of the book in the best case (TTR[200] of Relato de un náufrago, the shortest text). For this reason, I decided to compute the Standardized Type Token Ratio [STTR], a metric that contrasts the TTR sensitivity to sample size dividing the entire text into subsections of 1000 tokens (in this case), calculating the TTR for each subsection, summing up the TTR values and dividing the result by the number of subsections 7 . Before computing the STTR, I filtered and lowercased the text. This preprocessing operation together with the use of a different tokenizer (the NLTK one in my Python script) can explain the different number of total tokens found in Table 1 and in Table 3. This proceeding allowed me to have figures accounting on the lexical variety of 99,06% of Amuleto, 98,30% of Nocturno de Chile and 99,12% of Relato de un náufrago. Table 3. STTR STTR Number of subsections Number of tokens Amuleto 0.4204102564102564 39 39370 Nocturno de Chile 0.4426842105263158 38 38657 Relato de un náufrago 0.41893333333333327 30 30266 6 «La narradora menciona irónicamente, en la primera página de su relato, que a ella le enseñaron que las redundancias sobran y que solo debe bastar con el argumento. Sin embargo, ella no hará caso a esta enseñanza y estructurará su discurso en digresiones, repeticiones y dudas». Villalobos-Guízar, 2022, p. 131. 7 Link to the Python script I wrote to compute STTR: https://colab.research.google.com/drive/1Q1F17eBXfNHymPrKozNVozCU_lqNZv5B Comparing the figures of Table 2 and Table 3, we notice that the results of STTR for the three novels are much more similar to each other differing of a maximum of 3 centesimal values. Also, in Table 3, the less lexically varied novel is Relato de un náufrago whose both TTR[100] and TTR[200] scores where the highest in Table 2. Thus, we must conclude that there is no significant difference among the STTR figures of the three novels and therefore the null hypothesis can’t be disproved. These quantitative results contrast the opinion of the critics and lead us to two possible considerations. The first one is that the TTR and STTR metrics are not representative of the degree of repetitiveness of a novel and thus we should look for a more functional metric. The second one is that, starting from these quantitative results, critics could rethink and re-elaborate their considerations on the repetitions of Auxilio’s narration, maybe looking at the distribution of this stylistic feature in the text. In both cases, research in this direction would result in a better understanding of the novel and in an improvement of the metrics used for computational text interpretation. 3. Looking for landmarks in space and social networks: an analysis on Name Entities (NE) In Vallejos’s (2022) interpretation of the novel, Auxilio’s speech is related to the concept of literary testimony developed by Felman and Laub (1992). In this framework, the obsession of the narrator is characterized by repeated mentions of proper names, dates and places that could be interpreted as a way of giving credibility to the story by clinging to reassuring and stable landmarks. Therefore, in Amuleto this stylistic feature represents Auxilio’s reaction to the UNAM occupation trauma. But the same narrative mechanism can be also found in Nocturno de Chile, where the constant repetition of proper names and places can be interpreted as a reassuring way for Father Urrutia of staying on top of his sense of guilt, of dominating the joven envejecido. In fact, for both the narrators these landmarks represent a sort of safe haven to return to when they feel threatened or confused. To account for this stylistic trait in both Amuleto and Nocturno de Chile, I decided to study the frequency and the distribution of Name Entities (NE) in the three novels. In this case, the null hypothesis that I tried to falsify states that there’s no significant difference in the NE frequency and distribution among the three books. To automatically perform the Name-Entity Recognition (NER) task, I decided to upload, one at the time, the raw texts in their txt format into UDPipe 8 and process them with the UD Spanish AnCora (2.15) model. I then downloaded the tagged, lemmatized and parsed outputs in CoNLL-U format and processed them with another online tool developed by the Universal Dependencies project and designed for NE annotation: NameTag 9 . I chose the NameTag 3 Multilingual (250203) model for NER and I downloaded the output in the vertical csv format containing the retrieved NE only 10 . Subsequently, I aggregated the data of the vertical format counting the frequency for each NE category, namely Persons (PER), Locations (LOC), Organizations (ORG) and Miscellanea (MISC). After having repeated this process for all the three novels, I created Table 4 containing all these figures together with the total number of NE detected and the number of single tokens labeled as NE 11 . In fact, a single NE can be formed by more than one token: for example, México Distrito Federal is a single NE formed by three tokens. 8 UDPipe: https://lindat.mff.cuni.cz/services/udpipe/ 9 NameTag: https://lindat.mff.cuni.cz/services/nametag/ 10 For more information on the output formats: https://ufal.mff.cuni.cz/nametag/1/users-manual#run_ner_output_formats 11 Link to the Python code I wrote to aggregate and count this data: https://colab.research.google.com/drive/1tiKI0P9kkP2-U7sz0iMh9yFTYnzFTS6 Table 4. NE aggregated raw frequencies Amuleto Nocturno de Chile Relato de un náufrago PER 833 960 161 LOC 355 343 148 ORG 42 41 21 MISC 61 130 32 TOT NE 1291 1474 362 TOT NE TOKENS 1886 1964 573 Having the total number of single tokens labeled as NE and the total number of tokens of each book, I computed the percentage of NE tokens on all the tokens of each book to assess if there’s a significant difference in the proportion of NE tokens in the three books. The results of Table 5 show that, as highlighted by the critics, in Bolaño’s novels the narrators tend to mention more frequently proper names of people, places and organizations. Table 5. Percentage of NE tokens on the total tokens in each book Percentage of NE / total tokens Amuleto 4,15% Nocturno de Chile 4,36% Relato de un náufrago 1,64% I then looked at the internal distribution of NE for each novel. In fact, the critics mostly focus on names and places repetition in Auxilio and Father Urrutia’s speeches, whereas the NE labeled as organizations and miscellaneous aren’t considered in literary analysis, even if sometimes the tokens contained in these categories can be intended as a spatial reference or can be mislabeled. For example, looking at the NE’s list of Amuleto, the acronym DF (standing for Distrito Federal, the former name of Ciudad de México) is very often labeled as MISC or ORG. Also, in the same list the protagonist’s name, Auxilio, whose literal meaning is “help”, is sometimes mislabeled as MISC. Table 5. Normalized data of the NE token distribution for each label Amuleto Nocturno de Chile Relato de un náufrago PER 64,53% 65,13% 44,48% LOC 27,50% 23,27% 40,88% ORG 3,25% 2,78% 5,80% MISC 4,72% 8,82% 8,84% TOT 100% 100% 100% Nevertheless, the percentages contained in Table 6 show that in the three novels, the ORG and MISC labels together represent a maximum of 14,64% in Relato de un náufrago and a minimum of 7,97% in Amuleto of all the NES, highlighting the preponderance of proper names of people and places. Also, the NE distribution is almost identical in Amuleto and Nocturno de Chile with ~65% of PER and ~25% of LOC for both the book. On the other hand, in Maquez’s short novel the figures show another trend with 44,48% of PER and a very high 40,88% of LOC. This difference can be explained by both the particular style of Bolaño (always prone to cite names of other authors) and the content of Relato de un náufrago where the narrator, stranded on a deserted Caribbean Island, navigates the surrounding geography for giving spatial references to himself and the reader. To have a better visual understanding of this data distribution, I plotted them into a stacked bar chart 12 . Figure 2. Distribution of NEs across the three novels Discussion and conclusion In conclusion, in this paper I have used computational tools to support or discuss with quantitative data some qualitative interpretations on Roberto Bolano’s short novel Amuleto, using as a benchmark two other novels: Nocturno de Chile by the same author and Relato de un náufrago by Gabriel García Márquez. The first experiment consisted of assessing the accuracy of the WordCloud representation of the 100 MFWs in conveying the general meaning of the novels. The results prompted me to hypothesize that this tool can be very useful for getting a global idea of the content of a book if the author’s style is more standard and the narrative voice is not delusional. Secondly, I used the TTR and STTR metrics to assess if the stylistic redundancy highlighted by the critics in Amuleto have a negative impact on these measures of lexical variety. The results did not support my hypothesis, opening a discussion on the significance of computing the STTR of a literary text and on the possible future studies on the distribution of repetitions in Amuleto. Finally, I provided quantitative data supporting the relevance of a particular stylistic trait in Bolaño’s novels consisting of the repetition of people and places proper nouns. In fact, I studied the NE frequency and distribution across the three books showing that the number of NE tokens labeled as PER and LOC is significantly higher in Amuleto and Nocturno de Chile with respect to Relato de un náufrago. The main limitation to my work was the absence of parameters assessing the statistical significance of my considerations. For example, when discussing the results of STTR measurements on the three novels, I concluded that the difference between the figures of each book wasn’t relevant, but I couldn’t find any statistical metrics to support this point. I hope that future research can help me identify suitable metrics for this purpose. 12 Link to the Python code I wrote to plot the stacked bar chart: https://colab.research.google.com/drive/1DNm_G8WdO2r0sPbyhsH3ndiuFPVD0znj?usp=sharing References Novels • Bolaño, Roberto (1999, ed. 2007). Amuleto, Barcelona: Editorial Anagrama S.A. • Bolaño, Roberto (1999, ed. 2000). Nocturno de Chile, Barcelona: Editorial Anagrama S.A. • Márquez, Gabriel García (1955, ed. 2014). Relato de un náufrago, Barcelona: Debolsillo. Papers • Álvarez, Moira (2012). La voz de Auxilio en Amuleto de Roberto Bolaño, in «Revista de Crítica Literaria Latinoamericana», vol. XXXVIII, n. 75, pp. 419-440. • Amaro Castro, Lorena (2010). Delirio y margen como estrategias discursivas en dos narraciones de Roberto Bolaño, in «Atenea», n. 501, pp. 147-156. • Brunato, Dominique et alii (2020). Profiling–UD: a Tool for Linguistic Profiling of Texts, in «Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020)», pp. 7145-7151. • Felman, Shoshana et Laub, Dori (1992). Testimony: Crises of Witnessing in Literature, Psychoanalysis, and History. New York: Routledge. • Genette, Gérard (1972). Figures III, Paris: Seuil. • O’Bryen, Rory (2011). Memory, Melancholia and Political Transition in Amuleto and Nocturno de Chile by Roberto Bolaño, in «Bulletin of Latin American Research», vol. XXX, n. 4, pp. 473-487. • Rodríguez Villouta, Mili (2000). Roberto Bolaño: La historia de la historia, in «Mensaje» vol. XLIX, pp. 58-59. • Vallejos, Yasmina (2022). Amuleto y el poder del testimonio en el discurso de Auxilio Lacouture, in «Revista Iberoamericana», vol. LXXXVIII, n. 278, pp. 109-126. • Villalobos-Guízar, Valeria (2022). La voz de Auxilio: las potencialidades disidentes de la narración oral en Amuleto de Roberto Bolaño, in «Revista Iberoamericana», vol. LXXXVIII, n. 278, pp.127-143. Online tools and code • Shared folder containing the Python functions, the three novels in txt format, and other relevant material: https://drive.google.com/drive/folders/1jm14gDDsrP2kV1QCOrAWCtE6GFc25Th?usp=sharing • UDPipe: https://lindat.mff.cuni.cz/services/udpipe/ • ProfilingUD: http://www.italianlp.it/demo/profiling-ud/ • NameTag: https://lindat.mff.cuni.cz/services/nametag/