Era- and Genre-Specific Stop Word Lists for Low-Resource Computational Research: A Classical Latin Exemplum

2.1.1 Consideration & evaluation of existing Latin stop lists

In recent years, most Latin stop lists have been based on word frequency. By creating stop lists based exclusively on the frequency of words within an entire corpus of extant (digitized) literature, scholars may inadvertently remove words that are less common and/or more significant in certain eras, thereby erasing essential content from their analyses.

This is particularly true in the context of Latin literature given that a significant proportion of extant text comes from a Late Antique or Christian context. Over the past several years, the Classical Language Tool Kit (CLTK, Johnson et al., 2021) has offered several stop lists based on different methods of computing word frequency, including overall frequency, entropy, and variance.¹ These stop lists demonstrate disproportionate influence from Late Antique linguistic contexts due to the large amount of extant text from these time periods. On the other hand, the Perseus Digital Library (PDL; Crane, accessed 2024) and the International Organization for Standardization (ISO, Diaz 2016) offer stop lists that suffer from brevity and inconsistency. A summary of previous lists is presented in Table 2; the script used to create this table is available in the GitHub repository. Our revised stop list is available below in Table 3.

2.1.2 Removal of stop words with high semantic relevance to target corpus

We begin with the combined list of words from Table 2 and methodically remove and add words to create a targeted Latin stop list for poetry from the Augustan era. The words we remove can be sorted into three categories.

First, we remove words with a disproportionately high frequency in Late Antique and ecclesiastical Latin relative to Classical Latin. In most cases, these are references to or titles for the Christian god (e.g. dominus, deus, or rex). The word dominus is so common in Late Antique texts that it is the 53^rd most common word in all Latin literature.² In an ecclesiastical context, dominus means ‘lord [god]’. In a Classical Latin context, however, the word dominus (here, ‘master’ or [slave] ‘owner’) is much less frequent and carries more importance as a semantic feature for conceptual representation (as opposed to, e.g., the formulaic uses in ecclesiastical Latin). Removing dominus from the text of a Classical poet could constitute the erasure of slave narratives and essential literary constructs (servitium amoris, or love as slavery, Copley 1947; Lyne 1979). Similar considerations should be made for words like rex and deus.

Second, we remove words that are overrepresented in particular genres. Another inclusion on previous stop lists is the noun bellum (‘war’), which often occurs in historiographic texts like those of Julius Caesar. While bellum is a valid inclusion for stop lists targeted at historiography, as researchers creating a targeted stop list for Augustan poetry, we elected to remove it. Like slavery, war features in a metaphor of love as a military endeavor (militia amoris, Murgatroyd 1975). Another problem with previous stop word lists is their inconsistent inclusion of prepositions and pronouns. For example, some stop word lists include per (a common preposition meaning “through”), but not trans (a preposition meaning “across”). In our list, we correct these inconsistencies to include all pronouns and prepositions of similar semantic significance.

Third, we remove words that do not appear in Classical texts. This includes abbreviations such as od (a shortening for oculus dexter, ‘right eye’) and ed (possibly a shortened form of the pronoun idem, ‘the same’, or the verb edo, ‘eat’). Finally, we remove words that are redundant, inconsistent, or illogical inclusions. For example, nummus (‘coin’) has an overall frequency rank of 1560; the reason for its inclusion in any stop list is unclear. Other words (me, sit, omnes) are specific forms that would be lemmatized to their dictionary forms ego (‘I’), sum (‘be’), and omnis (‘all’).

2.1.3 Addition of stop words with low semantic relevance to target corpus

To correct inconsistencies in existing lists, we include all pronouns except nemo (‘no one’). We did not include nemo as a stop word because its usage in the nominative (a grammatical case in Latin) tends to be more pointed than that of other pronouns, which carry less and equivalent semantic significance. The generalized use in other cases reverts to forms of nullus (‘nothing’), which is included on the list. We also add words with low semantic weight and high frequency in Classical Latin. For example, facio (‘make/do’), do (‘give’), magnus (‘big/great’), multus (‘much’), and fero (‘carry’).³

2.1.4 Creation of Python script for easy implementation of specialized list

To make it easy for researchers to use our targeted stop lists, we provide a command-line based Python script (latin_stop_list.py) that can be implemented into an existing research pipeline.