Linguistic Value Construction in 18th-Century London Auction Advertisements: a Quantitative Approach Alessandra De Mulder1,∗ , Lauren Fonteyn2 and Mike Kestemont1 1 University of Antwerp, Antwerp, Belgium 2 Leiden University, Leiden, the Netherlands Abstract Georgian England was characterised by a buzzing consumer society in which advertising played a pro- gressively important role when it came to the (linguistic) value construction surrounding material goods. Increasingly, the perceived value of goods was not only determined by the intrinsic quality of the goods, but also by the socio-commercial discourse used to characterise them. Linguistic modi昀椀ers, such as ad- jectives, must have played an important role in this process – re昀氀ecting these socio-economical trends in text while also reinforcing them. Here, we focus on a diachronic corpus of over 5,000 pages of Lon- don auction advertisement pages, digitised via automated transcription and divided across four sample periods between 1742-1829. Prime methodological challenges include: (1) the noisiness of the available data because of imperfect transcription; (2) the coarseness of the available time stamps, and (3) the lack of suitable NLP so昀琀ware, such as lemmatizers or (shallow) syntactic parsers. Through the use of word embeddings, we try to alleviate the issue of spelling variation with reasonable success. We 昀椀nd that, over time, subjective or ‘evaluative’ modi昀椀ers have become more prominent in these advertise- ments than their objective or ‘descriptive’ counterparts – but there are di昀昀erent temporal patterns for di昀昀erent types of advertised objects Keywords advertisements, linguistic modi昀椀cation, frequentist statistics, spelling normalisation, time series 1. Introduction In Georgian England, a certain group, known as the beau monde, procured their place in so- ciety by publicly demonstrating their ties to one another, both in personal relationships and material expressions [17].1 They are considered the trailblazers of a new, fast-paced consumer society, although volumes have been written discussing the time, place and pace of consumer (r)evolution(s) [24, 36]. In any case, the interaction between people and their possessions al- tered on an unprecedented scale in eighteenth-century England. This translated into many CHR 2022: Computational Humanities Research Conference, December 12 – 14, 2022, Antwerp, Belgium ∗ Corresponding author. £ alessandra.demulder@uantwerpen.be (A. De Mulder); l.fonteyn@hum.leidenuniv.nl (L. Fonteyn); mike.kestemont@uantwerpen.be (M. Kestemont) ȉ 0000-0002-2612-420X (A. De Mulder); 0000-0001-5706-8418 (L. Fonteyn); 0000-0003-3590-693X (M. Kestemont) © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). CEUR Workshop CEUR Workshop Proceedings (CEUR-WS.org) Proceedings http://ceur-ws.org ISSN 1613-0073 1 All code and data that were used for the preparation of this paper are available from the following repository under a CC-BY-NC-SA license: https://doi.org/10.5281/zenodo.7252695. 92 changes in terms of how people behaved in and thought about the empire of things they lived in. Language immediately followed – or might even have preceded – these shi昀琀s; a whole new repertory emerged to name the new material world and to distinguish between quality and rub- bish, luxuries and necessities, and between what was fashionable and what was not [40, 19]. This was especially important when assessing the intrinsic value, i.e. material, and extrinsic value, i.e. design of objects. Previous research has shown the rising importance of design rather than material when acquiring, for example, silverware, a household staple that had lost many functions due to chi- naware’s rising popularity [7]. It makes sense to favour design over material durability when buying goods meant for display rather than intensive usage. However, this does not explain both the popularity and centrality of design as the decisive factor in an inherent breakable and yet daily used household good such as chinaware. This may be because the value of objects went beyond their external features: much value can be derived from how it was used by par- ticular owners, known in certain circles for their good taste [17, 16]. The well-heeled middling sorts would have invested in an extensive assortment of patterned and customised chinaware to serve tea and other refreshments in their well-furnished matching drawing room to their peers. The guests would have noted the decorative lexicon on their hosts’ porcelain and at- tributed value according to the beau monde’s unwritten rules to structure their society. The hypothetical hosts would have done as well when picking out china patterns which had already been con昀椀rmed as genteel by peers or opting for – slightly – diverging motifs emblazoned with their family crest to reinforce their own position as ‘tastemakers’. The cultural value of these objects played a crucial role in setting a price, even when their vanguard, previous owners passed, and their goods were auctioned o昀昀 to settle any remaining bills. Research into the ever-growing language of consumption has already shown that the emer- gence of advertising, in general, played a crucial role. Newspaper content, including adverts, created an imagined community of all stakeholders in the world of goods, from producer to seller to consumer [4]. Unlike nowadays, the language in advertisements re昀氀ected consumer sensitivities in this imagined community rather than instilling in readers what they should consume [22]. The language in notices became increasingly specialised [39], communicating about and valuing luxuries in order to consume them appropriately (i.e., so that they would be recognised and appreciated by peers). This implied that advertisers tapped into a vocabulary directly linked to dominant consumer values. Sellers had limited space to convey what set their wares apart from others on the market; for new products, this meant introducing what they did and why one needed to buy it. Second-hand household goods up for auction, however, needed little to no introduction, which allowed auctioneers to focus on the unique qualities of the goods: intrinsic material, certain fashionable 昀椀nishings or the previous owner, who would have been recognised as a tastemaker by peers [39]. This reveals the dominant consumer val- ues of the time, much like how buzzwords such as sustainable and organic have become an indispensable part of the present-day advertising vocabulary. Hence auction advertisements were more than dry lists of goods and some practical information; regarding them as such would not give them the credit they deserve and the unique insights into changing consumer sensitivities they bring. As opposed to present-day publicity, the tone of eighteenth-century adverts was above all, polite. Criticism of false promises in advertisements almost exclusively applied to those for 93 new products, which is also quite plausible since one did not have to explain to anyone what a chair was for, unlike the latest silver bullet for the “French disease” [42, 13]. On top of that, a good reputation was a seller’s most prized possession; it ensured long-term client relationships and, thus, their livelihood. The notices tended to focus on the professionalism and knowledge of the retailer and were mainly informative to underline this to potential customers [9, 22]. In turn, potential buyers put their best foot forward to show that they could pay, on the one hand, and had enough knowledge to value the goods, on the other. Tapping into the appropri- ate language repertoire, from the initial contact through an advertisement to any subsequent interactions between buyer and seller, was the ideal way to assure the other party of your good intentions [11, 12, 39]. Previous studies have already exploited the potential of these rich sources through keyword searches but were faced with regrettably poorly OCRed databases or resorted to limited manual sampling [18, 38]. This manual approach is time-consuming and allows most of these studies to cover only one or, at best, a few core concepts used to explain the typical consumer. We apply digital text analysis to – still very poorly – OCR’ed auction advertisements to contribute insights into eighteenth-century buyers’ buying habits. Research hypotheses This paper uses a bottom-up corpus linguistic approach over +5,000 pages of auction advertisements pages to examine the relationship between the objects (noun phrases) and their descriptors (modi昀椀ers) in the adverts through four sample periods between 1742-1829. We speci昀椀cally examine the modi昀椀cation process because modi昀椀ers such as adjec- tives are an important way to express appreciation or value assignment in language. Below, we shall describe how we applied a categorisation scheme to a selection of modi昀椀ers and objects. Following the historical research described above, we advance the following hypotheses: • h1: We expect clear diachronic shi昀琀s in the prominence of certain objects. Real estate will have become more frequent, for instance, due to the emerging practice of selling houses together with their household goods;2 . • h2: The presence of modi昀椀cation, taken as a proxy for linguistic value construction, will have increased over time; • h3: The use of subjective or ‘evaluative’ modi昀椀ers will have increased relatively more strongly over time in comparison to their objective or ‘descriptive’ counterparts; • h4: Instances of (pre)modi昀椀cation will likely have become more complex (i.e. longer) over time through the process of ‘modi昀椀er stacking’. 2. Materials 2.1. Origin This paper’s source material mainly consists of advertisements from the Daily Advertiser and the Morning Chronicle. These are supplemented by pages from other London newspapers, which also contained auction advertisements and were available for the same period. The used auction advertisements were manually selected from 5 newspapers in the Burney Collection 2 This was due to rising duties on speci昀椀c goods and changing market functioning because of an oversupply a昀琀er the French Revolution amongst other reasons, which made solely selling one type of good less pro昀椀table[21] 94 and the British Newspaper Archive across four sample periods: 1742-1743 (SP1), 1773 (SP2), 1799-1800 (SP3) and 1828-1829 (SP4). Pages contain one to 51 advertisements, and the corpus comprises ∼9.6M tokens (∼640K of which are unique). Table 1 Overview of advertisement page counts across newspapers and advertisement pages for all sample periods. Abbreviations: DA (Daily Advertiser), MC (Morning Chronicle), DP (Daily Post), LDP (London Daily Post) & GA (London Daily Post & General Advertiser), PL (Public Ledger). *= only 1743 included. Sample period / newspapers DA MC DP LDP & GA PL Row total SP1 1742–1743 (Burney) 583 513 260* 1,356 SP2 1773 (Burney) 763 323 1,086 SP3 1799–1800 (Burney) 630 778 1,408 SP4 1828–1829 (BNA) 625 701 1,326 Column total 1,976 1,726 513 260 701 5,176 Figure 1: Morning Chronicle 03/07/1800 with an excerpt of the automatic transcription Figure 1 shows a typical auction notice where “elegant furniture” that belonged to a gentle- man was o昀昀ered for sale by Mr Postan in the Morning Chronicle. The advertisement started and ended with practical information such as the time and place of the sale and where further information like catalogues could be found. The actual listing of the goods always occurred according to a rigid scheme, usually starting with a formulaic sentence such as “elegant house- hold furniture, plate”, followed by what were considered the showstoppers of the upcoming auction, “昀椀ne linen, an eight-day chime table clock by Mitchell, 昀椀ne toned harpsichord by Kirk- man”. This could be followed with information about the previous owner, as is the case here, the goods belonged to a gentleman from Richmond. The main course of every auction adver- tisement was the listing of other goods o昀昀ered for sale and their descriptions; we can read that 95 Table 2 Indicative number of tokens (and unique tokens) per newspaper (to be multiplied by 100K). DA MC DP LDP & GA PL Row total SP1 6.30(0,41) 1.4(0.12) 0.59(0,07)* 8.03 (0.59) SP2 23.43(1.15) 2.86(0.25) 26.29(1.40) SP3 10.38(0.79) 28.23(1.65) 38.60(2.43) SP4 17.31(2.74) 12.07(1.29) 29.38(4.03) Total 40.10(2.35) 48.39(4.63) 1.14(0.12) 0.59(0.06) 12.07(1.29) 102.30(8.45) “capital dome bedsteads, with elegant chintz cotton furnitures; mahogany four-post and tent bedsteads, with rich cotton furnitures, three sets of French window curtains to correspond; prime goose feather beds and bedding” are going under the hammer at 11 o’clock the next day. We collected the data by manually browsing through all available digital issues of seven volumes of the newspapers mentioned in the tables above (Tables 1 and 2). We selected the pages with advertisements from each issue and manually cropped them, so they only contained auction advertisements. These were quite straightforward to single out as all auction notices started with “for sale by auction”, “to be sold by hand”, “to be sold by the candle”, or other markers such as mentioning “auction room”, “catalogues”, etc., as the example above clari昀椀es. This was done in the Burney Collection for the 昀椀rst three sample periods, 1742-1800, and in the British Newspaper Archive for the last period, namely 1828-1829. Sample periods were chosen based on the availability of the core newspapers, Daily Adver- tiser and Morning Chronicle, to have some consistent data to account for newspaper-speci昀椀c variations. We opted for three sample periods in the eighteenth century and one at the begin- ning of the nineteenth to pinpoint which developments can be traced back to the eighteenth century. It is well attested that this was a crucial century for newspapers due to many changes in printing techniques, amongst other developments, and we can see this manifesting in a rise of advertisements before the true boom in the nineteenth century [32]. The 5,176 advertisement pages were automatically transcribed into (unstructured) plain text using Transkribus with a model trained for eighteenth-century printed English texts. This CITLAB HTR+ model was trained on 13,083 words and reached a character error rate (CER) of 0.15 % on the training set and 1.89 % on the validation set. The base model was an HTR model called “French18thCPrint” [1], which was trained on 38,487 words with a CER of 0.09 % on the training set and 0.74 % on the validation set. However, the data proved extremely noisy, as there was no segmentation between individual adverts, no distinction between < �㕠 >/< Ā > due to marginal di昀昀erence between a printed < �㕠 > and < Ā > as shown in Figure 2, and spotty layout analysis. A manual check of the 583 pages of the Daily Advertiser in 1742-1743 brought to light that there were more signi昀椀cant layout issues (e.g., erroneous column detection) in 8% of the pages and 316 instances of more minor layout issues (e.g., one or two lines in the wrong order). 96 Figure 2: Extract from Daily Advertiser (12/10/1742), highlighting examples of s/f similarities. 2.2. Modifiers and objects: classification In our analysis, we distinguish between ‘objects’, i.e. material goods that were put on sale in the auctions represented in the corpus, and ‘premodi昀椀ers’, any elements that can occur before objects to describe – or ‘modify’ – them.3 Put simply, the phrase colourful glass shards, we can identify one object (i.e. shards) and two premodi昀椀ers (i.e. colourful and glass). We designed and applied a custom classi昀椀cation scheme for both premodi昀椀ers and objects. First, for the premodi昀椀ers, we manually constructed a list of 282 commonly occurring premodi昀椀ers in the material. Each of these premodi昀椀ers was identi昀椀ed through a unique headword (type) in a normalised, modern spelling, but collapsing the di昀昀erence between the characters < �㕠 > and < Ā >, e昀昀ectively treating them as allographs. Two annotators then independently categorised the premodi昀椀ers along as either ‘evaluative’ (E) or ‘descriptive’ (D). If the premodifying element functioned as a classi昀椀er (e.g. kidderminster carpet is a type or class of carpet) or an “objective epithet” (e.g. large, enamelled, woollen), whereas “subjective epithets” (e.g. antique, fashion- able, valuable) were marked as evaluative [10]. 4 In other words, modi昀椀ers that either do or do 3 While this study focuses on premodi昀椀cation; it should be noted that not all modi昀椀cation is premodi昀椀cation. Still, premodi昀椀cation is expected to be more common in light of the language-external context of advertising. The price of placing an advertisement was based on conventions of “moderate length”, which ties in with what we see in the source material: there was little to no length variation throughout the research period [32]. This obviously had implications on the word choices that advertisers made. They had to convey what goods were for sale, why they were worth buying and when and where this was all happening, all without exceeding the customary word limit. This is presumably why Mr Postan (Figure 1) opted to print “satin-wood and mahogany drawing-room and parlour chairs” and “elegant pier and pembroke tables” instead of using postmodi昀椀cation, as in “drawing-room and parlour chairs made of satin-wood and mahogany” and “pier and pembroke tables that are elegantly fashioned”. 4 Note that in some linguistic literature, the objectivity-subjectivity of epithets is considered a cline: while ‘good- looking’ is a clear example of a property assigned to people or objects in a subjective manner, properties such as ‘clean’ are “more collectively assessable”. Still, they are not as objectively assessable as properties like ‘blue’ or ‘wooden’ [10]. For simplicity’s sake, we chose to work with a binary annotation system where any degree of subjectivity was classi昀椀ed as evaluative. 97 Table 3 Initial instances of confusion between the two annotators in the binary modifier classification (282 headwords in total): Evaluative (E) vs Descriptive (D). D E D 167 4 E 17 94 not apply to an object –- for example, a table is either made of mahogany, or it is not –- were considered descriptive. This o昀琀en applies, for instance, to words referring to textiles, such as chintz and serges, which appeared as modi昀椀ers for upholstered furniture. Furthermore, words such as looking, which only appeared in 昀椀xed combinations such as looking glass, are consid- ered descriptive because they specify the type of object. Evaluative modi昀椀ers, by contrast, are modi昀椀ers that express a particular, more subjective evaluation of an object. Unlike descriptive modi昀椀ers, they cannot be de昀椀ned through a straightforward yes or no question: whether or not a chair is elegant is a matter of personal taste and di昀케cult to prove objectively. Similarly, it is di昀케cult or even impossible to provide an objective de昀椀nition that helps determine whether an object is commodious or valuable is true or false. A昀琀er manual annotation by two annotators, we applied the established metric Cohen’s �㔅 [8] in the scikit-learn reference implementation [31] to estimate the inter-annotator agreement. The resulting �㔅 statistic is a scalar that ranges between -1 and 1: larger positive values imply a strong agreement, but values closer to zero (or negative scores) mean that the agreement might be due to chance [2]. For the binary modi昀椀er classi昀椀cation, we obtained �㔅=.8407, indicating a “strong agreement” [20] between the two annotators, highlighting the relative straightforward- ness of this task. In Table 3, we report the initial disagreement between the two annotators in the form of a confusion matrix. As is clear from the confusion matrix, there was relatively more disagreement regarding the assignment of the evaluative class label, which arguably makes sense from an interpretive perspective. During group discussions in the adjudication phase, the annotators resolved instances of disagreement and agreed on a single label for each head- word. Ultimately, this yielded the following (somewhat skewed) distribution: 182 descriptive and 100 evaluative modi昀椀ers. These are listed in the appendix. Second, we repeated a similar procedure for a set of relevant and common objects of modi昀椀- cations in the data, i.e. goods mentioned in the advertisements. We started from a set of 139 ob- jects that were categorised into nine categories: ‘NA’ (for Not Applicable5 ), ‘clothing/fabrics’, ‘furniture’, ‘appliances/utensils’, ‘tableware’, ‘animal accessories’, ‘haberdashery’, ‘animal’, ‘in- strument’, ‘accessories’, ‘decoration’, and ‘real estate’. This categorisation system was based on prior work in consumption history and aimed for reasonably balanced groups in the object set. Categorising objects is a widespread practice in material culture studies, which are usu- ally centred around the function of objects in order to gauge the use, practices and meaning behind certain (clusters of) objects. Common categories for the eighteenth-century consumer society are, for instance, kitchenware, tableware, seating and table furniture, hot-drink-related 5 NA was assigned when the noun phrase was not an object, a word like ’assortment’, a street name or a nonsense word. 98 Table 4 Initial instances of confusion between the two annotators in the 12-class object classification (139 head- words in total). appliances/utensils animal accessories clothing/fabrics haberdashery instrument accessories decoration real estate tableware furniture animal NA accessories 5 0 0 0 0 0 0 0 0 0 0 0 animal 0 2 0 0 0 0 0 0 0 0 0 0 animal accessories 0 0 2 0 0 0 0 0 0 0 0 0 appliances/utensils 0 0 0 8 0 0 0 0 0 0 0 0 clothing/fabrics 0 0 0 0 5 0 1 1 0 0 1 0 decoration 1 0 0 3 0 11 0 0 0 2 1 0 furniture 0 0 0 0 0 0 24 0 0 1 0 0 haberdashery 0 0 0 0 0 0 0 2 0 0 0 0 instrument 0 0 0 0 0 0 0 0 4 0 0 0 NA 1 0 0 2 0 0 0 0 0 10 1 0 real estate 0 0 0 2 0 0 2 0 0 1 40 0 tableware 0 0 0 0 0 0 0 0 1 0 0 5 objects, interior decoration, sleeping furniture, and so on [35, 34]. However, all of these studies usually rely on object-speci昀椀c narratives, while we opted for the most general object category possible. This allowed us to bring broader trends to light. The objects were categorised by two annotators using the historical thesaurus of the English Oxford Dictionary for object de昀椀nitions and prede昀椀ned simpli昀椀ed object classi昀椀cations of Overton, Weatherill and Muldrew, amongst others [29, 43, 27]. In case of doubt (such as books), we looked at how they were described and assigned the most similar category. Books tended to be closely described like other ob- jects in the decoration category, which makes sense since they o昀琀en were collected and richly (re)bound to be shown o昀昀 [37]. The headwords were again independently categorised by two annotators. In this case, the test yielded �㔅=.8182, indicating a lower but still substantial agreement, especially when con- sidering up to 12 categories. The confusion matrix in Table 4 displays the initial disagreement between the annotators. In an adjudication phase, cases of disagreement were again resolved. The resulting distribution of object labels was as follows (a昀琀er excluding the NA class): real es- tate (46), furniture (25), decoration (18), clothing/fabric (10), appliances/utensils (9), tableware (6), accessories (5), instrument (4), animal/accessories (4). To enable the pragmatic approach outlined below, we made sure that there was no overlap between the sets of object and modi昀椀er headwords; ambiguous cases were not included. The objects are also listed in the appendix. 2.3. Orthographic normalisation The materials under scrutiny contained a considerable amount of orthographic variation, which is partially due to naturally occurring historical spelling variants. The bulk of varia- 99 tion is, however, caused by artefacts from the imperfect digitisation procedure (see above for the automated transcription procedure followed). An example of a noisy single ad entry from the corpus is shown in the illustration above (Figure 2). This noise impedes the straightforward identi昀椀cation of the headwords from the modi昀椀er and object sets described in the previous sec- tion. In spite of English’s current dominance in the world of natural language processing, we are working with a historical language variant, which generally tends to be under-resourced. We decided to apply a pragmatic normalisation routine based on word embeddings in com- bination with straightforward string distance heuristics. FastText word embeddings [6] o昀昀er an advantage over the earlier generation of embeddings, such as Word2vec [25], in that they do not represent words as an atomic, symbolic index but that they take into account a to- ken’s subword information, in the form of the token’s constituent n-grams. This makes them more 昀氀exible in the face of out-of-vocabulary items (in which our material can be expected to abound). FastText word embeddings could therefore o昀昀er a baseline for identifying super昀椀cial spelling variants (and in昀氀exions) of the objects’ headwords and modi昀椀ers we aim to study.6 We 昀椀rst applied a lightweight naive preprocessing to the material, restoring hyphenated words at line endings, removing all non-alphabetic characters and splitting the entire corpus into space-free character strings or tokens. We divided the materials into 100,762 segments of 100 consecutive tokens we fed as sentences to a reference implementation of the FastText algorithm in Gensim [33]. We considered the entire vocabulary of tokens and trained a model with a dimensionality of 500 for 20 epochs. To recognise a new token as a potential instance of one of our headwords (objects and modi昀椀ers), we would use the FastText model to retrieve all the headwords with a cosine distance < .3 from the new token. If this set was non-empty, we suggest the headword at the minimal Levenshtein distance from the new token as a spelling normalisation replacement for the input token. Note that this routine is naive because it oper- ates at the type level and considers no contextual features.7 We evaluated the e昀昀ectiveness of this spelling normalisation procedure on a manually cor- rected sample of advertisements (amounting to 3,545 tokens in total), which contained ran- dom selections from all four periods in the corpus. Note that we collapse all instances of non- replacement into a single class for this purpose (‘NA’). Also, because of the transcription rou- tine adopted, we treated the characters < �㕠 > and < Ā > as allographs and mapped all instances of < �㕠 > to < Ā >. Tables 5 and 6 show a random sample of examples of concrete replacements, both for correct and incorrect interventions. The conventional classi昀椀cation statistics are re- ported in Table 7. We achieve an encouraging macro-averaged F1 score (0.877) in the upper eighties while maintaining a reasonable balance between precision (0.907) and recall (0.862). For instance, the random sample of incorrect replacements mainly concerns unusually noisy spelling variants for which the correct headword was not properly retrieved (‘NA’). That preci- sion is higher than recall seems acceptable for our application that involves su昀케cient material: in this context, we favour the correct replacement of spelling variants over missing a few oth- ers. We applied this spelling normalisation to the entire corpus before proceeding. Before this 6 For an example of a study showing the potential of FastText and Word2vec Skipgram embeddings in identifying di昀昀erent types of spelling variants, see [28]. 7 For this reason, our method may be better suited for mid- to high frequency (rather than low-frequency) words and spelling variants, as the quality of word type vectors is a昀昀ected by frequency. 100 Table 5 Random sample of correct replacements (out-of-context tokens). token gold silver houfhold houfehold houfehold defks defk defk rtanding 昀琀anding 昀琀anding botles bottlef bottlef handkerchiefs handkerchief handkerchief serges fergef fergef houfhold houfehold houfehold fachionable fafhionable fafhionable aluable valuable valuable mualin muflin muflin Table 6 Random sample of incorrect replacements (out-of-context tokens); ‘NA’s are dominant, indicating that the correct headword could not be identified. token gold silver finale fingle NA pramttes premifef NA idengs refidence NA irigh irifh NA righ rich NA dameatle dome昀琀ic NA atxiroxar 昀琀aircafe NA houfe houfehold NA aparinus fpaciouf NA darillay dwelling NA Table 7 Conventional classification metrics for the spelling normalisation procedure on a manually corrected sample of 3,545 tokens. accuracy 0.980254 F1 (macro) 0.877163 precision 0.907394 recall 0.862167 procedure, the four time periods amounted to unique 661,117 word types, distributed over 10,076,512 token instances – meaning that each type occurred on average about 15.24 times in the corpus. A昀琀er the normalisation, the number of types was reduced to 521,516. In the nor- malised corpus, the average type would therefore have a token frequency of 19.32, suggesting that we were indeed able to reduce the orthographic noise in the material. 101 Table 8 Tabular overview of the mean scores for the normalised frequencies per category across SP1-SP4, in- cluding the average �㔏 score. era SP1 SP2 SP3 SP4 �㔏 category Desc 0.214359 -0.406092 0.026862 0.164870 0.076923 Eval -0.461974 -0.219575 0.462137 0.219411 0.285795 3. Analysis 3.1. Time series for modifiers and objects To test our main research hypotheses, we calculated the relative frequency of each item from our modi昀椀er set for each of the four time periods in the data. We normalised the per-headword frequencies via standard scaling (mean subtraction and dividing by unit variance) to ensure that more common items would not dominate the analysis. We produce boxplots for the two categories of modi昀椀ers per time period. Based on Figures 3 and 4, we see that there is a consider- able diachronic variance over time in both the evaluative and descriptive categories.8 However, the evaluative set of modi昀椀ers seems to display a more consistent increase in frequency over time than the descriptive modi昀椀ers, which show a pronounced drop in the second time period, with surprisingly many outliers (SP2). Figure 3: Per-category boxplots of �㔏 scores across SP1-SP4 for modifiers in the Descriptive category. 8 For a more detailed view on whether temporal e昀昀ects di昀昀er depending on the type of object being advertised, we refer to Figure 7 below. 102 Figure 4: Per-category boxplots of �㔏 scores across SP1-SP4 for modifiers in the Evaluative category. 1.00 0.75 per category 0.50 0.25 0.00 Distribution of 0.25 0.50 0.75 1.00 Desc Eval Object category Figure 5: Per-category boxplots of �㔏 scores across SP1-SP4 for modifiers in the Desc(riptive) and Eval(uative) category. The central tendency is markedly higher for the evaluative modifiers. To obtain a more quantitatively informed assessment of this situation, we calculated the per- word Kendall �㔏 statistic for each individual time series: this is a non-parametric rank correlation coe昀케cient test that can be used the measure the consistency in the increase or decrease of a scalar over a series of time points. With only four time points, however, it is impossible to 103 reach statistical signi昀椀cance, so we focus on the �㔏 primarily as a metric. This statistic will output a scalar in the [-1, +1] range, with large positive values indicating a strong increase, large negative values indicating a strong decrease and values relatively closer to 0 indicating a stable result (neither increase nor decrease). The resulting distribution of �㔏 values is shown in each category in (Figure 5). The average �㔏 statistic (cf. Table 8) is positive for both categories (�㔏 =0.076 for descriptive, �㔏 =0.285 for evaluative), indicating that linguistic premodi昀椀cation grew relatively more impor- tant over time in this material. However, the range of values seems notably higher for the evaluative modi昀椀ers indicating that these gained even more prominence over time. To com- pare the range of �㔏 values in the two categories (which are both not normally distributed), we applied a non-parametric Mann-Whitney �㕈 test to verify the (directed) hypothesis that the eval- uative �㔏 -values are indeed larger overall than the descriptive ones. For this dataset (�㕛þÿ�㕠ý = 182; �㕛ÿ�㕣�㕎�㕙 =100), the resulting test statistic (�㕈 =7189.5) and the associated �㕝-value (�㕝=0.001) o昀昀er rea- sonable grounds not to reject this hypothesis. We conclude that premodi昀椀cation grew more important over time, and this was relatively more true for the set of evaluative modi昀椀ers con- sidered here than their descriptive counterparts. We applied analogous measurements to the frequency of normalised tokens in the object categories we distinguished and recorded their scaled relative frequencies across the sample periods SP1-SP4. Next, we calculated Kendall’s �㔏 for the time series associated with each nor- malised token and aggregated them at the category level. Figure 6 shows the distribution of the �㔏 ’s across these categories. The category ‘clothing/fabric’ is the only one to display a negative trend, indicating that its prominence in the advertisements decreased over time. Next, the me- dian presence of ‘appliances/utensils’, ‘decoration’, and ‘furniture’ remained stable. The rest of the categories have a mean �㔏 that is more outspokenly positive; goods belonging to the fol- lowing categories have clearly gained prominence over time: ‘animal/accessories’, ‘tableware’, ‘instruments’, and ‘real estate’. 3.2. Complexity in modification We now turn to h4, the hypothesis related to ‘modi昀椀er stacking’: did instances of premodi昀椀ca- tion become longer in this material and thus gained complexity over time? Because we lack a proper syntactic parse of the data, we adopted a pragmatic approach and collected all instances of an object mention that was directly preceded by an uninterrupted series of modi昀椀ers (i.e. se- ries of at least one modi昀椀er in front of an object). We gauged their length across the four time periods and presented a number of relevant summary statistics in Table 9. A clear majority of the premodi昀椀cation instances has a length of one, and longer series are much less common. Interestingly, we found no clear diachronic trend regarding these length measurements, and we must therefore refute h4 for the time being. This suggests that the increase in evaluative modi昀椀ers is not due to a simple stacking process, where evaluative modi昀椀ers simply supple- ment descriptive modi昀椀ers. Rather, these numbers suggest that the descriptive modi昀椀ers were more likely replaced altogether by more evaluative ones. 104 0.0004 Median frequency 0.0002 0.0000 0.00075 Mean frequency 0.00050 0.00025 0.00000 1 Distribution of 0 1 clothing/fabric furniture appliances/utensils decoration animal/accessories tableware accessories instrument real estate Figure 6: From top to bottom: (1) median frequency and (2) mean frequency of objects in each category per time period (le昀琀 to right: SP1-SP4); (3) Per-category boxplots of �㔏 scores across for objects in several categories. 3.3. Interaction between time and object categories Above, we presented preliminary evidence that the use of evaluative modi昀椀ers would have increased over time in the corpus. However, this trend will likely show a di昀昀erent development across the various categories of objects we distinguish in the corpus. To verify this hypothesis, we extracted a subset of the corpus containing all instances of modi昀椀ers immediately preceding a headword from one of the object categories. This makes it likely that the modi昀椀er was reigned by the adjacent object.9 For each instance in the subset, we record the original token of the modi昀椀er, the normalised headword, the category of the ensuing object and its headword. This resulted in a data set of 168,417 modi昀椀cation instances, of which a random sample is presented in Table 10. This subset contains a much simpli昀椀ed and partial view of the data set but apart 9 Note that, at present, reliable deep or shallow parsers are not readily available for this period – though progress may be underway [30]. 105 Table 9 Length (in tokens) of premodification instances across time: mean length, standard deviation, and proportion of instances of length 1 (f1), length 2 (f2), etc. None of these summary statistics display an obvious trend. mean std prop(f1) prop(f2) prop(f3) prop(f4) period SP1 1.377 0.676 0.712 0.219 0.054 0.013 SP2 1.360 0.617 0.699 0.255 0.035 0.009 SP3 1.469 0.778 0.663 0.238 0.073 0.017 SP4 1.348 0.643 0.726 0.216 0.046 0.009 Table 10 Random sample of instances from the premodification subset extracted, consisting of each modifier occurring immediately preceding a headword from one of the object categories. Our binomial model predicts whether a modifier will be evaluative as the dependent target variable. token normalised period head-cat head-word mod handfonie handsome 2 real estate apartment E houfhold household 0 furniture furniture D capital capital 1 real estate mansion E excellent excellent 3 furniture press E arable arable 3 real estate land E from the high precision of this extraction process, the measurements in the previous section demonstrate that we nevertheless cover a large portion of all modi昀椀cation instances. Using a conventional Generalised Linear Model (of the binomial family), we apply a statis- tical model that aims to predict whether the modi昀椀er will be evaluative (i.e., we applied the coding: 0 = descriptive; 1 = evaluative). For each instance, we have two predictors available as main predictors: time period (0-3, encoded as an ordered, categorical factor) and the type of ob- ject it modi昀椀es (categorical factor). We ran three versions of the base model, each of increasing complexity.10 We compare the variants of the model using their AIC score and Akaike Weights (i.e., conditional probabilities which indicate how much statistical importance we should attach to di昀昀erences in AIC values [41]; for an application in Linguistics, see [15]); additionally, mod- els are compared using the Anova test. This information is presented in Tab. 11. We considered adding token-level random e昀昀ects to this simple model, i.e. 昀椀tting random intercepts to the (1) normalised headword for the modi昀椀ers and (2) the objects for each instance. However, for (1), this would not be insightful because the modi昀椀er headword perfectly predicts the dependent outcome variable (evaluative vs descriptive). With (2), this extension proved di昀케cult because this random e昀昀ect had to be included as a nested variable since the headword of the object per- fectly predicts the category of the object included in the interaction term. The resulting model frequently failed to converge properly, arguably because the contribution of the random e昀昀ect towards the optimisation objective was ultimately too small. Following the concerns formu- lated by [3], we did not pursue this mixed e昀昀ect approach further. 10 None of the models displayed overdispersal. 106 Table 11 Comparison of three variants of the binomial model, showing the model index, its underlying formula (R syntax), the AIC and Akaike Weights, as well as the significance of the pairwise Anova (with the likelihood ratio test) for the comparison with the previous model. model formula AIC Weight Anova (LRT) 1 mod ∼ period 207983.6 0.0 0.0 2 mod ∼ period + head.cat 198723.6 0.0 �㕝 < 2.2ÿ − 16 *** 3 mod ∼ period * head.cat 196284.7 1.0 �㕝 < 2.2ÿ − 16 *** We note that the subsequent models always yield a better 昀椀t of the data, as indicated by the decreasing AIC scores. The weights in Table 11 indicate a very high probability (of 1.0) that model 3 outperforms models 1 and 2 (both of which were assigned a weight of 0.0). Each subse- quent model is, moreover, invariably a signi昀椀cant improvement over the previous, according to the Anova test. The simplest model (1), where the time period is the sole predictor, predicts a solid increase in the probability of encountering an evaluative modi昀椀er. This was probably to be expected based on the previous section, but the additional experiments show that this sim- ple view is naive. Adding the object type of the modi昀椀er’s head (model 2), for instance, yields a better 昀椀t of the data in which, surprisingly, the e昀昀ect of time reverses: this indicates that the evaluative trend probably played out di昀昀erently in di昀昀erent modi昀椀cation contexts. Model 3 concretely models this interplay as a statistical interaction between time period and object category: the interaction model proves to be an improvement concerning the additive model, urging us to include the interaction and disregard the e昀昀ect of the individual predictors. The previous paragraph contains important insights: apparently, the object categories in- vited di昀昀erent levels of evaluativeness in their modi昀椀cation, and this association can only be properly modelled by taking the interaction with the period properly into account. In Figure 7, we plot the e昀昀ect of time period conditional on the object category of the modi昀椀er for the most complex model 3. This lattice plot shows that the increase in evaluative modi昀椀ers has been relatively stronger for speci昀椀c object categories, such as accessories, clothing/fabric and instruments. In other categories, this evolution was less outspoken (e.g. decoration) or even negative, such as for real estate and tableware. This suggests that linguistic developments in consumer trends showed di昀昀erent rationales in speci昀椀c market segments. One caveat, how- ever, is that under our approach, many multiword units in the real estate and tableware cate- gory would be of the type glass bottles, where we treated glass as a descriptive modi昀椀er and bottles as the headword. This might explain why we should expect no surge in the use of more evaluative modi昀椀ers in this position, and why some object categories are more strongly asso- ciated with, e.g. descriptive modi昀椀ers. At the same time, this situation also does not explain why one should see such a consistent drop in the use of evaluative modi昀椀ers. 4. Discussion Object categories such as ‘decoration’, ‘furniture’ and ‘utensils/appliances’ were continually o昀昀ered for sale in the adverts because their socio-cultural worth (i.e. the previous owner) 107 period predictor effect plot 0 1 2 3 head.cat = accessories head.cat = anim/acce head.cat = appl/uten 0.5 0.4 0.3 0.2 0.1 head.cat = clot/fabr head.cat = decoration head.cat = furniture 0.5 0.4 mod 0.3 0.2 0.1 head.cat = instrument head.cat = real estate head.cat = tableware 0.5 0.4 0.3 0.2 0.1 0 1 2 3 0 1 2 3 period Figure 7: E昀昀ect plot (resulting from Model 3) for time as a predictor, conditional on the category of the object headword. remained more stable [17]. A more trend-sensitive object category such as ‘fabric/clothing’ declined throughout the research period, most likely due to rapid changes in fashion on the one hand and competitive access to upholsterers, which allowed buyers to customise their purchases to their liking and interior [23, 26]. Next, we look at the rising object categories, namely ‘animal/accessories’, ‘real estate’, ‘tableware’ and ‘instruments’. We can explain the 昀椀rst two based on our hypothesis that houses were increasingly sold with household e昀昀ects and gardens, including stables and their inhabitants, which is clearly re昀氀ected in the source material [21]. The remarkable surge in ‘tableware’ and ‘instruments’ is likely due to their status as elite markers; some genuine chinaware and a harpsichord clearly indicated gentility. The auctioneer thus placed these goods prominently in his adverts to assure potential buyers of the quality and prestige of the furnishings o昀昀ered for sale. Besides, these goods became simply a much more widespread commodity throughout (elite) society throughout the eighteenth century [23, 21, 108 14]. It is not surprising that changes in modi昀椀er usage are object-speci昀椀c, i.e. descriptions which highlight the material properties remained more important for structural goods such as real estate and fragile goods such as tableware where materiality de昀椀ned functions, e.g. silver for cold- and porcelain for warm drinks [5]. Other furnishings such as accessories, clothing/fabric, instruments and - to a lesser extent - decoration, where a signi昀椀cant part of the allure lay in their outward appearance, were increasingly evaluative. These goods were material culture items that peers noticed in a drawing room and where design mattered the most to the well- heeled middling sorts. In the end, we draw the following conclusions. Regarding h1, we indeed see that some ob- ject categories become more prominent throughout the research period (‘animal/accessories’, ‘tableware’, ‘instruments’, and ‘real estate’), mainly due to the growing practice of selling houses with their household goods, stables etc as well as fashion changes. When it comes to h2, we see that modi昀椀cation was on the rise overall; we even 昀椀nd con昀椀rmation for h3, stating that evaluative modi昀椀ers grew relatively more frequent over time than descriptive modi昀椀ers. Finally, we could not 昀椀nd evidence for the hypothesis regarding modi昀椀er stacking (h4): this indicates that evaluative modi昀椀ers must have replaced descriptive modi昀椀ers across the board (instead of supplementing them). These results tie in with the rising importance of design in literature, with noted exceptions in some categories. This was precisely the case when the evaluative modi昀椀ers were employed to describe object categories whose value relied on their unique properties and design to emphasise their worth and thus construct their value linguis- tically. The adverts ultimately served to draw potential buyers to the viewing and auction where consumers could gauge the true value of the goods by feeling, smelling and interacting with the objects o昀昀ered for sale beyond the linguistic constructions that were examined in this paper. Acknowledgments The authors would like to thank Bruno Blondé, Enrique Manjavacas, Wout Saelens, Bas Spliet, the anonymous reviewers for their help and comments and the Research Council of the Uni- versity of Antwerp for supporting the project: Fashioning ’old and new’. Secondary markets, commodity value conventions and the dawn of consumer societies in Western Europe (18th - 19th centuries). References [1] E. H. project (National Library Netherlands). French 18th Century Print. 2019. url: https: //readcoop.eu/model/french-18th-century-print/. [2] R. Artstein and M. Poesio. “Inter-Coder Agreement for Computational Linguistics”. In: Computational Linguistics 34.4 (2008), pp. 555–596. [3] D. Bates, R. Kliegl, S. Vasishth, and H. Baayen. Parsimonious Mixed Models. 2018. doi: 10.48550/arxiv.1506.04967. url: https://arxiv.org/abs/1506.04967. 109 [4] M. Berg. Luxury and Pleasure in Eighteenth-Century Britain. Oxford and New York: Ox- ford University Press, 2005. [5] B. Blondé. “Tableware and Changing Consumer Patterns: Dynamics of Material Culture in Antwerp, 17th - 18th centuries”. In: Majolica and glass from Italy to Antwerp and Be- yond: the Transfer of Technology in the 16th - early 17th Century. Ed. by J. Veeckman, S. Jennings, C. Dumortier, D. Whitehouse, and F. Verhaeghe. Antwerpen: Stad Antwerpen, 2002, pp. 295–311. [6] P. Bojanowski, E. Grave, A. Joulin, and T. Mikolov. “Enriching Word Vectors with Sub- word Information”. In: arXiv preprint arXiv:1607.04606 (2016). [7] H. Cli昀昀ord. “A Commerce with Things: The Value of Precious Metalwork in Early Mod- ern England”. In: Consumers and Luxury: Consumer Culture in Europe, 1650–1850. Ed. by M. Berg and H. Cli昀昀ord. Manchester: Manchester University Press, 1999, pp. 147–168. [8] J. Cohen. “A Coe昀케cient of Agreement for Nominal Scales”. In: Educational and Psycho- logical Measurement 20.1 (1960), pp. 37–46. [9] N. Cox and K. Dannehl. Perceptions of Retailing in Early Modern England. Farnham: Ash- gate Publishing, 2007. [10] K. Davidse and T. Breban. “A Cognitive-Functional Approach to the Order of Adjectives in the English Noun Phrase”. In: Linguistics 57.2 (2019), pp. 327–371. doi: doi:10.1515/li ng-2019-0003. url: https://doi.org/10.1515/ling-2019-0003. [11] D. Davis. A History of Shopping. London: Routledge, 2007. [12] R. Diaz-Bone. “The Methodological Standpoint of the ‘Économie des Conventions’”. In: Historical Social Research / Historische Sozialforschung 36.4 (2011), pp. 43–63. doi: https: //doi.org/10.12759/hsr.36.2011.4.43-63. [13] B. Elliot. A History of English Advertising. London: Business Publications Limited, 1962. [14] C. Fairchilds. “The Production and Marketing of Populuxe Goods in Eighteenth-Century Paris”. In: Consumption and the World of Goods. Ed. by J. Brewer and R. Porter. London and New York: Routledge, 1993, pp. 228–248. [15] L. Fonteyn. “Assessing Theory with Practice: an Evaluation of Two Aspectual-Semantic Classi昀椀cation Models of Gerundive Nominalizations”. In: Corpus Linguistics and Linguis- tic Theory 16.2 (2020), pp. 275–302. doi: doi:10.1515/cllt-2017-0057. url: https://doi.org /10.1515/cllt-2017-0057. [16] V. de Graiza. Irresistible Empire. America’s Advance through Twentieth-Century Europe. Boston: Harvard University Press, 2006. url: https://www.hup.harvard.edu/catalog.php ?isbn=9780674022348. [17] H. Greig. The Beau Monde. Fashionable Society in Georgian London. Oxford: Oxford Uni- versity Press, 2013. [18] E. Hart. “A British Atlantic World of Advertising? Colonial American ‘For Sale’ Notices in Comparative Context”. In: American Periodicals 24.2 (2014), pp. 110–127. doi: https: //doi.org/10.1353/amp.2014.0014. 110 [19] M. Kwass. “Between Words and Things: ‘La Querelle du luxe’ in the Eighteenth Century”. In: Mln 130.4 (2015), pp. 771–782. doi: https://doi.org/10.1353/mln.2015.0065. [20] J. R. Landis and G. G. Koch. “The Measurement of Observer Agreement for Categorical Data”. In: Biometrics 33.1 (1977), pp. 159–174. [21] B. Learmount. A History of the Auction. London: Barnard & Learmount, 1985. [22] D. Lyna and I. Van Damme. “A Strategy of Seduction? The Role of Commercial Adver- tisements in the Eighteenth-Century Retailing Business of Antwerp”. In: Business History 51.1 (2009), pp. 100–121. doi: 10.1080/00076790802604475. [23] A. McCants. “Porcelain for the Poor: the Material Culture of Tea and Co昀昀ee Consumption in Eighteenth-Century Amsterdam”. In: Early Modern Things. Objects and their Histories, 1500-1800. Ed. by P. Findlen. London: Routledge, 2013, pp. 316–341. [24] N. McKendrick, J. Brewer, and J. H. Plumb. The Birth of a Consumer Society: the Commer- cialization of Eighteenth-Century England. London: Europa Publications, 1982. [25] T. Mikolov, K. Chen, G. Corrado, and J. Dean. “E昀케cient Estimation of Word Representa- tions in Vector Space”. In: arXiv preprint arXiv:1301.3781 (2013). [26] F. Montgomery. Textiles in America, 1650-1870: a Dictionary Based on Original Documents: Prints and Paintings, Commercial Records, American Merchants’ Papers, Shopkeepers’ Ad- vertisements, and Pattern Books with Original Swatches of Cloth. New York: Norton, 1984. [27] G. Muldrew. Food, Energy and the Creation of Industriousness: Work and Material Culture in Agrarian England, 1550-1780. Cambridge: Cambridge University Press, 2011. [28] D. Nguyen and J. Grieve. “Do Word Embeddings Capture Spelling Variation?” In: Proceed- ings of the 28th International Conference on Computational Linguistics. Barcelona, Spain (Online): International Committee on Computational Linguistics, 2020, pp. 870–881. doi: 10.18653/v1/2020.coling-main.75. url: https://aclanthology.org/2020.coling-main.75. [29] M. Overton, J. Whittle, D. Dean, and A. Hann. Production and Consumption in English Households, 1600-1750. Abingdon: Routledge, 2004. [30] “Parsing Early Modern English for Linguistic Search”. In: Proceedings of the Society for Computation in Linguistics 5 (2022). doi: https://doi.org/10.7275/twww-ef90. [31] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. “Scikit-learn: Machine Learning in Python”. In: Journal of Machine Learning Research 12 (2011), pp. 2825–2830. [32] J. Raven. The Business of Books: Booksellers and the English Book Trade, 1450-1850. New Haven: Yale University Press, 2007. [33] R. Řehůřek and P. Sojka. “So昀琀ware Framework for Topic Modelling with Large Corpora”. In: Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks. Val- letta, Malta: Elra, 2010, pp. 45–50. [34] W. Ryckbosch. “A Consumer Revolution under Strain: Consumption, Wealth and Status in Eighteenth-Century Aalst (Southern Netherlands)”. PhD Dissertation. University of Antwerp, 2012. 111 [35] W. Saelens. “The Comforts of Energy? Consumer Culture and Energy Transition in Eighteenth-Century Gent and Leiden (1650-1850)”. PhD Dissertation. University of Antwerp/ Vrije Universiteit Brussel, 2021. [36] J. Sear and K. Sneath. The Origins of the Consumer Revolution in England: from Brass Pots to Clocks. London: Routledge, 2020. [37] J. Stobart. “In and Out of Fashion? Advertising Novel and Second-Hand Goods in Geor- gian England”. In: Fashioning Old and New. Changing Consumer Patterns in Western Eu- rope (1650-1900). Ed. by B. Blondé and I. Van Damme. Turnhout: Brepols, 2009, pp. 133– 144. [38] J. Stobart. “Selling (Through) Politeness: Advertising Provincial Shops in Eighteenth- century England”. In: Cultural and Social History 5.3 (2015), pp. 309–328. doi: https://do i.org/10.2752/147800408X331416. [39] J. Stobart. “The Language of Luxury Goods: Consumption and the English Country House, c.1760-1830”. In: Virtus: Yearbook of the History of the Nobility 18 (2011), pp. 89– 104. [40] F. Trentmann. Empire of Things: How We Became a World of Consumers, from the Fi昀琀eenth Century to the Twenty-First. London: Allen Lane, 2016. [41] E.-J. Wagenmakers and S. Farrell. “AIC Model Selection Using Akaike Weights”. In: Psy- chonomic Bulletin & Review 11.1 (2004), pp. 192–196. doi: 10.3758/bf03206482. url: http s://doi.org/10.3758/BF03206482. [42] C. Walsh. “The Advertising and Marketing of Consumer Goods in Eighteenth Century London”. In: Advertising and the European City. Historical Perspectives. Ed. by C. Wischer- mann and E. Shore. Farnham: Ashgate Publishing, 2000, pp. 79–95. [43] L. Weatherill. Consumer Behaviour and Material Culture in Britain, 1660-1760. London: Routledge, 1988. A. Modifiers Descriptive adapted - ebony - white - turky - sedan - wilton - couch - domestic - elbow - library - madeira - chintz - scotch - silk - double - golden- - green - plated - servants - writing - dimity - carved - glass - window - red - steel - moreen - household - blue - bronze - sleeping - hollands - housekeepers - private - brussels - mohair - brick-built - four-post - royal - cotton - brass - french - muslin - parlour - nankeen - tin - bowed - built - irish - cheney - metal - repair - four-stall - check - dresden - eight - walled - card - gilt - printed - enamelled - leasehold - bordered - attached - culinary - damask - satin - arched - several - winged - public - english - 昀椀eld - german - farming - furnished - dwelling - goose - serges - walnut-tree - wainscot - sconces - crimson - pembroke - double-key’d - bay - detached - various - snu昀昀 - harrateen - three-stall - standing - kitchen - yellow - toned - persian - stained - brilliant - glazed - single - genoa - copper - inlaid - sundry - calico - camblet - iron - indigo - jamaica - circular - pier - general - japan - fowling - eight- day - pewter - breakfast - drawing - musical - pearl - japanned - marseilles - womens - 112 worsted - brown - carpeting - grey - velvet - wrought - dressing - chestnut - mahogany - horse-hair - wood - lisbon - india - spanish - nag-tail - marble - uphol昀琀ery - pair - livery - woollen - looking - wearing - stage - cornices - coloured - four - black - chelsea - foreign - singularly - ornamental - oval - cabriole - broad - eating - coach- - silver - copyhold - singular - stone - russia - 昀氀emish - miscellancous - diamond - dutch - feather - draught - shaving - complete - variety - drinking - dining - panned - italian - oriental - rosewood - worsted-damask - chinese - billiard Evaluative fancy - extensive - useful - requisite - celebrated - antique - splendid - old - lady’s - modern - improvable - beautiful - truly - richly - clean - very - cheerful - great - plain - exceedingly - well-chosen - large - scarce - valuable - convenience - comfortable - well- bred - superior - neatly - pleasing - long - spacious - genteel - larger - neat - 昀椀tted - light - compact - highly - easy - condition - elegant - pleasure - remarkable - 昀椀ne - proper - airy - convenient - masters - strong - prime - 昀椀nished - important - improved - capital - new - superb - taste - commodious - noble - well-built - excellent - select - fashionable - 昀椀ne-toned - curious - suitable - eminent - original - little - admired - magni昀椀cent - exceeding - stout - lo昀琀y - genuine - ladies - narrow - eligible - common - numerous - desirable - necessary - seasoned - handsome - roomy - principal - arable - new-built - clever - quality - super昀椀ne - gentleman’s - much-improved - substantial - rich - exquisite - family - grand - ancient B. Objects accessories jewellery - bracelet - locket - trinket - earring animal/accessories horse - saddle - harness - pony appliances/utensils barrel - butts - hearth - pistol - chimney - mangle - utensil - 昀椀re-arm - stove clothing/fabric counterpane - matress - mercery - shawl - habderdashery - handkerchief - sheet - clothes - hose - drapery decoration boxes - chandelier - pillar - picture - lamp - lustre - candelabra - globe - plant - carpet - cut-glass - vase - shell - screen - books - frame - chimney-glass - candlestick furniture settee - cellarets - drawers - desk - bedstead - cabinet - chair - couch - library-case - pantry - cabinet-work - sideboard - bureau - dining-tables - furniture - canopy - secretaire - commode - wardrobe - closet - chaise - chest - bookcase - press - sofa instrument instrument - harpsichord - piano - pianoforte real estate cistern - dining-room - lawn - out-building - estate - garden - buildings - farm-yard - villa - chamber - drawing-room - staircase - orchard - hall - stabling - wash-house - timber - land - house - bedchamber - farm-house - brewhouse - residence - premises - mansion - dwelling-house - lots - cellar - stable - pipe - cottage - garret - court - coach-house - bath - tenenement - warehouse - bed-room - messuage - attic - counting-house - apartment - farm - stall - cellaring - mansion-house tableware dish - china - bottles - decanter - porcelain - glasses 113