In research on language, neighborhoods are a conglomeration of words that are highly similar to one another along a critical characteristic. Most commonly, neighbors are defined on the basis of shared linguistic features such as orthography, phonology, or semantics. Because a word’s neighborhood size (i.e., the number of neighbors it has; also called neighborhood density) can have an impact on a variety of linguistic tasks and processes, it has become an important psycholinguistic metric. However, in spite of the focus on neighbors in psycholinguistic research, neighbors are inconsistently identified, particularly across languages. These inconsistencies, which often arise as a result of researchers employing different databases, make it difficult to compare the effects of neighborhood density across studies. The current paper has two goals: (1) to introduce a centralized database of neighborhood information for five commonly-studied languages – Dutch, English, French, German, and Spanish – and provide a single corpus through which neighborhoods can be indexed cross-linguistically; and (2) to compare general properties of neighborhoods across these five languages using this database in order to determine where and how languages differ in respect to their neighborhoods.

In the current paper, we examined two types of linguistic neighborhoods – orthographic and phonological. Orthographic neighborhoods are often defined according to Coltheart, Davelaar, Jonasson, and Besner’s [1] N metric, which refers to the number of words that can be constructed by substituting one letter of the target word. For example, the word log has hog, lug, and lot as orthographic neighbors. Phonological neighborhoods are calculated similarly, but instead of depending on grapheme substitution, phonological neighbors are constructed by substituting one phoneme of the target word [2]. Fish (/fι∫), for example, has dish (/dι∫/) and fig (/fιg/) as phonological neighbors. These “substitution neighbors” have historically been the focus of the literature and have dominated investigations of neighborhood size. However, research has also investigated the effects of addition (formed by the addition of a grapheme or phoneme, for example and has hand as an orthographic addition neighbor) and deletion (formed by the deletion of a grapheme or phoneme, for example bend has end as an orthographic deletion neighbor) neighbors [3].

The effects of phonological and orthographic neighborhood density on language processing have been well documented across a variety of tasks [4]–[11] and across multiple languages [12]–[15]. However, in spite of the prevalence of neighborhood effects, the nature of these effects is subject to debate. For example, neighborhood density may affect recognition and production processes differently [16], [17], and effects may vary depending on the language of presentation [13], [18], [19] (but see [14]). The ongoing debate surrounding neighborhood density effects, particularly across languages, underscores the need for resources that allow researchers to consistently identify orthographic and phonological neighbors across studies. For some languages, even the most basic descriptive data are not available, forcing researchers to continually recreate basic neighborhood and frequency statistics. Furthermore, even when descriptive statistics are available [13], [15], [20], [21], direct cross-linguistic comparisons are often not reported or possible.

While there have been some attempts to create consistent corpora from which neighborhood information can be derived, these corpora vary across languages. For example, N-Watch, a database of English neighborhood information [22], defines phonological neighbors according to the substitution of a single phoneme in any word position. BuscaPalabras, a database of Spanish neighborhood information [8], and E-Hitz, a database of Basque neighborhood information [23], define phonological neighbors according to those same rules, but also include words that differ by the addition or deletion of a phoneme from any word position.

The goal of this paper is therefore to introduce CLEARPOND, the Cross-Linguistic Easy-Access Resource for Phonological and Orthographic Neighborhood Densities, a catalog of neighborhood density across languages. Perhaps the most comprehensive psycholinguistic database to date is WordGen [21], which queries the CELEX and Lexique databases to provide searchable datasets for Dutch, English, German, and French. While WordGen controls for factors such as written word frequency, orthographic neighborhood size, bigram frequency, and word length, it is missing a number of relevant features including information on phonological neighbors, neighborhood frequency, and the ability to index neighbors across languages. The database that we present here has been controlled for word frequency to ensure that consistent and comparable tokens are sampled from each language, and provides data regarding word length, neighborhood density, and neighborhood frequency. We also provide measures of foreign neighborhoods (i.e., the number of Spanish neighbors of an English word, or English neighbors of a Spanish word, etc.) for use in bilingual comparisons. Neighborhoods are defined both orthographically and phonologically, with stimuli derived from film and television subtitle corpora that capture spoken word frequencies. Finally, we have defined neighborhoods by substitution, addition, and deletion. It is our intent that CLEARPOND will provide a standard from which neighborhood data can be easily extracted and that it will provide a comprehensive tool for psycholinguistic researchers.

To examine phonological and orthographic neighborhood densities across languages, we selected corpora for the following languages: Dutch (SUBTLEX-NL) [24], English (SUBTLEX-US) [25], French (Lexique) [26], German (SUBTLEX-DE) [27], and Spanish (SUBTLEX-ESP) [28]. Misspellings, including culturally-defined spellings (e.g., British “colour”), and foreign language intrusions (e.g., the English word “mind” in the Spanish corpus) were removed by cross-referencing each subtitle corpus with a dictionary in that language. Because all five corpora use the same source-material (i.e., film and television subtitles) to derive frequency data, they are highly comparable and well suited for cross-language comparisons. To increase similarity among the corpora, homographs were removed from the French corpus to match the parameters of the Dutch, English, German, and Spanish corpora (none of which distinguish between the different meanings of homographs). French homographs were reduced to a single entry, and the frequency per million of the collapsed entry was created by adding the frequency per million of each of the homographs. For example, the French word est is the third person singular form of the verb meaning “to be,” and has a frequency of 19,417 per million; est is also the French word for the cardinal direction East, which has a frequency of 81 per million. We collapsed these two entries into a single entry, est, that had a frequency of 19,498 per million.

Using large corpora (the subtitle lexicons range from 74,286 to 441,132 tokens) can lead to overestimations of neighborhood size compared to people’s actual working vocabularies. By only including words above a certain frequency threshold, the effect of very low frequency words (which are unlikely to be in people’s everyday, working vocabularies) on neighborhood calculations is reduced. In the present study, a frequency threshold of 0.34 per million was used, based on the standard used by Davis [22]. This frequency cutoff yielded a corpus size of 27,751 for English, which compares favorably to English vocabulary size estimates for educated adults (20,000 word families) [29]. However, the frequency cutoff yielded different corpus sizes across languages (Dutch: N = 31,691; English: N = 27,751; French: N = 34,113; German: N = 45,027; Spanish: N = 41,968), which would limit our ability to make cross-linguistic comparisons. Larger corpora are likely to inflate neighborhood size estimates, as a larger overall sample pool results in a larger pool of potential neighbor-candidates. To alleviate this concern, corpus size was equated across languages by including the 27,751 most frequent words in each language (based on the smallest corpus, English) in all further comparisons. Figure 1a (left) shows that when corpus size was equated, the languages had comparable average frequencies (Dutch: 32.58, SEM = 3.10; English: 32.72, SEM = 3.18; French: 30.87, SEM = 2.64; German: 33.74, SEM = 2.74; Spanish: 33.87, SEM = 3.02), while Figure 1a (right) indicates that the languages differed in average frequency when corpus size was instead defined by a frequency threshold. In addition, frequency distributions (Figure 1b) were comparable across languages when corpus size was equated. Together, these results provide support for the ability to make direct comparisons between the size-equated corpora.

Because the present analysis derived orthographic and phonological neighborhoods from the same subtitle corpora, we were able to make direct comparisons between the two neighborhood types. The differences that emerge in the relationships between these neighborhoods across languages can potentially be used to illuminate differences in language transparency. Transparency, or orthographic depth, is a measure of how closely a language maintains a one-to-one grapheme-phoneme correspondence; the more transparent a language, the more the graphemes and phonemes are tightly matched. For example, in the most transparent of languages, each phoneme would map to only one grapheme and vice versa (e.g., the Spanish phoneme/m/is always represented by the grapheme m, and the m grapheme always corresponds to the phoneme/m/). Conversely, opaque languages are those in which grapheme-phoneme mappings are less consistent; multiple graphemes can represent the same phoneme (e.g., English k and c can both represent the phoneme/k/), and more than one phoneme may be represented by a single grapheme (e.g., English g can represent the phonemes/g/and/ /). Because the grapheme-phoneme mappings of transparent languages are consistent, in these languages, many orthographic neighbors are also phonological neighbors. When phonemes and graphemes are consistently matched, the phonetic transcriptions of words mirror the orthographic structure. Therefore, when a single grapheme substitution (or addition or deletion) results in the creation of a new word, it is likely that the new word similarly differs from the original in only one phoneme.

Our analyses suggest that, in addition to indexing language transparency as a strict match between grapheme-phoneme correspondences, there may be a relationship between a language’s transparency and the degree of similarity between the language’s orthographic and phonological neighborhoods. For example, Spanish and German (both considered to be transparent languages [33]), demonstrate a high degree of similarity in the distributions of their orthographic and phonological neighborhoods. However, the similarity between orthographic and phonological neighborhoods is not quite as tightly coupled in German as it is in Spanish, likely because, German contains specific consonant clusters (e.g., sch) that correspond to single phonemes (e.g.,/∫/). Accordingly, there is higher similarity between graphemic and phonemic word lengths in Spanish than in German, Dutch, or English (Figure 14). French, a language with a high number of silent letters and digraphs, has the largest difference between graphemic and phonemic word length.

In addition to revealing differences between phonological and orthographic neighborhoods, our data illustrate differences in how substitution, addition, and deletion neighbors are used across languages.

In our analysis of foreign neighbors, we restricted comparisons to English and each other language (Dutch, French, German, and Spanish) to facilitate ease of comparisons, and because English is one of the most commonly learned second languages [34]. Foreign orthographic neighbors were found to make relatively substantial contributions to overall neighborhood size, constituting between 13–20% of a word’s total neighbors on average. Within-language neighbors still dominated overall neighborhood size, likely because languages have different orthotactic rules and requirements for the formulation of valid words. The result is that words in each of the languages we examined were more similar in orthographic form to other words within the same language than they were to foreign words.

Compared to foreign orthographic neighbors, foreign phonological neighbors were very rare. The effect of foreign phonological neighbors on overall neighborhood size was quite low, and the percentage of a word’s neighbors that derived from a foreign language was even lower, between 1–8%. These results are consistent with those of Vitevitch [32], who conducted an analysis of foreign phonological neighbors across Spanish and English and found that the two languages share relatively few neighbors.

One potential reason for the small number of foreign neighbors is that though the five languages we investigated share an alphabetic system (aside from accented letters), they contain phonological systems that are much more distinct. Because the orthographic structure of a language is anchored by that language’s writing system, orthography does not vary much over time. Conversely, a language’s phonetic structure has much more freedom to vary over time and across geographical space; the accumulation of these phonological changes likely contributes to the languages’ phonological distinctiveness, thereby reducing the number of foreign phonological neighbors.

While comparisons of foreign neighbors can be used for purposes of stimuli construction and to validate cross-linguistic comparisons, it is important to note that our data should not be interpreted as a measure of the bilingual mental lexicon. In order to make true claims about the nature of bilingual lexical representations based on corpus analyses, it would first be necessary to procure a bilingual corpus in which frequency values are representative of usage when a single individual speaks two languages. To our knowledge, such a corpus does not exist. If bilingual corpora can be obtained, it would be worthwhile to conduct neighborhood analyses using those lexical entries.

The corpus analysis presented in the current study provides a novel tool for researchers who study language processing. It enables comparisons between orthographic and phonological neighbors and within and across five languages.

While neighborhood information for some languages has been made available in the past [13], [15], [20], [21], the database that we present here provides comparable corpora and analyses across languages. We also expand upon the past examinations of foreign neighbors in Spanish and English [32] by supplying foreign neighborhood data for four language pairs – English-Dutch, English-French, English-German, English-Spanish – and by including both orthographic and phonological neighbors. Our future efforts will focus on developing a comparable corpus derived from written word data using written-word databases, such as Google Ngram (http://books.google.com/ngram) to complement our present work on spoken language.

In sum, the current paper presents a unified database for indexing neighborhood information derived from spoken corpora. These data provide cross-linguistic metrics that are crucial for designing experiments of spoken and written language processing. We have made our database available in searchable form (see Figure 15 for a screenshot of the web interface) at http://clearpond.northwestern.edu; it is also freely available for download.