Elamur Research⌕
☰
home / concepts / Lexicography

📒 Lexicography

Concept · 30 key papers · all concepts

Lexicography is the discipline of compiling, describing and designing dictionaries and related lexical resources. It covers both the practical craft of writing entries and the theoretical study of how word knowledge should be selected, structured and presented to human users or to machines.

Also: dictionary making, dictionary compilation, practical lexicography, theoretical lexicography (metalexicography), electronic lexicography (e-lexicography), computational lexicography

What it is

Lexicography has two sides. Practical lexicography is the work of producing a concrete dictionary: choosing a word list, dividing words into senses, writing definitions, picking examples and adding labels. Theoretical lexicography, often called metalexicography, studies dictionaries as objects. It asks about their typology, their users, the structure of entries and the criteria by which dictionaries are judged.

The basic unit is the entry. A typical entry contains several parts:

Atkins gives a fuller inventory of entry components for bilingual dictionaries. It includes the following (Atkins 1996):

Today the field also covers resources that no human reads directly. These include lexical databases, machine-readable lexicons for parsers and taggers, and bilingual word lists produced by algorithms.

The key idea

The central methodological idea of modern lexicography is to ground the description of a word in evidence of its actual use, and increasingly to let statistics organise that evidence. Church and Hanks proposed the association ratio, an adaptation of pointwise mutual information that measures how strongly two words attract each other in a corpus (Church & Hanks 1990). Rychlý later proposed a score tuned to lexicographers' needs (Rychlý 2008):

I(x,y)       = log2 [ P(x,y) / ( P(x) · P(y) ) ]
logDice(x,y) = 14 + log2 [ 2 · f(x,y) / ( f(x) + f(y) ) ]

In the first line, P(x) and P(y) are the probabilities of encountering words x and y in the corpus, and log2 is the base-2 logarithm. P(x,y) is the probability that x is followed by y within a small window. Because it counts order, the ratio is asymmetric, unlike true mutual information. A value well above zero means the pair occurs together more often than chance would predict.

In the second line, f(x) and f(y) are the corpus frequencies of each word and f(x,y) is the frequency of the pair. The constant 14 sets the ceiling of the score: a pair whose words only ever occur together scores 14. Scores drop below zero only for very weak pairs (Rychlý 2008).

Church and Hanks used a window of 5 words and discarded pairs whose joint count f(x,y) was 5 or lower. They left the ratio asymmetric on purpose, so that word order is preserved. In 1988 Associated Press text of 44 million words, save…from appeared 176 times against 18 in the reverse order (Church & Hanks 1990). The resulting ranked lists serve the lexicographer as an index. Concordance lines can be sorted by strong collocates and then grouped into senses based on usage (Church & Hanks 1990).

The same principle extends to selecting examples. GDEX ranks corpus sentences automatically to propose good dictionary examples (Kilgarriff et al. 2008).

History and founding papers

Dictionary making is old. The European tradition was shaped by landmarks such as Samuel Johnson's single-author dictionary (1755) and the Oxford English Dictionary. The OED was compiled on historical principles by many editors and volunteer readers. Ladislav Zgusta's Manual of Lexicography (1971) is often cited as a founding text of the theoretical side. The corpus turn is usually traced to John Sinclair's COBUILD project and the corpus-based dictionary it published in 1987.

Computational work on dictionaries began with machine-readable versions of printed ones. Alshawi parsed the sense definitions of the machine-readable LDOCE (Longman Dictionary of Contemporary English). These definitions are written in a restricted vocabulary of about 2000 words. His aim was to extract the semantic head and other classification features (Alshawi 1987). On 500 random definitions, the head was correct in 387 cases, or 77% (Alshawi 1987).

Jensen and Binot mined Webster's Seventh for cues such as "used for" to decide prepositional-phrase attachment (Jensen & Binot 1987). Russell and colleagues built a dictionary and morphological analyser for English parsing (Russell et al. 1986). Church and Hanks's association ratio, described under the key idea above, then tied corpus statistics directly to lexicographic practice (Church & Hanks 1990).

Lexical databases form a parallel line of work. WordNet, developed at Princeton by George Miller's group, organised English vocabulary as a network of word senses linked by semantic relations (Miller 1994). It became a widely used resource for disambiguation. Stetina and Nagao, for example, used a WordNet-based semantic distance to resolve attachment. They reached 88.1% accuracy on about 3,100 test quadruples, that is, verb–noun–preposition–noun tuples (Stetina & Nagao 1997).

FrameNet, based on Fillmore's frame semantics, added syntactic realisations of meaning backed by annotated sentences. Boas designed a German counterpart linked to English through shared frames (Boas 2002).

Atkins, whose inventory of entry components appears above, argued that electronic bilingual dictionaries were still printed books in disguise. She proposed describing entries along six dimensions (Atkins 1996):

The automatic construction of bilingual dictionaries from existing dictionaries began with pivot methods. Tanaka and Umemura built a bilingual dictionary through a third language (Tanaka & Umemura 1994). The approach was later applied to Japanese–Chinese via English (Kaji et al. 2008).

How it is used today

Corpus-based dictionaries now operate at scale. The Leipzig Corpora Collection offers frequency lists, example sentences and significant co-occurrences. It had reached 136 online languages at the time of its report (Goldhahn et al. 2012).

Collaborative lexicography has produced Wiktionary, which has been examined as a possible rival to expert-built lexicons (Meyer & Gurevych 2012). Its translation tables also feed automatic methods. Ács triangulated through pivot languages across 43 Wiktionary editions. A classifier, trained on automatically collected examples, then judged which triangulated candidate pairs were correct and reached an F1 of about 0.93. Coverage of a basic-vocabulary lexicon rose from 67.4% to 80.2% (Acs 2014). Probabilistic inference over many dictionaries has been used to compile a massive multilingual one (Mausam et al. 2009).

Electronic lexicography depends on shared encoding standards. TEI-Lex0 reduces the TEI dictionary guidelines to a baseline for exchanging retrodigitised dictionaries, meaning dictionaries converted from printed editions into digital form. One of its rules is that every form element must carry a type, such as lemma, inflected, paradigm or variant (Bański et al. 2017).

Revisions of Moroccan, Syrian and Iraqi Arabic dictionaries are kept in a relational database. They are exported as XML that conforms to the Lexical Markup Framework (ISO 24613) (Graff & Maamouri 2012). Bilingual dictionaries have also been used to transfer FrameNet to French (Mouton et al. 2010).

Dictionaries are also valuable to NLP as supervision and as knowledge sources:

"Dictionary learning" in sparse autoencoders, which decomposes neural activations into feature directions, is a separate and unrelated use of the word (Rajamanoharan et al. 2024).

Open questions

Sense division is the first. Purely distributional scores cannot tell compositional pairs from idioms (Church & Hanks 1990). Boas argues that splitting senses by frame gives finer distinctions than EuroWordNet-style senses (Boas 2002). Still, there is no agreement on how fine senses should be or how they should map across languages.

The second issue is how far automatic construction can be trusted. Unsupervised bilingual induction assumes that embedding spaces in different languages are nearly isomorphic, that is, close enough in shape for one to be mapped onto the other. That assumption often fails. With a fully unsupervised method, English–Finnish reached a P@1 of 0 in every frequency band and word group analysed. P@1 is the share of test words whose top-ranked translation is correct (Søgaard et al. 2018). Common benchmarks for this task may also give misleading results (Kementchedjhieva et al. 2019). Pivot methods are vulnerable to polysemous pivot words (Acs 2014).

The third issue is quality control at scale. When there are hundreds of corpora, manual checking is impossible, so monitoring has to rely on statistical signals (Goldhahn et al. 2012).

Finally, the role of the dictionary is changing. As language models absorb lexical knowledge, it is unclear whether dictionaries will remain mainly reference works for people or become structured inputs for machines. Atkins's call to design entries around user needs (Atkins 1996) may need to be rethought for readers that are not human.

Key papers

The works this page cites, the most cited first; citations as counted by OpenAlex plus the papers of this library that cite the work, read 2026-10-23.

More papers

Topics: 📖 Lexical semantics & word embeddings · 🧷 Idioms & multiword expressions

Shelves: 🗣️ Linguistics & languages

This page in Markdown → https://md.research.elamur.ai/free/en/concepts/lexicography/