Corpus-based Distributional Data for Linguistic Generalization:

Frequency and more?

Alvin Cheng-Hsien Chen 陳正賢
National Taiwan Normal University

Feb. 17, 2023

Corpus-based Apporach to Linguistic Analysis

  • Frequency counts have been one of the central distributional properties extensively used in corpus linguistic analysis.
  • These include the frequency counts of words, multiword combinations, constructions/structures, or some other linguistic structures/phenomena.

Corpus-based Distributional Properties

  • Two types of distributional properties are most often extracted from corpus data:
    • occurrences
    • cooccurrences
  1. Gries, Stefan Th. 2017. Quantitative Corpus Linguistics with R: A Practical Introduction. Second Edition. New York and London: Routledge.
  • The occurrences of a particular linguistic unit
    • e.g., morphemes, words, grammatical patterns
  • The cooccurrences as the conditional distribution of linguistic units/phenomena
    • How often does a morpheme cooccur with a particular word?
    • How often does a word cooccur with a particular construction?
    • How often does a word/construction cooccur with a particular register/genre?
  • These occurrences and cooccurrences data often serve as empirical support for linguistic hypotheses
  1. Stefanowitsch (2019) defines Corpus Linguistics as “the investigation of linguistic research questions that have been framed in terms of the conditional distribution of linguistic phenomena in a linguistic corpus”.

In addition to frequency, what else?

  • Studies have shown that speakers are sensitive to more subtle distributional/statistical properties in language use.
  • Frequency counts may only capture one aspect of our native intuition (i.e., Which linguistic structure is more frequently used?)
  1. Ellis, Nick C, Matthew Brook O’Donnel & Ute Römer. 2014. The processing of verb-argument constructions is sensitive to form, function, frequency, contingency, and prototypicality. Cognitive Linguistics 25(1). 55–98.
  • Usage-based studies have highlighted a few important aspects of our statistical native intuition:
    • Exclusivity
    • Dispersion
    • Directionality
  • We will illustrate these ideas with the examples of two-word collocations.
  1. Gablasova, D., Brezina, V., & McEnery, T. (2017). Collocations in corpus-based language learning research: Identifying, comparing, and interpreting the evidence. Language Learning, 67, 155–179.

A Quick Example

  • Data Source:
    • English: Corpus of Contemporary American English (COCA)
    • Chinese: Academia Sinica Balanced Corpus of Mandarin Chinese (Sinica Corpus)
  • Target: Two-word combinations (Contiguous Bigrams)
  • Research Question: Representative Collocations

A Naive Start

  • Create a frequency list of two-word combinations based on the native corpus



While the frequency list may reflect our intuition, they may contribute little to the research question.


  • High-frequency of the two-word combinations may not necessarily reflect the strong lexical associations.
  • It is indeed a statistical evaluation. Many statistical metrics can be used, e.g., MI and LLR.

Now what if we add exclusivity to the analysis:


  • A linguistic structure can be evenly distributed across the documents of the corpus or specific to only a few documents/authors/registers etc.
  • Domain-general vs. Domain-specific?
  • Many quantitative metrics can be used, e.g., range, inverse document frequency, delta P.

Now if we add dispersion to our analysis:


  • When two words are attracted to each other (i.e., high exclusivity), their lexical association may be asymmetrical (e.g., apply for vs. at home)
  • Many metrics can be used: delta P, surprisal, transitional probability.

If we add directionality to our analysis:

An Integrated Analysis

Frequency and more?

  • To identify representative two-word combinations, one may need to consider distributional properties at multiple dimensions.
  • A representative combination may need to demonstrate considerably high frequency and exclusivity with wide dispersion.
  • The directionality of combination’s lexical association may lead to more diverse representative sets of combinations as well.

L2 Acquisition of Multifaceted Native Intuition

  • We have observed a similar development in L2 data as well.
  • L2 learners demonstrate increasing sensitivity to the multifaceted statistical distributional properties in their L2 production (written essays).
  1. Chen, A. C.-H. (2022).Assessing L2 learner’s multifaceted collocation competence: A comparative analysis of Chinese and English L2 learners’ essays. EUROCALL2022, Reykjavik,16–19 August 2022.
  2. Chen, A. C.-H. (2021). Acquisition of L2 collocation competence: A corpus analysis of exclusivity, directionality, dispersion and novel usage. Taiwan Journal of TESOL,18(1),29–61.

L1 vs. L2 Comparative Study

  • English Data
    • Learner: International Corpus Network of Asian Learners of English (ICNALE)
    • Native: Corpus of Contemporary American English
  • Chinese Data
    • Learner: The Test of Chinese as a Foreign Language (TOCFL)
    • Native: Academia Sinica Balanced Corpus of Mandarin Chinese (Sinica Corpus)

Structure of Analysis

  • All two-word combinations in L2 argumentative essays across different proficiency levels (CEFR)
  • L1 Corpus as the proxy for native intuition (Exclusivity, Dispersion, Directionality)
  • Evaluate the sophistication of L2 two-word combinations at different proficiency levels
  • As L2 proficiency level grows, the exclusivity of their two-word combinations grows as well.

  • Learners’ two-word combinations also become more and more sophisticated when their lexical associations are examined in bi-directional order (Forward-directed sophistication is more prominent.)

  • As L2 proficiency level grows, the dispersion of their two-word combination develops as well.
  • Advanced learners demonstrate greater capacity of domain-specific combinations.


  • We cannot ignore the multifaceted nature of the corpus-based distributional properties.
  • In particular, common corpus-based analyses (e.g., collocations, collexemes, keywords, lexical bundles/chunks) may need to consider dispersion and directionality in more comprehensive ways.
  • The choice of a proper metric is always a non-trivial issue in CL.

Thank you!