Word Usage Graphs (WUGs)

Word Usage Graphs represent usages of a word as nodes in a graph which are connected by weighted edges representing semantic proximity.

Word Usage Graphs

Word Usage Graphs (WUGs) represent usages of a word as nodes in a graph which are connected by weighted edges representing (human-annotated) semantic proximity.  We list various WUG resources below. These can be exploited in various ways, e.g. as resources of thousands of word use pair semantic proximity judgments or as clustered graph representations of word use sets. As such they provide various possibilities to evaluate computational lexical semantic models (e.g. contextualized embeddings, word sense disambiguation/discrimination) with additional aspects such as variation over time or dialect.

Find the idea behind WUGs described in this blog article. Also, consider the DURel annotation tool to annotate your own WUGs and enriched WUGs with cluster definitions. Find further information on the format and code to process the data below.

Each resource contains the following data:

  • readme: containing information specific to the resource.
  • annotators: a list of all (anonymized) annotators.
  • guidelines: used to train annotators.
  • data: for each word including
    • uses: with the main use in the column 'context' as a spelling-normalized version of 'context_tokenized' (see code below). The column 'indexes_target_token' gives the character indexes of the target token in 'context' (Python list ranges as used in slicing), while the column 'indexes_target_sentence' gives the character indexes of the target sentence (containing the target token) in 'context'. Similarly, the columns 'indexes_target_token_tokenized' and 'indexes_target_sentence_tokenized' contain the token indexes of target token and sentence in the tokenized, lemmatized and POS-tagged version of the word use (after splitting at whitespaces). We provide meta-information on each use (POS, date, etc.) and tokenized, lemmatized and POS-tagged versions if available.
    • judgments: human semantic proximity judgments of use pairs.
    • senses (optional): sense descriptions, if available. Annotators had the choice to assign uses to a generic sense description 'andere' ('others').
    • judgments_senses (optional): judgments of sense descriptions for uses, if available. If no judgment was provided by an annotator the column 'identifier_sense' contains the string 'None'.
  • graphs: WUGs derived from the annotated data as Python NetworkX objects (see code). The folder full/ contains the full graphs with all nodes and edges without clustering. opt/ contains the latest cleaned and optimized version including a clustering obtained with optimized parameters (see code). Additionally, we provide versions of the graphs related to previous publications (e.g. semeval/).
  • clusters: clusterings extracted from the respective graphs.
  • plots: two-dimensional visualizations of the respective graphs. Clusters are visualized as nodes with the same color (additionally marked by common numbers in nodes to avoid confusion of colors). Black edges mark median judgments with values above the chosen threshold for clustering (e.g. 2.5), while gray edges mark median judgments below that threshold. We provide various versions of the plots with different edge labels (e.g. edges labeled with all their judgments or only the median judgment respectively). Nodes contain the context of the corresponding word use with additional meta-information. Nodes can be moved by clicking and dragging. We also provide visualizations of subgraphs corresponding to each grouping of the nodes as specified in the use files (e.g. time periods) and pairwise aligned versions for each combination of groupings. Mozilla Firefox is often a good choice to open the visualizations.
  • stats: statistics derived from the respective graphs including clustering statistics. We also report statistics concerning the comparison of clusterings between groupings including change scores as well as annotator agreement.

Additional data may be provided for individual resources, as specified in their readme.

All text files are tab-separated and use UTF-8 encoding. Quotes do not mark strings, but are part of the original contexts.

We provide code to create, process and cluster the graphs in the WUG repository.

Dominik Schlechtweg. 2023. Human and Computational Measurement of Lexical Semantic Change. PhD thesis. University of Stuttgart.


  • language: German
  • groupings: 1800-1899, 1946-1990
  • judgments: 24k
  • download
  • reference: Sinan Kurtyigit, Maike Park, Dominik Schlechtweg, Jonas Kuhn, Sabine Schulte im Walde. 2021. Lexical Semantic Change Discovery. Proceedings of the Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing.
  • language: German
  • groupings: 1750-1800, 1850-1900
  • judgments: 4k
  • download
  • reference: Dominik Schlechtweg and Sabine Schulte im Walde. submitted. Clustering Word Usage Graphs: A Flexible Framework to Measure Changes in Contextual Word Meaning.


  • language: English
  • groupings: none
  • judgments: 9k
  • note: find scripts to convert the data under code below
  • download
  • reference: Katrin Erk, Diana McCarthy, Nicholas Gaylord. 2013. Measuring Word Meaning in Context. Computational Linguistics 39 (3), pp. 511-554.
  • language: English
  • groupings: 1910-1920, 1930-1940, 1950-1960, 1970-1980, 1990-2000
  • judgments: 16k
  • download
  • reference: Mario Giulianelli, Marco Del Tredici, and Raquel Fernández. 2020. Analysing lexical semantic change with contextualised word representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3960–3973, Online. Association for Computational Linguistics.



Other languages

  • language: Mandarin Chinese
  • groupings: 1954-1978, 1979-2003
  • judgments: 61k
  • download
  • reference: Jing Chen, Emmanuele Chersoni, Dominik Schlechtweg, Jelena Prokic, Chu-Ren Huang. 2023. ChiWUG: A Graph-based Evaluation Dataset for Chinese Lexical Semantic Change Detection. Proceedings of the 4th International Workshop on Computational Approaches to Historical Language Change 2023 (LChange'23).

Related Resources

This image shows Dominik Schlechtweg

Dominik Schlechtweg


Junior research group leader

This image shows Sabine Schulte im Walde

Sabine Schulte im Walde

Prof. Dr.

Akademische Rätin (Associate Professor)

To the top of the page