My primary interest in metaphor concerns its annotation in corpora, thus the problem of metaphor identification: How does one know when something is a metaphor and when it isn’t? This depends on what definition of metaphor (in language) one assumes. Thus, in my thesis project, I consider both of these questions in detail, in particular what implications they have for the case in which one wishes to annotate metaphor in historical corpora.
I was also a member (2022–2026) of subproject B01 in CRC 1475 “Metaphors of Religion”, which had to do with metaphor in Middle High German. In this subproject, we developed annotation guidelines (Dipper et al. (2024); Dipper et al. (2025)) and produced metaphor annotations in historical and modern texts.
The historical development that a lexeme undergoes has important implications for which metaphors are possible, since metaphors depend on which literal meanings are available, and these are always changing: Lexemes may gain or lose new senses, senses may broaden or narrow, etc. Thus this research area interacts with those relating to historical data and computational lexical semantics.
I was involved in the development of various corpora of historical German, most notably the Reference Corpus of Middle High German (ReM) (Klein et al. (2016); Roussel et al. (2024)), the Reference Corpus of Early New High German (ReF), and the Anselm Corpus.
My work in these projects involves the processing of diplomatic transcriptions, which provide a variety of useful information, e.g.: about the legibility of the text, tokenization information, ligatures, diacritics, abbreviations, etc. I write and maintain software to convert the corpus data between formats and pay particular attention to representing the source data faithfully and appropriately for each use case. With the ReM v2 series, this includes GraphML, for use with ANNIS, and TEI XML, from which the PDF versions for reading are generated.
My work in computational lexical semantics includes tasks such as word sense
induction and lexical semantic change detection, both of which come into play
in the course of my dissertation project, since the task of
annotating metaphor (especially in historical corpora) depends on information
pertaining to literal (or conventional) meanings and how these change over
time. My erwv2cdo program implements algorithms for both of
these tasks.
Another area of interest is in non-distributional representations of meaning.
My wn-vsa project involved constructing representations for
synsets from the relational structure of a wordnet using the techniques of
vector symbolic architectures (which see below), as opposed to corpus data.
Vector symbolic architectures (VSA, also known as hyper-dimensional computing, HDC) are a family of approaches for representing symbols and structurally complex entities with vectors and performing computations by means of specially chosen operations on these vectors.
I’m interested in methods incorporating VSAs, because they offer the promise of a seamless integration of complex, structured symbolic representations with the robustness and graded similarity that vector representations provide. I also see in these methods the promise of a more lightweight, resource-efficient, and transparent approach to processing linguistic data that stands in marked contrast to the currently dominant paradigm.
Apart from the aforementioned wn-vsa project (Roussel
(2023)), in my vsatreebank
project (Roussel
(2026), I use these techniques to
encode an entire treebank to implement an example-based search application
that, due to the properties of VSAs, is lightweight and allows for fuzzy
matching.
For a while (ca. 2016–2021) I worked on the automatic resolution of discourse
deixis. Discourse deixis involves two main classes of referring expressions in
English and German, demonstrative pronouns and shell nouns. To a lesser extent,
the personal pronoun it (eng.) or es (deu.) also plays a role. Such
expressions are used to refer to abstract entities, such as events, processes,
propositions, etc., which are typically realized linguistically with
non-nominal expressions, in contrast to the usual antecedents of anaphors. In
our survey on the topic, we give
an extensive overview of the topic from a computational perspective, and our
nna-datasets project (Roussel et al.,
2018) provides annotated data and some
analysis.
I also designed and implemented a resolution system for my master’s thesis project, which involved a combination of lexico-syntactic patterns and pattern-specific classifiers.
The full list of my publications is here.