CLaCC History

Research at the intersection of language, cognition and computation at Uds has a long history, starting with CRC 100 “Electronic Language Processing” (SFB 100: Elektronische Sprachforschung) in the early 1970s. In subsequent years, researchers from linguistics, informatics and psychology were continuously involved in collaborative projects funded by Deutsche Forschungsgemeinschaft (see Figure below for a selection of the main projects) as well as BMBF and EU programmes. 

In the most recent project, CRC 1102 “Information Density and Linguistic Encoding” (2014-2026), we investigated the extent to which linguistic encoding decisions are driven by the goal of speakers to modulate the distribution, and density, of information across the linguistic signal. Against the background of information theory (Shannon), language is viewed from the perspective of (bounded) rational communication, according to which interlocutors strive to optimize the encoding of their utterances to (i) successfully convey their intended message, and (ii) optimize their cognitive effort. At the start of CRC 1102, there was converging evidence that the ease of processing linguistic material is correlated with its contextually determined predictability, which can be formally expressed as information density, or “surprisal”. Surprisal has been shown to be a strong predictor of linguistic encoding choices across linguistic levels, from the phonetic and grammatical to the semantic and discourse levels and across languages. Research in CRC 1102 has revealed that information density is not uniquely determined by linguistic context, but develops dynamically within specific situations, across interactions and over longer periods of time, with interlocutors continuously adapting to each other and the environment.

All 17 projects in CRC 1102 relied on a range of empirical methods and techniques, employing both controlled on-line and off-line experimental paradigms with human subjects as well as records of authentic language use (corpora), often in combination and complemented with the use of (large) language models. These were used for estimating surprisal, which in turn informed formal and computational models.