Jun.-Prof. Dr. Titus von der Malsburg
Psycholinguistics and Cognitive Modeling
Address:
University of Stuttgart
Institute of Linguistics
Keplerstraße 17
70174 Stuttgart
Profiles:
Github
Google Scholar
OSF
ORCID
News
- New article in SSR: Mézière, Yu, Reichle, MacArthur, von der Malsburg (2023). Scanpath regularity as an index of reading comprehension [ PsyArXiv ]
- AMLaP 2023 talk: Laurinavichyute, Yadav, von der Malsburg, & Vasishth (2023). The role of goal in sentence processing.
- CogSci 2023 article: Ranjan, von der Malsburg (2023). A bounded rationality account of dependency length minimization in Hindi. [ arXiv ]
- Preprints:
- Laurinavichyute & von der Malsburg. (2023). Agreement attraction in grammatical sentences and the rule of the task [ PsyArXiv ]
- Pankratz, von der Malsburg, & Vasishth (2022). Shannon entropy is a more comprehensive and principled morphological productivity measure than the standard alternatives. [ PsyArXiv ]
- Laurinavichyute & von der Malsburg. (2023). Agreement attraction in grammatical sentences and the rule of the task [ PsyArXiv ]
Bio
I investigate how the human brain makes sense of language. How is each word that we hear or read combined with our understanding so far? What sources of knowledge are recruited in this process? And how are they combined, especially when they are in conflict? To answer questions like these, I use experimental and computational methods ranging from eye-tracking and event-related brain potentials to large-scale crowd-sourcing, corpus methods, Bayesian data analysis, and cognitive modeling.
I am a tenure-track professor at the Institute of Linguistics at the University of Stuttgart. Previously, I held postdoctoral positions at University of Potsdam, UC San Diego, and the University of Oxford. I also served as a visiting professor for two years at the University of Potsdam. I earned my PhD in Cognitive Science from the University of Potsdam under the supervision of Shravan Vasishth and Reinhold Kliegl. In addition to my academic work, I've worked as a software engineer, founded two companies, and collaborated with sound artists and composers at the intersection of science and art.
My name is a bit unusual and can cause confusion. To clarify: My first name is “Titus” and my surname is “von der Malsburg”. Despite common misconceptions, my name should not be written as “Von Der Malsburg”, just “Malsburg”, or “van der Malsburg” (with “van” instead of “von”). Nor should “von der” be treated as a middle name. All these forms are incorrect. For historical and practical reasons, it’s customary to list my name alphabetically under “M” not “v”. In BibTeX entries, it’s best to write {von der Malsburg}, Titus
and to include the field sortname = {Malsburg}
when I’m the first author of a publication.
Some of my research interests: Incremental sentence comprehension; reading comprehension, reading strategies, and scanpaths; syntactic ambiguity and ambiguity resolution; implicit linguistic gender biases; cue-based parsing, working memory; locality and word order; expectation and prediction; surprisal and information density; cognitive biases and bounded rationality; resumptive pronouns; dog whistles, plausible deniability, and social meaning; experimental methods, especially web-based and eye-tracking; corpus methods and large language models; Bayesian and frequentist statistics; cognitive modeling
Media coverage:
- New York Times (25.01.2020): She’s the Next President. Wait, Did You Read That Right?
- In Her Words, New York Times newsletter on women, gender and society (04.02.2020): Surprise! Language is sexist.
- WNYC Studios, The Takeaway: America's weekday conversation (podcast, 29.01.2020): Could "She" Be President? A Look at Pronoun Bias in Politics
- MIT News (08.01.2020): “She” goes missing from presidential language. Also reported by ScienceDaily, AAAS EurekaAlert!, and Phys.org.
- Kunstjahr 2013 – Die Zeitschrift, die Bilanz zieht, Lindinger + Schmid. Article about my science/art cross-over project with sound artist Christoph Illing.
Lab
My group currently consists two postdocs, Dr. Anna Prysłopska and Dr. Sidharth Ranjan, Candy Adusei, who supports us in the eye-tracking lab as a student assistant, and myself. Together, we use a range of tools to conduct research, including a TRACKPixx3 eye-tracker (2Hz, binocular, similar to EyeLink 1000+) and three GazePoint trackers (two GP3 HD and one GP3) which are mostly used for eye-tracking education. (We also have an SMI iViewX system, but no one knows if it still works.)
Working with Us: If you are interested in joining our team as a doctoral student or postdoc, we encourage you to get in touch so we can discuss potential opportunities. Before doing so, we kindly ask that you conduct some research to see whether our research interests align with yours. Our group emphasizes high technical quality, robustness of research findings, replicability, and reproducibility. This means that lab members are expected to possess or develop proficiency in statistics and in software tools such as Rmarkdown or LaTeX/Knitr, and version control systems like git. Basic programming skills in Python or R will also be beneficial. When reaching out, ensure that your message explicitly connects your research interests with the specific research conducted in our group. This will help us better understand how you could potentially contribute and thrive as a member of our team.
Publications
Click the button below to download a BibTeX file listing all my publications. This file can be imported into common bibliography managers such as Zotero, Jabref, Mendeley, etc.
Submitted manuscripts
Journal articles
Proceedings articles
Conference talks

Pioneering work in reading research has shown that scanpaths in reading can be informative about sentences processing (Frazier, Rayner, 1982). Nevertheless, scanpaths have not gained much traction in reading research. One reason for that may have been a lack of suitable analytical tools. Here, we summarize three recent studies in which we used a new scanpath measure to analyze gaze data from two experimental studies (von der Malsburg, Vasishth, 2011, 2012) and one corpus study (von der Malsburg, Kliegl, Vasishth, under revision). The experiments investigated how readers process temporarily ambiguous sentences. We showed that readers do not always commit to one of the alternative interpretations, and that readers with low working-memory capacity do so less often. Contrary to what was reported earlier, we found that reparsing instead of targeted repair is a common strategy to recover from incorrect interpretations. Interestingly, these results did not emerge in an analysis using traditional word-based eyetracking measures showing their limitations. In the corpus study, we demonstrated how syntax, oculomotor constraints, and age of reader jointly determine the regularity of scanpaths. We argue that, taken together, these results establish the scanpath as an informative and tractable object of investigation in reading research.
Recent studies (Dimigen et al., JEP:G, 2011; Kretzschmar et al., NeuroReport, 2009) demonstrated the feasibility of investigating fixation-related potentials (FRPs) and that the results are similar to brain potentials recorded during rapid serial visual presentation (RSVP). We conducted an experiment to assess if this correspondence also holds for rhythmic brain activity. To pit our results against ults gained with RSVP, the experiment used sentences that are known to produce reliable effects both in the time domain and in the frequency domain (Hagoort, Science, 2004). Participants (N=48) read freely through true ('The Thames flows through London') and false statements ('The Hudson flows through London'). Hagoort et al. report an N400, a negative-going deflection in the ERP with a peak around 400 ms, as well as increased theta (4-7 Hz) and gamma (30-70 Hz) activity. During our experiment, participants' eye movements were monitored and their EEG was recorded from 32 electrodes and later evaluated contingent to the first fixation on the critical word. We analyzed the canonical eye-tracking measures with linear mixed-effects models and the EEG with cluster-permutation tests (Maris & Oostenveld, J Neurosci Methods, 2007) to control for multiple comparisons. As expected, we observed a negativity in the FRP with a centro-parietal distribution and a peak latency of approximately 400 ms. Also as expected, this N400 lined up with increased first fixation durations, gaze durations, and regression rates in the eye movement record (Dambacher & Kliegl, Brain Res, 2007). Crucially, fixation-related power spectra showed synchronization in the delta range (1-3 Hz) at central electrodes and desynchronization in the upper alpha range (11-13 Hz) at occipito-parietal sites relative to a pre-fixation baseline. None of these effects is reported by Hagoort et al. which suggests that fixation-related EEG changes are at least not fully comparable to those observed in RSVP. One reason for the diverging results could lie in different processing demands: In RSVP, readers must retrieve earlier parts of the sentence from working memory because they cannot make regressions. Increased theta activity reflects this more effortful memory access (Klimesch, Brain Res Rev, 1999). In natural reading, readers can easily move their eyes back to resolve the processing difficulty which facilitates memory access. Our findings question the comparability of results acquired with serial presentation vs. natural reading.
Recent research (Dimigen et al., JEP:G, 2011; Kretzschmar et al., NeuroReport, 2009) shows that fixation-related potentials (FRPs) yield similar results as brain potentials recorded during rapid serial visual presentation (RSVP). We conducted an experiment to see if this correspondence also holds for oscillatory brain dynamics. Participants (N=48) read true ("The Thames flows through London") and false factual statements ("The Hudson flows through London"). Such violations are known to elicit an N400, a negative-going deflection with a peak around 400 ms, and increased theta and gamma activity (Hagoort, Science, 2004). As expected, we see an N400 in the FRP and increased fixation durations and regression rates in the eye movement record. Moreover, a cluster-permutation test (Maris & Oostenveld, J Neurosci Methods, 2007) for fixation-related power spectra shows synchronization in the delta range (1-3 Hz) and desynchronization in the upper alpha range (11-13 Hz) but no theta or gamma effects. This is at odds with prior findings and suggests that fixation-related oscillatory EEG changes are not fully comparable to those observed in RSVP. One reason for the diverging results may lie in different processing demands: In RSVP, readers must retrieve earlier parts of the sentence from memory because they cannot make regressions.

Conference posters and short talks

During reading our brain predicts upcoming words. If predictions are correct, words can be processed faster when they are finally fixated. It has been amply shown that Predictability (the variable that estimates the probability of guessing the next word) have an impact on how we move our eyes across the text and that it modulates brain potentials associated with word processing. On the one side, more predictable words are fixated for shorter periods of time than less predictable words. On the other side, more predictable words correspond to less N400 amplitude. This knowledge comes from separated EEG and eye movement experiments, but in the last few years, co-registration experiments enabled us to test these hypotheses together in more natural contexts. With the aim of investigating different sources of predictions during reading, in previous studies, we showed that mnemonic predictions (i.e. predictions performed purely on long term memory, like when reading a proverb or a song lyric) and predictions done purely on the linguistic context have different impact, both on gaze duration and on the N400. Here, we asked participants to read proverbs and common sentences while we recorded EEG and eye movements simultaneously. Firstly, we analysed brain activity aligned to fixation onset (fixation-related potential, FRPs) showing differences between Proverbs and Common sentences in late potentials evoked by low- and high-Predictable words. Secondly, we analysed oscillations aligned to fixation onset (fixation-related spectral perturbations, FRSPs) showing differences between sentence type only in low-frequency bands after 200ms. These results extend our knowledge of the differences between the mechanisms involved in the prediction of the following word.



Language production and comprehension draw on wide-ranging knowledge and beliefs, including general world knowledge and contextually variable information. Pronominal references to role nouns with diverse gender biases provide a window into the interplay of these sources of information: violations of stereotypical gender elicit surprise (e.g., referring to a surgeon as she), but comprehenders can accommodate to non-stereotypical genders within discourse. In two experiments, we investigate how gender expectations are reflected in production and comprehension of pronominal references to role nouns. Our results indicate that female gender is consistently underused in English pronoun production, and under-inferred in English pronoun comprehension.





Standard analyses of eye movements in reading test a set of canonical dependent measures calculated for multiple regions. Although the resulting multiple comparisons increase the rate of false positive results, it is accepted standard practice not to correct for that. We investigated false positives rates through computer simulations and tested how much statistical power has to be sacrificed to control them. In 100.000 iterations, we generated six realistic data sets of eye movements for a hypothetical experiment with typical parameters. The true effect sizes of the manipulation ranged from 0 ms to 40 ms. Four standard measures were analyzed using linear mixed models: first fixation duration, gaze duration, go-past time, and total viewing time. In the data sets with no true effect, the rate of false positives was 12.1%, i.e. much higher than the conventionally accepted 5%. A Bonferroni correction reduced false positives to 3.2% and was therefore slightly more conservative than required. The reduction in power due to the Bonferroni correction was moderate, e.g., from 90% to 75% for effect sizes of 5 ms. Contrary to conventional wisdom, the Bonferroni correction seems to be an appropriate tool for controlling false positives in reading experiments.

Recent research demonstrated the feasibility of analyzing fixation-related brain potentials (FRPs) recorded during natural reading (Kretzschmar et al., 2009; Dimigen et al., 2011). Two questions arise from these studies: (1) Are effects observed with fixation-triggered EEG signals comparable to those observed in standard RSVP designs? (2) Does the combined analysis of EEG and fixation data provide additional insights into reading and comprehension processes that are not available with either method alone? Both above-mentioned studies used material known to robustly elicit strong N400 effects. The present study (N=50) examined responses to a more subtle manipulation representative of common experimental designs: we manipulated the distance between anaphoric expressions (pronouns, verb ellipsis) and their antecedents. Differences were examined using a non-parametric Monte Carlo test (Maris & Oostenveld, 2007). Increased distance of the antecedent elicited an early frontocentral negativity in response to verb ellipses (88ms-134ms, p<0.001) and a negativity at frontocentral and parietal electrodes on the words following pronouns (98ms-186ms, p<0.001). We discuss these results in the context of earlier findings and argue that adopting FRP-methodology requires factoring in complex visuomotor contingencies that are not yet fully understood and that evoke ERP effects different from those seen in RSVP designs.
Seit langem wird debattiert, wie reguläre und irreguläre Vergangenheitsformen repräsentiert und verarbeitet werden (Rumelhart & McClelland, 1986; Pinker & Prince, 1988). Das Dual- Mechanism-Modell (DMM; Pinker & Prince, 1988; Clahsen, 1999) nimmt an, dass reguläre und irreguläre Formen von zwei verschiedenen Mechanismen verarbeitet werden. Vertreter des Single-Mechanism Ansatzes gehen alternativ von einem einzigen Mechanismus aus, der sowohl der Verarbeitung von regulären als auch irregulären Verben dient.


Artistic works
See this video about a sound installation I created together with sound artist Christoph Illing.
During reading, we rapidly construct meaning from sequences of rather cryptic symbols. A multitude of processes are involved in making meaning happen, however, most of them are conveniently tucked away from the reader’s conscious experience allowing them to read effortlessly without having to worry about any of the practicalities such as where to place the gaze next and for how long. The present work reflects on the marvelous feat that is reading. In an experiment-like situation, we create a perceptual short-circuit that unlocks the otherwise unconscious processes involved in reading. To this end, the test subject is exposed to a written libretto while a computer tracks their eye movements and translates them to sound in real time. The artistic implementation is based on parametric synthesis (“mapping”) and model-based sonification.
“Blog”
Better described as random technical notes resulting from excessive procrastination. I have a lot more of those but I’m too embarrassed to show them.
- Cohen’s d in simple linear regression models (Sept 2022)
- How to calculate worker compensation for Amazon Mechanical Turk (Dec 2016)
- A potential pitfall when running Ibex experiments on Amazon Mechanical Turk (Feb 2016)
- What’s the difference between predict vs simulate in lme4? (May 2015)
Software
- Scasim
- An R package that implements our measure for scanpath similarity along with some tools for data prerocessing and visualization of scanpaths. The measure is described in von der Malsburg & Vasishth (JML, 2011).
- Saccades
- An R package for detecting saccades and fixations in raw eyetracking data. Implements the velocity-based algorithm proposed by Engbert & Kliegl (Vis Res, 2003).
- edfR
- An R package for reading EDF files generated by EyeLink eye-trackers.
- helm-bibtex
- A bibliography manager for Emacs.
- TEDview
- A program for visualizing discrete temporal data, e.g., events taking place in an incremental dialogue system. See von der Malsburg, Baumann, Schlangen (2009) for details.
- py-span-task
- A program for testing reading or operation span. Implements the recommendations given by Conway, Kane, Bunting, Hambrick, Wilhelm, & Engle (Psych Bull & Rev, 2005).