ParHistVis: Visualization of Parallel Multilingual Historical Data

Aikaterini-Lida Kalouli, Rebecca Kehlbeck, Rita Sevastjanova, Katharina Kaiser, Georg A. Kaiser, Miriam ButtACL 2019

TL;DR

This paper introduces ParHistVis, a novel, free, and interactive visualization tool designed to help researchers analyze complex historical linguistic data. By combining visual analytics and computational linguistics, ParHistVis tackles the challenges of sparse, high-dimensional historical corpora, enabling deeper insights into language change.

What Problem Does This Paper Address?

Studying language change through parallel corpora offers significant advantages, allowing for direct comparison across time and languages. However, historical linguistic data is often sparse and high-dimensional, making it difficult to exploit with traditional statistical methods. Manual inspection is often infeasible due to the sheer volume of data, and existing visualization tools are not well-suited for parallel, multilingual, diachronic, and synchronic linguistic datasets. This gap hinders researchers from fully understanding the intricate interactions that drive language evolution.

Methods

ParHistVis is a web-browser-based tool implemented in JavaScript. It utilizes a Visual Analytics approach, allowing users to upload their data in a simple tabulated format. The tool offers several visualization views:

  • Matrix View: This is the initial view, presenting the data in a matrix where languages and time periods form the horizontal axis, and the corpus instances are on the vertical axis. This view allows for direct comparability and qualitative inspection. Users can filter data by selecting specific columns (features).
  • Aggregated Matrix View: This view provides a higher-level overview by aggregating data points. It offers a standard aggregation of the entire corpus or custom aggregations based on patterns in the data's unique IDs. A colormap indicates the frequency of features within the aggregated dimension.
  • Streamgraphs: To overcome the limitations of aggregation (where patterns can disappear), streamgraphs are used. These are stacked area graphs that visualize changes in feature frequencies over time or another aggregated dimension, highlighting trends and outliers.
  • Sankey Diagrams: To explore potential interactions between different linguistic features (dimensions), Sankey diagrams are employed. These visualize "flows" between selected dimensions, with the width of the flow representing the strength of the correlation. This allows users to define and investigate specific interaction hypotheses.

Datasets

The paper uses a subset of a parallel, multilingual corpus of French and Spanish Bible translations from the 12th, 16th, and 20th centuries. This corpus was semi-automatically annotated for word order in interrogatives, interrogative pronouns, verbs of speaking, and particles used with interrogatives. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Bridging the Gap: ParHistVis effectively bridges the gap between the rich information contained in parallel historical corpora and the analytical capabilities of researchers, especially those with limited computational expertise.
  • Multi-faceted Visualization: The combination of matrix, aggregated matrix, streamgraph, and Sankey diagram views provides a comprehensive toolkit for exploring different facets of linguistic change, from individual data points to complex interactions.
  • User-Centric Design: The tool's emphasis on being "free, easy-to-use, and interactive" is crucial for broader adoption in historical linguistics, where computational tools are not always standard.
  • Hypothesis-Driven Exploration: The Sankey diagram, in particular, empowers researchers to explore specific hypotheses about feature interactions, which is vital for understanding the complex drivers of language change, especially in sparse data.

Key Results

The paper demonstrates the utility of ParHistVis through a case study on word order change in Romance wh-interrogatives. The visualizations allowed the researchers to:

  • Observe the emergence and increasing frequency of complex inversion in Modern French (whSNPVSCI stream).
  • Identify the stable frequency of diachronic non-adjacency of the wh-element and the verb when a particle is present (whPtcVS stream).
  • Gain insights into how certain interrogative pronouns (e.g., "why") allow for more variation in sentence structure and particle usage in both Spanish and French.

The tool's effectiveness is shown by its ability to facilitate "at-a-glance observations" and enable researchers to "recognize and evaluate patterns in an otherwise too multifactorial dataset."

Limitations and Future Work

The authors acknowledge that the tool's effectiveness relies on the quality and annotation of the input data. While the tool facilitates the observation of patterns, it does not make predictions about potential interactions. Future work could involve further development of automated pattern detection or more sophisticated statistical integration within the visualization framework. The paper also implicitly suggests that expanding the range of linguistic phenomena that can be visualized would be beneficial.

Summary updated

BibTeX
@inproceedings{kalouli2019parhistvis,
  author = {Aikaterini-Lida Kalouli and Rebecca Kehlbeck and Rita Sevastjanova and Katharina Kaiser and Georg A. Kaiser and Miriam Butt},
  title = {{ParHistVis: Visualization of Parallel Multilingual Historical Data}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure