Abstract In this work, we present to the NLP community, and to the wider research community as a whole, an application for the diachronic analysis of research corpora. We open source an easy-to-use tool coined: DRIFT, which allows researchers to track research trends and development over the years. The analysis methods are collated from well-cited research works, with a few of our own methods added for good measure. Succinctly put, some of the analysis methods are: keyword extraction, word clouds, predicting declining/stagnant/growing trends using Productivity, tracking bi-grams using Acceleration plots, finding the Semantic Drift of words, tracking trends using similarity, etc. To demonstrate the utility and efficacy of our tool, we perform a case study on the cs.CL corpus of the arXiv repository and draw inferences from the analysis methods.
Introduction
Language is dynamic, shifting alongside societal and technological advancements. In scientific research, this "diachronic" (across time) change is particularly rapid. A researcher reading a paper from 1994 might encounter the word "model" and find it refers to something entirely different than it does in 2024. While methods exist to analyze these shifts, they often require significant coding effort and complex data pipelines. DRIFT addresses this gap by providing an accessible, unified toolkit for visualizing research trajectories and identifying emerging or stagnant fields.
System Architecture and Methods
DRIFT is built using the Streamlit framework, offering an interactive web interface with two primary modes: Train Mode for preprocessing new corpora and Analysis Mode for generating visualizations.
The core engine utilizes several advanced techniques:
- Temporal Word Embeddings with Compass (TWEC): Ensures year-by-year word vectors exist in a comparable space by training a "compass" on the entire dataset.
- Productivity and Frequency Analysis: Categorizes terms as Growing, Consolidated, or In Decline based on entropy.
- Acceleration Plots: Measures the semantic shift between word pairs (e.g., the convergence of "Attention" and "Translation").
- YAKE & LDA: Used for automated keyword extraction and topic modeling to identify dominant themes.
Case Study: arXiv cs.CL and cs.CV
To demonstrate DRIFT, the authors analyzed 27,384 abstracts from the arXiv cs.CL (Computation and Language) and cs.CV (Computer Vision) subdomains spanning 1994 to 2021. Abstracts were chosen over full-text PDFs to minimize noise from equations and jargon while capturing the core contributions of the research.
Key Insights: The Evolution of "Model"
One of the most compelling demonstrations of DRIFT is the visualization of "Semantic Drift." For instance, the term "model" has undergone a radical shift in meaning within the NLP community.
In the late 1990s, "model" was closely associated with statistical methods like Bayesian, Markov, and HMM. By 2021, its semantic neighborhood shifted entirely toward transformer-based architectures like BERT, RoBERTa, and XLNet. Such visualizations allow researchers to identify "quiescent" topics that might be due for a modern revival or validate current "hype" cycles, such as labeling "BERT" as a growing term and "LSTM" as a declining one.
Conclusion and Future Work
DRIFT provides an efficient, open-source solution for large-scale literature reviews. By integrating the faiss library, it achieves clustering speeds up to 10 times faster than standard implementations. While currently focused on abstracts, future iterations aim to include full-text analysis, citation-based metrics, and statistical knowledge graphs to further enhance interpretability across diverse scientific domains.
Summary updated