TL;DR
This paper investigates how the structure of online communities influences semantic change in words. By analyzing 45 Reddit communities and their social networks, the researchers found that while individual community features like size or stability don't predict semantic change, a complex three-way interaction between these features, alongside word frequency and its change, significantly impacts how word meanings evolve. This work offers a novel computational approach to understanding semantic change within specific social contexts.
What Problem Does This Paper Address?
The study of semantic change, how word meanings evolve over time, has a long history. However, traditional linguistic approaches often treat meaning at a language-wide level. This paper highlights a crucial insight from linguistics: a word's meaning is fundamentally tied to the community in which it's used. This raises a key question: how does semantic change differ across different communities, and what community characteristics drive these differences? Existing sociolinguistic work often focuses on individual social networks (ego networks) and linguistic variation (different forms for the same meaning), rather than semantic shift (changes in meaning for a single form). This paper bridges this gap by using computational methods to quantify semantic shift at the community level and explore its relationship with community structure.
Methods
The core of this research lies in two main computational techniques:
- Distributional Semantics for Semantic Change: The authors employ a diachronic skip-gram with negative sampling (SGNS) model, a standard technique for learning word embeddings from text. They train models on word usage data from two different time periods (one year apart) within specific communities and a generic corpus. By comparing the word vectors from these two periods, they can quantify semantic change.
- Naïve and Rectified Cosine Change: Initially, they measure semantic change using "naïve cosine change," which is the angular distance between word vectors in the two time periods. However, recognizing that this metric can be biased by word frequency and context variability, they introduce a "rectified change" score. This score is a t-statistic that controls for noise by comparing the observed change to a null distribution derived from shuffled corpora. This is a key methodological contribution, providing a more robust measure of genuine semantic shift.
- Social Network Analysis: They construct social network graphs for each community based on user interactions (comments). From these graphs, they derive community-level features like clustering coefficient (how interconnected nodes are within a network), size (number of active members), and stability (similarity of membership between two time periods).
Datasets
The primary dataset consists of comments from 45 randomly selected subreddits from Reddit, active between 2015 and 2017. They also created a generic Reddit corpus by sampling comments from all subreddits. For each community and the generic corpus, they constructed separate datasets for 2015 and 2017, creating a one-year gap for diachronic analysis. This dataset is novel in its combination of social media text with detailed community network structures for studying semantic change.
Key Insights and Ideas
- Community Structure Matters, But Complexly: Individual community features like size, stability, or clustering coefficient, when considered in isolation, do not significantly predict semantic change. This suggests that semantic change isn't driven by a single factor but by a more intricate interplay.
- The Power of Three-Way Interactions: The most significant finding is the robust three-way interaction between community size, stability, and clustering coefficient. This indicates that the effect of one feature depends heavily on the levels of the other two, highlighting the nuanced relationship between community structure and semantic evolution.
- Frequency is a Strong Predictor: Word frequency and changes in frequency are consistently significant predictors of semantic change, aligning with previous research. However, this study is the first to demonstrate this effect while controlling for noise in semantic change measurement.
- Short-Term vs. Long-Term Change: While frequency's effect on semantic change is significant in the short-term (one-year period), the paper speculates that its influence might be smaller compared to long-term historical change, warranting further investigation.
- Rectified Change is Crucial: The introduction and use of the rectified change metric are vital. It effectively mitigates biases present in simpler measures like naïve cosine change, allowing for a more accurate assessment of genuine semantic shifts.
Key Results
- All three word-level features (frequency, change in frequency, and generic rectified change) were found to be significant predictors of semantic change.
- The three-way interaction between community size, stability, and clustering coefficient was significant (p < 0.014), indicating a complex interplay of these factors.
- Individual community-level features (size, stability, clustering) were not significant on their own.
- The paper reports that the three-way interaction between word-level predictors (frequency, change in frequency, and rectified change) was highly significant (p < 0.05), with a p-value too small to represent, indicating its strong influence.
Limitations and Future Work
The authors acknowledge several limitations:
- Single Vector Representation: The diachronic SGNS model assigns a single vector to each word, making it insensitive to polysemy (words with multiple meanings). Changes in the relative frequency of different senses might be reflected as semantic shift, even if individual senses haven't changed.
- Meaning-in-Use vs. Lexical Meaning: Distributional semantics captures meaning-in-use, but might not fully represent a word's potential meaning or common ground semantics that aren't explicitly activated in the corpus.
- Topic-Driven Change: Changes in community topics can influence word usage patterns without necessarily reflecting a true semantic shift (e.g., the word "2016" changing meaning as the year passes from future to past).
- Generalizability: The results are based on Reddit communities and a one-year timeframe. Generalizing these findings to other platforms or longer periods requires further research.
Future work suggested includes:
- Investigating the types of semantic change (e.g., broadening, narrowing).
- Exploring how communicative resources are used to create new word uses.
- Examining the circumstances of individual semantic changes, such as the role of central network positions in innovation.
- Investigating the correlation between semantic change and explicit word meaning negotiation within communities.
- Exploring contextualized word representations (like BERT) to better handle polysemy.
Summary updated