Better Together: Combining Language and Social Interactions into a Shared Representation

Yi-Yu Lai, Chang Li, Dan Goldwasser, Jennifer NevilleTextGraphs @ NAACL 2016 · DOI

TL;DR

This paper tackles the disconnect between analyzing social network structures and the language used within them. They propose a novel method to create a unified representation that captures both social connections and textual content, leading to improved performance on social relationship prediction tasks.

What Problem Does This Paper Address?

Social network analysis and natural language processing (NLP) are powerful tools for understanding human interactions, but they've largely been studied in isolation. Social network analysis focuses on the structure of connections (who is connected to whom, how often), while NLP focuses on the content of communication (what people are saying). This paper argues that a richer understanding of social interactions can be achieved by combining these two perspectives. The challenge is to find a way to represent both the social graph and the textual content in a single, coherent framework, rather than treating them as separate, independent signals.

Methods

The core idea is to create a "TS-Infused Social Graph" where the strength of connections (edges) between nodes is modified based on the textual similarity of the messages exchanged. This is achieved in a few key steps:

  1. Textual Content Representation: They explore two ways to represent the text associated with interactions:
  2. Topic Models: Representing documents as distributions over topics.
  3. Word Embeddings: Representing documents as the average of word embeddings within them.
  4. Edge Weighting: The cosine similarity between the textual representations of neighboring nodes is used to weight the edges in the social graph. This means stronger textual similarity leads to a stronger social connection in their model.
  5. Node Embedding: They then use the LINE (Large-scale Information Network Embedding) algorithm, a standard technique for learning low-dimensional vector representations of nodes in a graph. Crucially, they modify LINE to use the text-weighted edges instead of just raw interaction counts. This allows the embedding to capture both social proximity and textual similarity.
  6. Feature Extraction: For downstream tasks, they extract features from these node embeddings. This includes various distance-based similarity measures between node vectors and additional features like unigram representations or word embeddings of the text sent by each node.

Datasets

The paper evaluates their method on two datasets:

  • Purdue Facebook Network: This dataset consists of public Facebook network data from March 2007 to March 2008, including approximately 3 million post activities. They specifically extract mutually "top friend" pairs and one-way "top friend" pairs for their experiments.
  • Avocado Email Collection: This dataset comprises 279 email accounts from which job titles and departments were extracted. Accounts are grouped into three categories (executives, engineering, business), and the task is to predict if two accounts belong to the same group.

Key Insights and Ideas

  • Synergy of Social and Linguistic Signals: The most significant insight is that combining social network structure with textual content leads to better performance than using either alone. This suggests that social interactions are a product of both who we know and what we talk about.
  • Textual Similarity as a Social Signal: Using textual similarity to weight social graph edges is a clever way to inject linguistic information directly into the social graph structure, making the graph representation more nuanced.
  • Context Matters for Prediction Tasks: The paper highlights how the best approach can depend on the specific prediction task. For distinguishing strong vs. weak friendships (Facebook dataset), messages exchanged between individuals are highly indicative. For predicting group membership within a company (Avocado dataset), the broader interactions with colleagues seem more important.

Key Results

The paper reports that their TS-Infused Social Graph approach consistently outperforms baseline methods that only consider network structure or text independently.

  • Facebook Dataset: Their method, particularly using Topic Models (GETM) and Word Embeddings (GEWE) to infuse the social graph, shows significant improvements over baseline embeddings. For example, GETM achieves an F1 score of 82.69% when considering network connectivity and topic models, compared to 75.09% for a network-only approach.
  • Avocado Dataset: Similarly, their infused graph representations show advantages. GETM achieves an F1 score of 66.84% for predicting group membership, outperforming the baseline network-only embedding (65.75%).

Limitations and Future Work

  • Computational Cost: While not explicitly detailed as a limitation, embedding large graphs and processing extensive text data can be computationally intensive.
  • Choice of Embedding Method: The paper uses the LINE embedding technique. Future work could explore other graph embedding methods.
  • Task-Specific Optimization: The authors suggest that further tailoring the embedding and feature extraction process to specific downstream tasks could yield even better results. They also mention exploring different ways to aggregate textual content (e.g., per-message vs. aggregated per-user).

Summary updated

BibTeX
@inproceedings{lai2016better,
  author = {Yi-Yu Lai and Chang Li and Dan Goldwasser and Jennifer Neville},
  title = {{Better Together: Combining Language and Social Interactions into a Shared Representation}},
  booktitle = {TextGraphs @ NAACL},
  year = {2016},
  doi = {10.18653/v1/w16-1405},
}
Figure