TL;DR
This paper introduces a framework to link informal social media posts (tweets) with formal news articles, even when they are written in different languages. By applying non-linear manifold alignment (Procrustes analysis) to word embeddings, the authors successfully bridge the "vocabulary gap" between slang-heavy tweets and structured journalism, outperforming traditional topic-modeling baselines.
What Problem Does This Paper Address?
When a major event breaks, information flows through two very different channels: traditional news outlets and social media platforms like Twitter. Linking these sources is valuable—it allows editors to gauge public opinion on specific facts and helps users find context for short, noisy tweets.
However, two major hurdles exist:
- The Vocabulary Gap: Tweets are riddled with slang, abbreviations, and informal syntax, whereas news articles are diligently edited. Even if they discuss the same topic, the statistical distribution of their words differs wildly.
- The Language Barrier: News is global and multilingual. A user tweeting in German might be reacting to an English news report, but most existing linking systems are restricted to a single language.
Methods
The authors propose Transformed Word Embeddings (TWE) to align the semantic spaces of these two domains.
Monolingual Transformation
First, they generate separate word embeddings (using GloVe) for a tweet corpus and a news corpus. To bridge the gap, they use Procrustes analysis, a manifold alignment technique that transforms the tweet embeddings to match the news embeddings through three steps:
- Translation: Centering the datasets at a common origin.
- Scaling and Rotation: Using an orthogonal matrix to minimize the distance between common words in both sets (e.g., ensuring the vector for "crisis" in tweets aligns with "crisis" in news).
Cross-Lingual Transformation
To handle different languages (e.g., English tweets vs. German news), they use a two-step approach:
- Bilingual Embeddings: They use the Bilbowa tool to learn a shared semantic space for news articles in two languages (English and German) using parallel corpora like Europarl.
- Alignment: They then apply the Procrustes transformation to map tweets from language A into the bilingual news space of language A, which inherently provides a bridge to language B.
Datasets
The study utilized several collections for training and evaluation:
- Training Corpora: Large-scale collections of English and German news (over 1.2M documents combined) and tweets (approx. 167k documents). The specific sizes are detailed in the table below:
- Dataset-1 (New): A gold-standard dataset created by the authors focusing on the "Grexit" topic. It includes 18 tweets and 18 news articles in both English and German, with 628 human-annotated relevance judgments.
- Dataset-2 (Existing): A dataset from Krestel et al. [4] containing 1,600 binary relevance judgments across 17 news topics.
Human annotators rated pairs on a scale of 0 to 2, as defined here:
Key Insights and Ideas
- The "Nearest Neighbor" Disconnect: Before transformation, the authors found that the top 10 neighbors of a word in tweet-space were very different from its neighbors in news-space. Their Rank Biased Overlap (RBO) analysis confirmed that word usage is significantly distinct across these platforms.
- Non-Linearity Matters: While previous work tried linear transformations, this paper argues that handling disturbances like scaling and rotation through manifold alignment is crucial for capturing the relationship between informal and formal text.
- T2N vs. N2T: The authors tested transforming tweets to news (T2N) and vice versa (N2T). Interestingly, T2N generally yielded slightly better performance, suggesting it is more effective to project noisy data into a structured space than the other way around.
Key Results
- Monolingual Superiority: On Dataset-2 (English), the TWE approach combined with an SVM classifier achieved 83.1% accuracy, outperforming the "Boosting" baseline (82.5%) and significantly beating LDA-based topic models (79.1%).
- Cross-Lingual Success: In the challenging task of linking English tweets to German news, the TWE method achieved a Pearson correlation ($r$) of 0.1181, nearly 44% higher than the LDA-PTM baseline ($r = 0.0821$).
- Refinement: The transformation improved the RBO measure (a proxy for alignment quality) by 24.4% for English and 21.2% for German.
Limitations and Future Work
The authors acknowledge that while their method outperforms baselines, the absolute Pearson correlation values remain relatively low (around 0.1). This highlights the extreme difficulty of pairwise similarity tasks between such disparate text types.
Additionally, the evaluation was performed on a relatively small human-annotated dataset. Future work aims to scale these experiments to larger datasets and explore even more sophisticated non-linear transformation techniques to further close the vocabulary gap.
Summary updated