Introduction
In Information Retrieval (IR), term weighting is the process of determining which words in a document are the most significant. Traditionally, this task is approached via statistical methods like TF-IDF, which are robust but treat words as an "unordered bag," or linguistic methods that require extensive manual effort to define rules for every edge case. This paper identifies a "missing middle": a system that can automatically learn how words influence one another based on their context—order, distance, and relationship—without requiring hand-coded linguistic rules.
Methodology
The authors propose a recursive weighting scheme inspired by graph-based algorithms like TextRank. In this model, a document is treated as a social network of words where a word's importance is determined by both its intrinsic value and the "prestige" of its neighbors.
Recursive Weighting Scheme
The importance of a word $w$ is defined recursively: $$s(w) = (1 - \lambda)d_w + \lambda \sum_{u \in r(w)} s(u)c_{w,u}$$ Where $d_w$ is the "default" score (e.g., its TF-IDF value), $s(u)$ is the score of a neighboring word, and $c_{w,u}$ is the Context Function, which measures the influence word $u$ has on word $w$. The damping factor $\lambda$ balances intrinsic importance against contextual influence.
Learning the Context Function
Unlike previous models with fixed relationship values, this approach uses a parametric Context Function learned from examples. By modeling features like word distance and global rarity (IDF) through "soft switches" (logistic sigmoids), the system tunes its understanding of context. To enable training, the authors developed softMRR, a differentiable version of the Mean Reciprocal Rank metric, allowing for standard gradient-based optimization.
Experimental Setup
The algorithm was evaluated on a single-word question-answering task using 525 Italian crossword clues. The dataset included 165 clues where the answer is a Named Entity (NE) and 360 clues for common nouns, verbs, or adjectives (nonNE). Documents retrieved via web search for each clue served as the corpus for term weighting.
Results and Discussion
The context-based approach consistently outperformed standard baselines. For Named Entity questions, the Mean Reciprocal Rank (MRR) jumped from 0.216 (TF-IDF) to 0.355 with the full context model—a 64% relative improvement.
As shown in the performance table below, the addition of context features (FS-B and FS-B*) significantly enhanced results across different base weighting schemes.
The system also demonstrated a higher success rate at various rank positions. For NE questions, the probability of finding the correct answer within the top 50 candidates rose from 60.4% to 83.3%. The model proved robust even when the number of retrieved documents varied, suggesting that the learned Context Function captures general properties of language rather than over-fitting to a specific dataset size.
Conclusion
By treating documents as graphs and learning the influence of word proximity, this adaptive algorithm successfully bridges the gap between statistical and linguistic IR methods. The use of a differentiable ranking metric allows the system to be tuned for specific tasks, achieving state-of-the-art results in crossword solving. Future work may explore replacing the parametric context function with neural networks to capture deeper semantic relationships.
Summary updated