Smoothing Blog Posts Toward Their Author's Persistent Vocabulary

Finding the blogs that are about a topic is a different task from finding pages that mention it. Given a two-word query like "home baking", a feed-search system should return the dozen blogs that keep writing about baking. It should not return the news blog that ran five baking stories during one holiday week. Systems built for the TREC Blog track handle this in one of two ways. Some glue a blog's entire history into one giant document. Others score each post separately and let the posts vote for their blog. Both struggle with the same problem: the query terms are often missing from an individual post even when the blog is plainly on topic. The usual fix is to pad the query with terms mined from Wikipedia. That fix is known to drift off topic on about a third of queries.

We propose a fix that works inside the probability model and requires no outside text. In language-model retrieval, each post gets a word distribution, and "smoothing" fills in probability for words the post happens not to contain. We would smooth each short post toward a model of its own blog's vocabulary, rather than toward the whole collection. That blog model would count only the words the author uses steadily over time and would discount words that appear in a single burst. A second, separate smoothing step against the collection would then down-weight common words. The blog structure thus supplies, at no extra cost, the "neighborhood" that document-expansion methods have to compute expensively. Weighting by persistence moves the distinction between topical authority and viral bursts out of a post-hoc rank multiplier and into the probability estimates themselves.

Core Claim

The claim is that a two-stage post model is enough to bridge most of the vocabulary gap in feed search without external expansion, and to separate sustained authorities from burst-driven generalists. In the first stage, a Dirichlet prior pulls the post toward a persistence-weighted model of its own feed. In the second stage, linear interpolation mixes in the collection background. Concretely, for post $p$ in feed $f$, split $f$'s history into $T$ equal time bins. Let $\pi_f(w)\in[0,1]$ be the normalized entropy of word $w$'s counts across those bins: it is 1 for perfectly steady use and near 0 for a single burst. The model is

$$P_{\text{pers}}(w\mid f)\propto (\epsilon+\pi_f(w))^{\gamma}\,c(w,f),\qquad P(w\mid p)=(1-\lambda)\frac{c(w,p)+\mu P_{\text{pers}}(w\mid f)}{|p|+\mu}+\lambda P(w\mid C).$$

Here $c(w,\cdot)$ is a raw count, $|p|$ is post length, $\epsilon$ is a small floor so that rarely used words are not zeroed out, and $\gamma$ controls how hard bursts are discounted. The parameter $\mu$ handles the post-sparsity (estimation) role and $\lambda$ handles the common-word (query-modeling) role, following the decoupling of Zhai and Lafferty. Feeds are then scored by aggregating their top-scoring posts, as in existing voting models.

What is new is the specific per-term structure. A post that contains "baking" but not "home" can borrow probability for "home" from its author, but only if the author uses "home" persistently. Fusing a macro-document score with a voting score after the fact cannot do this, because that fusion multiplies whole-query scores rather than individual term probabilities.

Relation to Prior Work

Two papers supply the method. Azzopardi et al. smooth each document toward a document-specific LDA topic prior and then interpolate with the collection. They found that the topic prior helped only inside this two-stage form. Tao et al. smooth each document toward its cosine-nearest neighbors (an approach called DELM) and report the largest gains on short documents. We keep the two-stage structure that both papers depend on. We replace the learned or computed prior with the authorship grouping the collection already provides, and we filter that prior by time. Both papers were evaluated only on static ad-hoc collections and never on feed aggregation.

Several blog-track papers motivate the work by documenting the gap. Elsas et al. and Arguello et al. show that macro-documents plus Wikipedia anchor-text expansion work well. Arguello et al. also report that popularity bias hurts roughly one-third of queries. Ernsting et al. show that external expansion can push out the original query terms (their "lactose" failure), and that persistence over time beats recency as a signal of authority. Xu et al. turn persistence into an entropy-based "Recurring-Degree" rank factor. Our $\pi_f$ borrows that entropy idea but applies it per word, inside the estimate.

Two further papers constrain the design. Weerkamp et al. show that post-index voting with a cheap pre-filter is both accurate and efficient. Our model must therefore fit into that pipeline rather than require re-indexing macro-documents. Their "two-stage" refers to a filter-then-rank architecture, not to smoothing, so the two are orthogonal. Smucker and Allan show that once term weighting is handled elsewhere, the best Dirichlet $\mu$ falls by half. That result gives us a mechanistic prediction to test. Shakery and Zhai found that two-stage smoothing beat Dirichlet for web topic distillation, but that link propagation amplified noise when relevant items were sparse. We therefore propagate evidence only along authorship links, which are dense and reliable, and not along hyperlinks.

Research Program

We would build the post model above on the TREC Blog06 post index, using title and full permalink text, since Arguello et al. found permalink text matters. Each feed's persistence-weighted vocabulary would be precomputed in one pass over its posts. The cost is linear in collection size, compared with the pairwise similarity computation DELM needs. Feed scores would come from the Blogger-style aggregation of Weerkamp et al. over a pre-filtered candidate set. As a second aggregator, we would sum the top $n$ posts, following Ernsting et al., who found a peak at $n=8$. This checks that the effect does not depend on one aggregator.

Beyond overall accuracy, the analysis would test three mechanisms. The first is the lexical gap: on queries where relevant feeds' top posts lack at least one query term, gains should concentrate there. The second is bursts: the gain should show up as non-relevant feeds whose matching posts fall into a single time window sliding down the ranking. The third is the smoothing-parameter shift: with $\lambda>0$, the best $\mu$ should be lower than in single-stage Dirichlet, mirroring Smucker and Allan.

Baselines

The do-nothing reference is single-stage Dirichlet smoothing against the collection on the same post index, with the same aggregation and with $\mu$ tuned. This shows whether anything beyond standard query likelihood helps.

The strongest existing system on TREC 2007 feed topics is the Title + Full post-index model of Weerkamp et al., which reports MAP 0.3556. On TREC 2008, the strongest is cmuLDwikiSP from Arguello et al.: a macro-document model with Wikipedia anchor-text expansion and spam filtering, reporting MAP 0.306 and P@10 0.434. Their small-document baseline cmuSD reports MAP 0.246. The Ernsting et al. MAP of 0.4741 is for post-level topical retrieval, not feed distillation, so it is not a comparable target. Elsas et al.'s top run CMUfeedW has no MAP figure in the file.

Because the published runs used different indexes and tuning, we would reimplement both systems on our index. All methods would get the same tuning budget, with parameters selected on the 2007 topics and tested on 2008, and vice versa. The published numbers serve only to check that the reimplementations are faithful. Beating a baseline means a statistically significant MAP gain under a paired test on both held-out topic sets, plus no loss in P@10. Matching the reimplemented Arguello system without using Wikipedia would already be a meaningful result.

Three ablations isolate the key ingredient:

  1. Unweighted feed prior ($\gamma=0$). This tests whether persistence weighting matters or whether smoothing toward the author is enough.
  2. Persistence as a post-hoc multiplier. Here the Recurring-Degree factor of Xu et al. is applied to the single-stage baseline's feed scores.
  3. Late fusion. This interpolates the reimplemented macro-document feed score with the voting score, tuned to the same budget.

Why not just…?

Why not just expand the query with Wikipedia anchor text? It works well on average and produced the strongest 2008 run. However, Arguello et al. report that popularity bias hurts about a third of queries, and Ernsting et al. show that expansion can push out the core query terms. Our query is never rewritten, so that failure cannot occur. This alternative is a genuine threat on average MAP. We would compare per-query win/loss counts against the reimplemented expansion system, and also run expansion on top of our model. If expansion still adds large gains there, our method is complementary rather than a replacement, and we would report it that way.

Why not just tune Dirichlet on macro-documents? Macro-documents are so long that Dirichlet smoothing barely affects them. Any gain from them may come from length effects rather than topical focus, which is the open question the file raises about Arguello et al. Monolithic macro-documents also blend burst vocabulary with steady vocabulary, and they must be re-indexed whenever a feed updates. The late-fusion ablation is the direct test. If tuned fusion matches our in-model smoothing, the per-term claim fails and we would say so.

Why not just apply DELM or an LDA prior to every post? Tao et al. and Azzopardi et al. are the closest methods, but they need pairwise neighbors or topic inference over 3.2 million posts. Neither was tested beyond ad-hoc collections of a few thousand to hundreds of thousands of documents. We would run DELM on the pre-filtered candidate posts only, where its cost is manageable. If cosine neighbors beat same-author neighbors there, the authorship prior is just a cheap approximation, and the contribution narrows to efficiency plus burst handling.

Why not just multiply in a persistence score after ranking? This is ablation 2 above. If it equals the full model, then putting persistence inside the smoothing buys nothing, and the burst-related part of the claim is falsified.

Proposed Experiments

The main experiment is Blog06 feed distillation on the TREC 2007 (45 topics) and 2008 (50 topics) sets, with cross-year tuning, reporting MAP, P@10, and per-query win/loss against every baseline.

The first decisive check targets bursts. We would take judged non-relevant feeds whose retrieved posts fall in a single time bin and compare their mean rank under the full model against $\gamma=0$ and against the post-hoc multiplier. The full model should push these feeds down further than either.

The second decisive check targets the vocabulary gap. We would split queries by whether relevant feeds' top posts contain every query term. Gains should come mostly from the partial-match group, where per-term borrowing applies and late fusion cannot help.

The third check repeats Tao et al.'s shortening test, truncating posts to 30% of their length. If the author prior is doing the bridging, its advantage should grow as posts get shorter.

The fourth check verifies the predicted drop in the best $\mu$ once $\lambda$ is free to vary.

Finally, we would run the faceted distillation topics on Blogs08, which is about ten times larger, using the 2,500-post candidate pruning of Xu et al.. No reported MAP was found on Lacuna for that system, so this run tests scalability rather than serving as a numeric target. If the full model loses to Wikipedia expansion on average but cuts the expansion's failure rate on the bad third of queries, the honest conclusion is that the two should be combined, and the paper would report that combination.