Which words win best answers and which questions spark debate online

On community question-answering platforms, two structural outcomes signal that a post has done its job: an answer is picked as “best,” and a question keeps drawing responses long after it is asked. A new exploratory study of Yahoo! Chiebukuro, Japan’s largest question-answering site, maps which words travel with each of those fates — and finds that the vocabulary of winning answers and the vocabulary of buzzing questions are largely different sets, drawn from opposite ends of the logistics world.

The study, published in Discover Networks by Hisatoshi Naganawa, analyzed a corpus of 40,871 questions and 78,641 answer-tier posts scraped from Yahoo! Chiebukuro using the Japanese keyword for “logistics.” The posts span April 2004 to July 2026, a 22-year window, and were tagged by structural role: 24,512 best answers, 36,314 other answers, 7,307 replies, and 10,508 thanks comments. Of all questions, 61 percent received at least one answer-tier post. Because the platform’s search interface caps retrievable results, the crawl was partitioned into one-day posting-date windows, each paginated to exhaustion, with duplicate question URLs removed before thread pages were fetched.

Every question and answer was machine-translated from Japanese to English using a local, offline OPUS-MT model run through a SentencePiece tokenizer and CTranslate2. The author, a native Japanese speaker, then manually reviewed 50 translated examples stratified across short, medium, and long posts. In 34 of the 50 pairs — 68 percent — all key content nouns survived translation, and in the remainder most did. Errors clustered in proper nouns and rare loanwords, and no translation mistake was found that changed a post’s topic. On that basis the pipeline was judged accurate enough for noun-level analysis, and artifacts such as contraction fragments and politeness formulas mis-tagged as nouns were filtered out as stop words.

The word-level analysis used the informative-Dirichlet-prior log-odds z-score of Monroe, Colaresi, and Quinn, a keyness statistic that uses the pooled corpus as its own background and penalizes rare words. For best-answer selection, the comparison was restricted to threads with at least two competing answers and an explicit best answer — 15,104 threads containing 35,980 passed-over answers. Because that pooled comparison weights threads by their number of answers, the headline nouns were re-tested with a within-thread permutation test: best and other labels were shuffled 500 times within each thread. For all fourteen reported nouns, the observed z-score exceeded every one of the 500 relabelings, reaching the permutation floor of p = .002.

The lexical signature of chosen answers was strikingly concrete. “Logistics” itself was the single strongest noun in either direction by a wide margin, at z = 38.1, followed by “distribution” (z = 7.7), “center” (z = 6.2), “warehouse” (z = 5.8), “base” (z = 5.4), “network” (z = 4.3), and “manufacturing” (z = 4.1). Passed-over answers, by contrast, were marked by generic or deferential vocabulary: “people” (z = −9.8), “job” (z = −7.6), “car” (z = −7.0), and “reference” (z = −7.0). Inspection of the Japanese originals showed that “reference” derived mainly from formulaic phrases such as “for your reference” and the platform’s quoted-answer footer, marking secondhand or deferential replies rather than literal citation.

Question engagement told a different story. Questions were split by total answers plus replies into a high group of 6,747 questions with four or more responses and a low group of 9,602 questions with one or none. High-engagement questions clustered around broad, debatable civic and infrastructure nouns: “car” (z = 6.5), “war” (z = 4.9), “tax” (z = 4.8), “party” (z = 4.7), “shinkansen” (z = 4.3), “road” (z = 4.2), “expressway” (z = 4.1), “child” (z = 4.1), and “highway” (z = 4.1). Low-engagement questions were dominated by narrow, transactional terms — “logistics” itself was the strongest low-engagement noun at z = −9.7, alongside “product,” “Taobao,” “company,” “warehouse,” and “center.” Taobao-related questions, on inspection, overwhelmingly concerned individual cross-border purchases: stalled parcel tracking, customs holds, forwarding agents, refund requests — single-answer queries rather than discussion starters.

To test whether these word patterns aggregate into coherent themes, the study trained a Word2Vec embedding model on the noun-tokenized corpus, reduced the vectors for the 800 most frequent nouns with PCA, and clustered them with HDBSCAN under a parameter rule fixed in advance. The chosen solution yielded twelve semantic clusters covering 302 nouns and 50.3 percent of noun-token frequency, with the largest cluster holding only 40 nouns. Validation checks showed every cluster’s mean pairwise cosine similarity exceeding random baselines by at least a factor of six, and an average adjusted Rand index of 0.82 across neighboring parameter settings. The partition’s main sensitivity was to PCA dimensionality, and roughly half of noun-token frequency remained unclassified background — including “logistics” itself, which co-occurs with everything in a keyword-filtered corpus and so has no locally dense neighborhood.

Cluster-level regressions then separated each theme’s association with each outcome from the simple effect of writing more. Best-answer selection was modeled with conditional logistic regression stratified by question, so each selected answer was compared only against its own thread competitors. Eleven of twelve clusters were associated with selection at FDR-corrected p < 0.05; the largest positive associations were infrastructure and strategy (odds ratio 1.34), nations and geopolitics (1.32), work and employment (1.28), delivery and ordering (1.26), and road and rail transport (1.26). The controls were informative in their own right: answer length dominated with an odds ratio of 2.59 per unit of log length, later answers were somewhat favored (1.14 per rank), and question-answer noun overlap was strongly negative at 0.13 — answers that merely restate the question’s vocabulary are passed over.

Engagement was decomposed into two stages so that the 16,104 questions without responses were not discarded. Whether a question was answered at all turned on largely different clusters: places and regions lowered the odds most (0.67), while cost, tax, and government was the only cluster raising them (1.17); seven clusters, including work and employment and delivery and ordering, were null. Once a question was answered, road and rail transport showed the largest positive association with response volume (an incidence rate ratio of 1.20), followed by cost, tax, and government (1.14), while the operational core clusters — work and employment (0.90) and delivery and ordering (0.92) — were negative. No cluster behaved the same way across all three outcomes; cost, tax, and government was the only one positive on all three.

The author is careful about what the design cannot show. Best-answer selection was measured on answer texts within threads, engagement on question texts across threads, so a cluster positive on one outcome and negative on the other is not evidence that a single post trades one for the other — the coefficients describe different documents written by different users for different purposes. All results are conditional associations, not causal effects: answerer reputation, question difficulty, and platform recommendation went unmeasured, usernames were not retained so repeat-poster dependence cannot be modeled, and the corpus covers a single keyword on a single Japanese platform. Because the asker chooses the best answer, the outcome reflects asker satisfaction rather than verified correctness. Within those limits, the picture is coherent: “answerability” and “discussability” appear to be distinct dimensions of content value, not opposite ends of one axis — concrete operational vocabulary helps an answer win its thread, while broad, debatable civic topics are what keep a conversation growing.

Subject of Research: Lexical and semantic-cluster associations with best-answer selection and question engagement in logistics discussions on a Japanese community question-answering platform

Article Title: Content associations with answer selection and engagement in supply chain community question answering

Article References: Naganawa, H. (2026). Content associations with answer selection and engagement in supply chain community question answering. Discover Networks, 2(1), Article 41. https://doi.org/10.1007/s44354-026-00055-w

Image Credits: AI Generated

DOI: 10.1007/s44354-026-00055-w

Keywords: community question answering, Yahoo! Chiebukuro, logistics, best-answer selection, user engagement, keyness analysis, Word2Vec, HDBSCAN, conditional logistic regression, machine translation, supply chain, text mining