How analysis of mobile app reviews problematises linguistic approaches to internet troll detection

Research on trolling spans early computer-mediated communication and hate speech studies through contemporary analyses of state-sponsored disinformation. A pressing task is accurate, rapid identification of troll accounts using linguistic signals. Prior work (Monakhov, 2020a,b) argues trolls’ discourse is constrained by two orthogonal pressures: delivering a target message repeatedly while concealing this intention. This leads not to simple repetition, but to anomalous distributions of repeated words and especially word pairs, because target words are deployed across unusually varied contexts. Monakhov proposed a simple metric, the troll coefficient q, defined as the ratio of the proportion of repeated content words to the proportion of repeated content word pairs; trolls tend to have higher q because non-troll writing exhibits more repeated word pairs (higher denominator). The algorithm exceeded 98% accuracy on tweets. Unexpectedly, when applied to app reviews, both suspicious positive reviews and genuine one-star reviews were flagged as troll-like. This suggests either domain limitations of the metric (unlikely) or that the communicative setting of reviews induces a collective “troll effect” via repatterning: writers adapt their wording after reading prior reviews. The study tests the hypothesis that troll coefficients of app reviews can be explained as a function of creation time, predicting a negative correlation between troll coefficient and temporal distance within review groups (recency shaping lexical choices). The paper presents observational analyses of >180k reviews (Study 1) and an experiment manipulating access to prior comments (Study 2).

The paper situates trolling within decades of scholarship on computer-mediated communication, hate speech, and contemporary propaganda/disinformation. Foundational works examined group processes, anonymity, and safety online; recent studies profile state-sponsored trolls and compromised accounts. Monakhov (2020a,b) provides the core linguistic framework for troll detection via distributional anomalies and introduces the troll coefficient (repeated words vs. repeated word pairs). The study also connects its findings to language creativity literature (heteroglossia, schema refreshing, dialogicality) to interpret observed collaborative repatterning in online commentary, noting tensions with standard priming expectations and highlighting a recency-driven, discourse-level effect.

Study 1: Data and preprocessing. The authors collected all one-star reviews for two health apps: Moscow’s Social Monitoring app (4072 Russian reviews) and the UK NHS app (1987 English reviews). For each review, they extracted stemmed content words and precise timestamps. Reviews were chronologically aligned from most recent to oldest. Sampling procedure: (1) initial window of 250 reviews formed S1; (2) window size incremented by one at each step, forming sets S1…S(n−250); (3) from each set, 250 reviews were randomly sampled to standardize sample size. For each sample, the average pairwise time distance (in days) among reviews was computed. Troll coefficient q was computed per sample as the ratio of the proportion of repeated content words to the proportion of repeated content word pairs among all content words/pairs in the sample. Distributions were compared for chronologically aligned vs randomly shuffled reviews. To generalize beyond two apps, the authors randomly selected 59 apps across categories and downloaded 3000 English reviews per app (all ratings). The same sampling and metric computation were applied. Correlations between average time distance and troll coefficient were computed per app; some time-sparse apps (>300 days average between adjacent reviews) showed positive correlations. Modeling. A causal diagram motivated control of confounders: sample range (index of sampling window), rate of change (review arrival rate), and average number of words per review. A Bayesian multiple linear regression predicted troll coefficients from predictors: sample range, rate of change, average time distance, number of words, plus interactions of sample range with rate of change and with average time distance. Noninformative priors were used; JAGS drew MCMC samples (3 chains; 2000 adaptation; 5000 burn-in; 50,000 posterior iterations). Posterior predictive checks compared replicated to observed distributions; the trained model on 7 apps (spanning negative to positive correlations) was evaluated out-of-sample on the remaining 52 apps. Study 2: Experiment. Russian-speaking participants described a single image in one or two sentences under two conditions: (1) no access to others’ comments (n=192), and (2) optional access to read prior comments on a scrolling webpage before posting (n=193). Recruitment used Yandex.Toloka; each participant submitted one comment within 10 minutes and was compensated $1. Comments required at least two content words. Data integrity checks matched platform and webpage logs. For analysis, sampling windows were 50 comments (due to smaller N). Hypotheses: (H1.1) higher troll coefficients in condition 2 vs 1; (H1.2) in earliest-to-latest order, positive association between sample index and troll coefficient in condition 2 but not condition 1; (H1.3) in latest-to-earliest order, an inverted U-shaped relation in condition 2 but not in condition 1. First-order linear models (troll coefficient ~ sample index) tested H1.2; second-order polynomial models tested H1.3. Group differences were tested with t-tests adjusted for content word counts.

Study 1 observational results: In chronologically aligned samples of one-star reviews, troll coefficients strongly negatively correlated with average temporal distance within samples: Russian r = -0.87, p < 0.0001; English r = -0.84, p < 0.0001. Random shuffling removed this dependence, producing uniform distributions. Across 59 apps (3000 reviews each), some apps exhibited positive correlations, particularly when reviews were very time-sparse (average separation > 300 days). Bayesian regression: All predictors had 95% probability intervals excluding zero, indicating usefulness for prediction. Reported mean coefficients (with 95% PIs) included: Intercept ≈ 6.173 (6.129, 6.217); Sample range ≈ 1.873e-04 (1.580e-04, 2.167e-04); Rate of change ≈ -9.450e-02 (-0.1336, -0.0553); Average time distance ≈ -2.855e-04 (-4.065e-04, -1.644e-04); Number of words ≈ -0.2010 (-0.2047, -0.1972); Sample range × Rate of change ≈ -3.794e-04 (-4.255e-04, -3.334e-04); Sample range × Average time distance ≈ -1.104e-06 (-1.236e-06, -9.724e-07). Posterior predictive checks indicated reasonable fit; out-of-sample predictions for 52 apps captured major distributional patterns despite errors. Interpretation: when average time distance and rate of change are at reference values (0.5 and 0.002), sample range increases troll coefficient, suggesting greater adaptation when discourse is perceived as continuous. Study 2 experimental results: Troll coefficients were significantly higher in the condition with access to prior comments than without: earliest-to-latest alignment t(257.59) = -37.2, p < 0.0001; latest-to-earliest alignment t(262.88) = -46.7, p < 0.0001 (values indicate lower coefficients in condition 1). In earliest-to-latest order, condition 2 showed a significant positive association between sample index and troll coefficient, while condition 1 showed no association. In latest-to-earliest order, condition 2 exhibited the predicted inverted U-shaped relation; condition 1 showed none. Overall, both observational and experimental data confirm that troll coefficients decrease as temporal distance increases and that visibility of prior text induces lexical diversification that mimics troll-like patterns.

The findings support the hypothesis that collaborative repatterning in online reviews/comments drives systematic variation in the troll coefficient as a function of temporal proximity and perceived conversational continuity. When contributors can see recent prior messages, they avoid verbatim repetition and seek novel lexical choices, increasing repeated single-word rates relative to repeated word-pair rates, thereby inflating q in ways similar to troll discourse. This is a discourse-level, time-dependent phenomenon, likely influenced by a recency effect that counters typical priming expectations. Conceptually, it aligns with notions of heteroglossia and everyday language creativity: participants strategically deploy resources to distinguish their contributions in a collaborative space. Consequently, linguistic troll-detection algorithms that rely on repetition distributions risk false positives in settings characterized by dense, ongoing commentary. However, by modeling expected q from temporal structure (time distance, rate of arrival) and text length, one can adjust for collaboration-induced effects, improving interpretability and reducing misclassification.

Across >180,000 app reviews and a controlled experiment, the study shows that the troll coefficient of grouped short texts is a function of temporal structure and access to prior contributions. When discourse is perceived as continuous—via short time distances or slow arrival rates—contributors engage in lexical diversification that elevates troll-like signals. A Bayesian regression using sample range, rate of change, average time distance, and text length predicts q reasonably well and generalizes across apps. Practically, this complicates linguistic troll detection by revealing a benign, collaboration-driven source of troll-like patterns. Nonetheless, modeling expected q from time alone offers a baseline to differentiate collaborative repatterning from manipulative behavior. Illustrative cases in five-star app reviews show deviations from predicted trends that may indicate coordinated fake activity (creative rewriting or reposting), warranting further investigation. Future work should refine dynamic models of q, test broader platforms and languages, and probe underlying cognitive mechanisms of recency-driven lexical choice.

• Domain and language scope: Observational data focus on Google Play app reviews; languages are primarily English and Russian. Generalizability to other platforms, genres, and languages remains to be validated.
• Model simplifications: The linear regression is an oversimplification of a likely more complex, dynamic process; assumptions include a constant review arrival rate per app and linear interaction forms.
• Measurement choices: The troll coefficient depends on content word identification and stemming; methodological choices may affect repetition counts across languages and apps. Fixed sampling windows (250 for Study 1; 50 for Study 2) may influence estimates.
• Experimental constraints: The experiment used a single image, Russian native speakers on one platform, and did not directly measure the extent of participants’ engagement with prior comments beyond condition assignment.
• Mechanism inference: Psychological mechanisms (e.g., recency effects vs. priming) are hypothesized but not directly tested; the study is not designed to establish cognitive causality.
• Detection implications: While temporal adjustment can reduce false positives, the approach does not definitively separate trolls from collaborative repatterning and may require integration with non-linguistic signals.