The words that describe a topic, by four measures:
"prob": the most probable terms, \(\beta_{kv}\). Frequent words shared by several topics come first."frex": FREX (Bischof and Airoldi 2012; Roberts et al. 2014) balances frequency in the topic and exclusivity to the topic. With \(F_{kv}\) the rank share of \(\beta_{kv}\) among the terms of topic \(k\), and \(E_{kv}\) the rank share of \(\beta_{kv} / \sum_j \beta_{jv}\) (exclusivity), $$\mathrm{FREX}_{kv} = \left(\frac{w}{F_{kv}} + \frac{1 - w}{E_{kv}}\right)^{-1},$$ with \(w\) =frexw, the weight of frequency."score": the score ofstm::labelTopics()(after theldapackage), which down-weights terms probable in every topic: $$\mathrm{score}_{kv} = \beta_{kv}\left(\log\beta_{kv} - \frac{1}{K}\sum_j \log\beta_{jv}\right).$$"lift"(Taddy 2012): the probability in the topic relative to the empirical frequency in the corpus, on the log scale, $$\log \mathrm{lift}_{kv} = \log\beta_{kv} - \log\frac{c_v}{\sum_u c_u},$$ with \(c_v\) the count of term \(v\). It favours rare, topic-specific terms; it needs the counts (countsorstm_data).
Usage
topic_terms(
model,
measure = "prob",
n_terms = 10L,
frexw = 0.5,
counts = NULL,
stm_data = NULL,
...
)Arguments
- model
A fitted
stmmodel without content covariate (seemarginal_topic_model()), or a list withbeta$logbeta[[1]](topics x terms log-probabilities) andvocab.- measure
Character vector, one or several of
"prob","frex","score","lift".- n_terms
Integer scalar, number of terms per topic and measure.
- frexw
Numeric scalar in [0, 1], weight of frequency against exclusivity in FREX (
stm::labelTopics()uses 0.5).- counts
For
"lift": a table withtermandcount, the empirical counts (e.g. fromterm_counts()).- stm_data
For
"lift", whencountsisNULL: the STM input the model was fitted on (documents,vocab), from which the counts are computed.- ...
Not used: an argument renamed in the development version of mintR gives an error naming its replacement.
Value
A data.table with topic, measure, rank, term and
value (the value of the measure). For lift, terms with a zero count
have no defined lift and are never returned. The mean FREX of a model
(an exclusivity-like summary) is
topic_terms(model, "frex")[, mean(value)].
Details
The FREX of topic_terms() follows the formula above; the FREX
labels of topic_labels() reproduce stm::labelTopics(), which
shrinks exclusivity towards the corpus word counts (James-Stein
estimator): the two can differ slightly.
References
Bischof, J. and Airoldi, E. M. (2012). Summarizing topical content with word frequency and exclusivity. Proceedings of ICML 2012, 201-208.
Roberts, M. E., Stewart, B. M., Tingley, D., Lucas, C., Leder-Luis, J., Gadarian, S. K., Albertson, B. and Rand, D. G. (2014). Structural topic models for open-ended survey responses. American Journal of Political Science, 58(4), 1064-1082. doi:10.1111/ajps.12103
Taddy, M. (2012). On estimation and selection for topic models. Proceedings of AISTATS 2012, 1184-1193.
Examples
logbeta <- log(matrix(c(0.5, 0.3, 0.2, 0.1, 0.2, 0.7), nrow = 2, byrow = TRUE))
mdl <- list(beta = list(logbeta = list(logbeta)), vocab = c("wage", "labour", "trade"))
topic_terms(mdl, n_terms = 2)
#> topic measure rank term value
#> <int> <char> <int> <char> <num>
#> 1: 1 prob 1 wage 0.5
#> 2: 1 prob 2 labour 0.3
#> 3: 2 prob 1 trade 0.7
#> 4: 2 prob 2 labour 0.2
counts <- data.frame(term = c("wage", "labour", "trade"), count = c(10, 5, 20))
topic_terms(mdl, c("frex", "lift"), n_terms = 2, counts = counts)
#> topic measure rank term value
#> <int> <char> <int> <char> <num>
#> 1: 1 frex 1 wage 1.0000000
#> 2: 1 frex 2 labour 0.6666667
#> 3: 2 frex 1 trade 1.0000000
#> 4: 2 frex 2 labour 0.6666667
#> 5: 1 lift 1 labour 0.7419373
#> 6: 1 lift 2 wage 0.5596158
#> 7: 2 lift 1 labour 0.3364722
#> 8: 2 lift 2 trade 0.2029408