In this article, we present a machine learning-based solution for matching the performance of the gold standard of double-blind human coding when it comes to content analysis in comparative politics. We combine a quantitative text analysis approach with supervised learning and limited human resources in order to classify the front-page articles of a leading Hungarian daily newspaper based on their full text. Our goal was to assign items in our dataset to one of 21 policy topics based on the codebook of the Comparative Agendas Project. The classification of the imbalanced classes of topics was handled by a hybrid binary snowball workflow. This relies on limited human resources as well as supervised learning; it simplifies the multiclass problem to one of binary choice; and it is based on a snowball approach as we augment the training set with machine-classified observations after each successful round and also between corpora. Our results show that our approach provided better precision results (of over 80% for most topic codes) than what is customary for human coders and most computer-assisted coding projects. Nevertheless, this high precision came at the expense of a relatively low, below 60%, share of labeled articles.

Automated text analysis methods have made possible the classification of large corpora of text by measures such as topic and tone. Here, we provide a guide to help researchers navigate the consequential decisions they need to make before any measure can be produced from the text. We consider, both theoretically and empirically, the effects of such choices using as a running example efforts to measure the tone of New York Times coverage of the economy. We show that two reasonable approaches to corpus selection yield radically different corpora and we advocate for the use of keyword searches rather than predefined subject categories provided by news archives. We demonstrate the benefits of coding using article segments instead of sentences as units of analysis. We show that, given a fixed number of codings, it is better to increase the number of unique documents coded rather than the number of coders for each document. Finally, we find that supervised machine learning algorithms outperform dictionaries on a number of criteria. Overall, we intend this guide to serve as a reminder to analysts that thoughtfulness and human validation are key to text-as-data methods, particularly in an age when it is all too easy to computationally classify texts without attending to the methodological choices therein.

Supervised machine learning methods are increasingly employed in political science. Such models require costly manual labeling of documents. In this paper, we introduce active learning, a framework in which data to be labeled by human coders are not chosen at random but rather targeted in such a way that the required amount of data to train a machine learning model can be minimized. We study the benefits of active learning using text data examples. We perform simulation studies that illustrate conditions where active learning can reduce the cost of labeling text data. We perform these simulations on three corpora that vary in size, document length, and domain. We find that in cases where the document class of interest is not balanced, researchers can label a fraction of the documents one would need using random sampling (or “passive” learning) to achieve equally performing classifiers. We further investigate how varying levels of intercoder reliability affect the active learning procedures and find that even with low reliability, active learning performs more efficiently than does random sampling.

Matching for causal inference is a well-studied problem, but standard methods fail when the units to match are text documents: the high-dimensional and rich nature of the data renders exact matching infeasible, causes propensity scores to produce incomparable matches, and makes assessing match quality difficult. In this paper, we characterize a framework for matching text documents that decomposes existing methods into (1) the choice of text representation and (2) the choice of distance metric. We investigate how different choices within this framework affect both the quantity and quality of matches identified through a systematic multifactor evaluation experiment using human subjects. Altogether, we evaluate over 100 unique text-matching methods along with 5 comparison methods taken from the literature. Our experimental results identify methods that generate matches with higher subjective match quality than current state-of-the-art techniques. We enhance the precision of these results by developing a predictive model to estimate the match quality of pairs of text documents as a function of our various distance scores. This model, which we find successfully mimics human judgment, also allows for approximate and unsupervised evaluation of new procedures in our context. We then employ the identified best method to illustrate the utility of text matching in two applications. First, we engage with a substantive debate in the study of media bias by using text matching to control for topic selection when comparing news articles from thirteen news sources. We then show how conditioning on text data leads to more precise causal inferences in an observational study examining the effects of a medical intervention.

Political scientists often wish to classify documents based on their content to measure variables, such as the ideology of political speeches or whether documents describe a Militarized Interstate Dispute. Simple classifiers often serve well in these tasks. However, if words occurring early in a document alter the meaning of words occurring later in the document, using a more complicated model that can incorporate these time-dependent relationships can increase classification accuracy. Long short-term memory (LSTM) models are a type of neural network model designed to work with data that contains time dependencies. We investigate the conditions under which these models are useful for political science text classification tasks with applications to Chinese social media posts as well as US newspaper articles. We also provide guidance for the use of LSTM models.

Word vectorization is an emerging text-as-data method that shows great promise for automating the analysis of semantics—here, the cultural meanings of words—in large volumes of text. Yet successes with this method have largely been confined to massive corpora where the meanings of words are presumed to be fixed. In political science applications, however, many corpora are comparatively small and many interesting questions hinge on the recognition that meaning changes over time. Together, these two facts raise vexing methodological challenges. Can word vectors trace the changing cultural meanings of words in typical small corpora use cases? I test four time-sensitive implementations of word vectors (word2vec) against a gold standard developed from a modest data set of 161 years of newspaper coverage. I find that one implementation method clearly outperforms the others in matching human assessments of how public dialogues around equality in America have changed over time. In addition, I suggest best practices for using word2vec to study small corpora for time series questions, including bootstrap resampling of documents and pretraining of vectors. I close by showing that word2vec allows granular analysis of the changing meaning of words, an advance over other common text-as-data methods for semantic research questions.

Though used frequently in machine learning, boosted decision trees are largely unused in political science, despite many useful properties. We explain how to use one variant of boosted decision trees, AdaBoosted decision trees (ADTs), for social science predictions. We illustrate their use by examining a well-known political prediction problem, predicting U.S. Supreme Court rulings. We find that our ADT approach outperforms existing predictive models. We also provide two additional examples of the approach, one predicting the onset of civil wars and the other predicting county-level vote shares in U.S. presidential elections.

There is an ongoing debate about whether human rights standards have changed over the last 30 years. The evidence for or against this shift relies upon indicators created by human coders reading the texts of human rights reports. To help resolve this debate, we suggest translating the question of changing standards into a supervised learning problem. From this perspective, the application of consistent standards over time implies a time-constant mapping from the textual features in reports to the human coded scores. Alternatively, if the meaning of abuses have evolved over time, then the same textual features will be labeled with different numerical scores at distinct times. Of course, while the mapping from natural language to numerical human rights score is a highly complicated function, we show that these two distinct data generation processes imply divergent overall patterns of accuracy when we train a wide variety of algorithms on older versus newer sets of observations to learn how to automatically label texts with scores. Our results are consistent with the expectation that standards of human rights have changed over time.

Automated text analysis allows researchers to analyze large quantities of text. Yet, comparative researchers are presented with a big challenge: across countries people speak different languages. To address this issue, some analysts have suggested using Google Translate to convert all texts into English before starting the analysis (Lucas et al. 2015). But in doing so, do we get lost in translation? This paper evaluates the usefulness of machine translation for bag-of-words models—such as topic models. We use the europarl dataset and compare term-document matrices (TDMs) as well as topic model results from gold standard translated text and machine-translated text. We evaluate results at both the document and the corpus level. We first find TDMs for both text corpora to be highly similar, with minor differences across languages. What is more, we find considerable overlap in the set of features generated from human-translated and machine-translated texts. With regard to LDA topic models, we find topical prevalence and topical content to be highly similar with again only small differences across languages. We conclude that Google Translate is a useful tool for comparative researchers when using bag-of-words text models.

We introduce a model that extends the standard vote choice model to encompass text. In our model, votes and speech are generated from a common set of underlying preference parameters. We estimate the parameters with a sparse Gaussian copula factor model that estimates the number of latent dimensions, is robust to outliers, and accounts for zero inflation in the data. To illustrate its workings, we apply our estimator to roll call votes and floor speech from recent sessions of the US Senate. We uncover two stable dimensions: one ideological and the other reflecting to Senators’ leadership roles. We then show how the method can leverage common speech in order to impute missing data, recovering reliable preference estimates for rank-and-file Senators given only leadership votes.

Despite the popularity of unsupervised techniques for political science text-as-data research, the importance and implications of preprocessing decisions in this domain have received scant systematic attention. Yet, as we show, such decisions have profound effects on the results of real models for real data. We argue that substantive theory is typically too vague to be of use for feature selection, and that the supervised literature is not necessarily a helpful source of advice. To aid researchers working in unsupervised settings, we introduce a statistical procedure and software that examines the sensitivity of findings under alternate preprocessing regimes. This approach complements a researcher’s substantive understanding of a problem by providing a characterization of the variability changes in preprocessing choices may induce when analyzing a particular dataset. In making scholars aware of the degree to which their results are likely to be sensitive to their preprocessing decisions, it aids replication efforts.

Measuring the polarization of legislators and parties is a key step in understanding how politics develops over time. But in parliamentary systems—where ideological positions estimated from roll calls may not be informative—producing valid estimates is extremely challenging. We suggest a new measurement strategy that makes innovative use of the “accuracy” of machine classifiers, i.e., the number of correct predictions made as a proportion of all predictions. In our case, the “labels” are the party identifications of the members of parliament, predicted from their speeches along with some information on debate subjects. Intuitively, when the learner is able to discriminate members in the two main Westminster parties well, we claim we are in a period of “high” polarization. By contrast, when the classifier has low accuracy—and makes a relatively large number of mistakes in terms of allocating members to parties based on the data—we argue parliament is in an era of “low” polarization. This approach is fast and substantively valid, and we demonstrate its merits with simulations, and by comparing the estimates from 78 years of House of Commons speeches with qualitative and quantitative historical accounts of the same. As a headline finding, we note that contemporary British politics is approximately as polarized as it was in the mid-1960s—that is, in the middle of the “postwar consensus”. More broadly, we show that the technical performance of supervised learning algorithms can be directly informative about substantive matters in social science.