The Structure of quanteda

Kenneth Benoit

See http://docs.quanteda.io for additional tutorials, examples, and general documentation.

Introduction

The logic of Quanteda’s design

Grammatical rules

The “grammar” of the package is split between three basic types of functions and data objects:

Constructors for core data types

The quanteda package consists of a few core data types, created by calling constructors with identical names. These are all “nouns” in the sense of declaring what they construct. This follows very similar R behaviour in many of the core R objects, such as data.frame(), list(), etc.

Core object types and their constructor functions:

Note that a core object class in quanteda is also the character atomic type, for which there is no constructor function, and is abbreviated as char in the function nomenclature.

Functions for manipulating core data types

Naming convention

All functions that begin with the name of a core object class will both input and output an object of this class, without exception.

This replaces the approach in versions up to 0.9.8.5 where a general methods such as selectFeatures() was defined for each applicable class of core object. This approach made the specific function behaviour unpredictable from the description of the general behaviour. It also made it difficult to get an overview of the functionality available for each object class. By renaming these functions following the convention of object class, followed by an underscore, followed by a verb (or verb-like statement), we could both separate the behaviours into specific functions, as well as clearly describe through the function name what action is taken on what type of object.

Advantages

In our view, the advantages of this clarity outweigh whatever advantages might be found from overloading a generic function. The functions corpus_sample(), tokens_sample(), and dfm_sample(), for instance, are clearer to understand and read from a package’s function index, than the previously overloaded version of sample() that could be dispatched on a corpus, tokenized text, or dfm object. Additionally, in the case of sample(), we avoid the namespace “conflict” caused by redefining the function as a generic, so that it could be overloaded. Our new, more specific naming conventions therefore reduce the likelihood of namespace conflicts with other packages.

Why are some operations unavailable for specific object types?

Because not every operation makes sense for every object type. Take the example of a “feature co-occurrence matrix”, or fcm. Construction of a feature co-occurrence matrix is be slightly different from constructing a dfm. Unlike the “Swiss-army” knife approach of dfm(), which can operate directly on texts, fcm() works only on tokens, since the definition of how the context of co-occurrence is defined is dependent on token sequences and therefore highly dependent on tokenization options. In addition, fcm() is likely to be used a lot less frequently, and primarily by more expert users.

Furthermore, many functions defined for fcm() objects should be unavailable, because they violate the principles of the object. For instance, fcm_wordstem() and fcm_tolower() should not be applied to fcm() objects, because collapsing these and treating them as equivalent (as for a dfm object) is incorrect for the context in which co-occurrence is defined, such as a +/- 5 token window.

Extensions of core R functions

Many simple base R functions – simpler at least than the example of sample() cited above – are still extended to quanteda objects through overloading. The logic of allowing is that these functions, e.g. cbind() for a dfm, are very simple and very common, and therefore are well-known to users. Furthermore, they can operate in only one fashion on the object for which they are defined, such as cbind() combining two dfm objects by joining columns. Similar functions extended in this way include print(), head(), tail(), and t(). Most of these functions are so natural that their documentation is not included in the package index.

Additions to core R(-like) functions

Additional functions have been defined for quanteda objects that are very similar to simple base R functions, but are not named using the class_action format because they do not return a modified object of the same class. These follow as closely as possible the naming conventions found in the base R functions that are similar. For instance, docnames() and featnames() return the document names of various quanteda objects, in the same way that rownames() does for matrix-like objects (a matrix, data.frame, data.table, etc.). The abbreviation of featnames() is intentionally modeled on colnames(). Likewise, ndoc() returns the number of documents, using the singular form similar to nrow() and ncol().

Workflow principles

quanteda is designed both to facilitate and to enforce a “best-practice” workflow. This includes the following basic principles.

  1. Corpus texts should remain unchanged during subsequent analysis and processing. In other words, after loading and encoding, we should discourage users from modifying a corpus of texts as a form of processing, so that the corpus can act as a library and record of the original texts, prior to any downstream processing. This not only aids in replication, but also means that a corpus presents the unmodified texts to which any processing, feature selection, transformations, or sampling may be applied or reapplied, without hard-coding any changes made as part of the process of analyzing the texts. The only exception is to reshape the units of text in a corpus, but we will record the details of this reshaping to make it relatively easy to reverse unit changes. Since the definition of a “document” is part of the process of loading texts into a corpus, however, rather than processing, we will take a less stringent line on this aspect of changing a corpus.

  2. A corpus should be capable of holding additional objects that will be associated with the corpus, such as dictionaries, stopword, and phrase lists. These will be named objects, that can be invoked when using (for instance) dfm(). This allows a corpus to contain all of the additional objects that would normally be associated with it, rather than requiring a set of separate, extra-corpus objects.

  3. Objects should record histories of the operations applied to them. This is for purposes of analytic transparency. A tokens object and a dfm object, for instance, should have settings that record the processing options applied to the texts or corpus from which they were created. These provide a record of what was done to the text, and where it came from. Examples are tokens_tolower(), dfm_wordstem(), and settings such as remove_twitter. They also include any objects used in feature selection, such as dictionaries or stopword lists.

  4. A dfm should always be a documents (or document groups) in rows by features in columns. A dfm object may be transposed but then it is no longer a dfm class object.

  5. Encoding of texts should always be UTF-8.

Basic text analysis workflow

Working with a corpus, documents, and features

  1. Creating the corpus

    Reading files, probably using readtext() from the readtext package, then creating a corpus using corpus(), making sure the texts have a common encoding, and adding document variables (docvars()) and metadata (metadoc and metacorpus).

  2. Defining and delimiting documents

    Defining what are “texts”, for instance using corpus_reshape() or grouping (corpus_segment()).

  3. Defining and delimiting textual features

    This step involves defining and extracting the relevant features from each document, using tokens(), the main function for this step, involves identifying instances of defined features (“tokens”) and extracting them as vectors. Usually these will consist of words, but may also consist of:
    • ngrams: adjacent sequences of words, not separated by punctuation marks or sentence boundaries; including
    • multi-word expressions, through tokens_compound(), for selected word ngrams as identified in selected lists rather than simply using all adjacent word pairs or n-sequences.

    tokens() returns a new object class of tokenized texts, a hashed list of index types, with each element in the list corresponding to a document, and each hash vector representing the tokens in that document.

    By defining the broad class of tokens we wish to extract, in this step we also apply rules that will keep or ignore elements such as punctuation or digits, or special aggregations of word and other characters that make up URLs, Twitter tags, or currency-prefixed digits. This will involve adding the following options to tokens:
    • remove_numbers
    • remove_punct
    • remove_symbols
    • remove_twitter
    • remove_url

    By default, tokens() extracts word tokens, and only remove_separators is TRUE, meaning that tokens() will return a list including punctuation as tokens. This follows a philosophy of minimal intervention, and one requiring that additional decisions be made explicit by the user when invoking tokens().

    For converting to lowercase, it is actually faster to perform this step before tokenization, but logically it falls under the next workflow step. However for efficiency, *_tolower() functions are defined for character,tokens, and dfm objects.

    Since the tokenizer we will use may not distinguish the punctuation characters used in constructs such as URLs, email addresses, Twitter handles, or digits prefixed by currency symbols, we will mostly need to use a substitution strategy to replace these with alternative characters prior to tokenization, and then replace the substitutions with the original characters. This will slow down processing but will only be active by explicit user request for this type of handling to take place.

    Note that that defining and delimiting features may also include their parts of speech, meaning we will need to add functionality for POS tagging and extraction in this step.

  4. Further feature selection

    Once features have been identified and separated from the texts in the tokenization step, features may be removed from token lists, or handled as part of dfm construction. Features may be:
    • eliminated through use of predefined lists or patterns of stop words, using dfm(x, remove = ) or tokens_remove()
    • kept through through use of predefined lists or patterns of stop words, using dfm(x, select = ) or tokens_select()
    • collapsed by:
      • considering morphological variations as equivalent to a stem or lemma, through stem option in dfm (or dfm_wordstem())
      • considering lists of features as equivalent to a dictionary key, either exclusively using dfm(x,dictionary = )or as a supplement to uncollapsed features throughdfm(x, thesaurus = )
      • tokens_tolower() or dfm_tolower() to consider as equivalent the same word features despite having different cases, by converting all features to lower case

    It will be sometimes possible to perform these steps separately from the dfm creation stage, but in most cases these steps will be performed as options to the dfm() function.

  5. Analysis of the documents and features

    1. From a corpus.

      These steps don’t necessarily require the processing steps above.
      • kwic()
      • textstat_readability()
      • summary()
    2. From a dfm – after dfm() on the processed document and features.

dfm, the Swiss Army knife

In most cases, users will use the default settings to create a dfm straight from a corpus. dfm() will combine steps 3–4, even though basic functions will be available to perform these separately. All options shown in steps 3–4 will be available in dfm().

dfm objects can always be built up using constituent steps, through tokenizing and then selecting on the tokens. quanteda integrates the magrittr %>% pipe operator, to make it easy to chain functions without creating intermediate objects:

We recognize however that not all sequences will make sense, for instance tokens_wordstem() will only work after tokenization, and will try to catch these errors and make the proper sequence clear to users.