buzz: python corpus linguistics
buzz is a linguistics tool for parsing and then exploring plain or metadata-rich text. This README provides an overview of functionality. Visit the full documentation for a more complete user guide.
buzz requires Python 3.6 or higher. A virtual environment is recommended.
pip install buzz[word] # or git clone http://github.com/interrogator/buzz cd buzz python setup.py install
buzz has an optional frontend, buzzword, for exploring parsed corpora. To use it, install:
pip install buzz[word]
Then, generate a workspace,
cd into it, and start:
python -m buzzword.create workspace cd workspace python -m buzzword
More complete documentation is available here, as well from the main page of the app itself.
A URL will be printed, which can be used to access the app in your browser.
buzz models plain text, or CONLL-U formatted files. The remainder of this guide will assume that you are have plain text data, and want to process and analyse it on the command line using buzz.
First, you need to make sure that your corpus is in a format and structure that buzz can work with. This simply means putting all your text files into a folder, and optionally within subfolders (representing subcorpora).
Text files should be plain text, with a
.txt extension. Importantly though, they can be augmented with metadata, which can be stored in two ways. First, speaker names can be added by using capital letters and a colon, much like in a script. Second, you can use XML style metadata markup. Here is an example file,
<meta aired="10.01.1999" /> MELFI: My understanding from Dr. Cusamano, your family physician, is you collapsed? Possibly a panic attack? <meta exposition=true interrogative-type="intonation" move="info-request"> TONY: <meta emph=true>They</meta> said it was a panic attack <meta move="refute" /> MELFI: You don't agree that you had a panic attack? <meta move="info-request" question=type="in" /> ...
If you add a
meta element at the start of the text file, it will be understood as file-level metadata. For sentence-specific metadata, the element should follow the sentence, ideally at the end of a line. Span- and token-level metadata should wrap the tokens you want to annotate. All metadata will be searchable later, so the more you can add, the more you can do with your corpus.
To load corpora as buzz objects:
from buzz import Corpus corpus = Corpus("sopranos")
You can also make virtual corpora from strings, optionally saving the corpus to disk.
corpus = Corpus.from_string("Some sentences here.", save_as="corpusname")
spaCy to parse your text, saving the results as CONLL-U files to your hard drive. Parsing by default is only for dependencies, but constituency parsing can be added with a keyword argument:
# only dependency parsing parsed = corpus.parse() # if you also want constituency parsing, using benepar parsed = corpus.parse(cons_parser="benepar") # if you want constituency parsing using bllip parsed = corpus.parse(cons_parser="bblip")
You can also parse text strings, optionally passing in a name under which to save the corpus:
from buzz import Parser parser = Parser(cons_parser="benepar") for text in list_of_texts: dataset = parser.run(text, save_as=False)
The main advantages of parsing with buzz are that:
- Parse results are stored as valid CONLL-U 2.0
- Metadata is respected, and transferred into the output files
- You can do constituency and dependency parsing at the same time (with parse trees being stored as CONLL-U metadata)
parse() method returns another
Corpus object, representing the newly created files. We can explore this corpus via commands like:
parsed.subcorpora.s1.files.e01 parsed.files parsed.subcorpora.s1[:5] parsed.subcorpora["s1"]
You can also parse corpora without entering a Python session by using the
parse --language en --cons-parser=benepar|bllip|none path/to/conll/files # or python -m buzz.parse path/to/conll/files
Both commands will create
path/to/conll/files-parsed, a folder containing CONLL-U files.
Loading corpora into memory
You can use the
load() method to load a whole or partial corpus into memory, as a Dataset object, which extends the pandas DataFrame.
loaded = parsed.load()
You don't need to load corpora into memory to work on them, but it's great for small corpora. As a rule of thumb, datasets under a million words should be easily loadable on a personal computer.
The loaded corpus is a
Dataset object, which is based on the pandas DataFrame. So, you can use pandas methods on it:
|text||1||1||My||-PRON-||DET||PRP$||2||poss||_||10.01.1999||_||2||O||_||True||intonation||info-request||_||1||14||MELFI||My understanding from Dr. Cusamano, your family physician, is you collapsed?||0|
|2||understanding||understanding||NOUN||NN||13||nsubjpass||_||10.01.1999||_||2||O||_||True||intonation||info-request||_||1||14||MELFI||My understanding from Dr. Cusamano, your family physician, is you collapsed?||1|
|3||from||from||ADP||IN||2||prep||_||10.01.1999||_||2||O||_||True||intonation||info-request||_||1||14||MELFI||My understanding from Dr. Cusamano, your family physician, is you collapsed?||2|
|4||Dr.||Dr.||PROPN||NNP||5||compound||_||10.01.1999||_||2||O||_||True||intonation||info-request||_||1||14||MELFI||My understanding from Dr. Cusamano, your family physician, is you collapsed?||3|
|5||Cusamano||Cusamano||PROPN||NNP||3||pobj||_||10.01.1999||_||3||B||PERSON||True||intonation||info-request||_||1||14||MELFI||My understanding from Dr. Cusamano, your family physician, is you collapsed?||4|
You can also interactively explore the corpus with tabview using the
The interactive view has a number of cool features, such as the ability to sort by row or column. Also, pressing
enter on a given line will generate a concordance based on that line's contents. Neat!
Exploring parsed and loaded corpora
A corpus is a pandas DataFrame object. The index is a multiindex, comprised of
token. Each token in the corpus is therefore uniquely identifiable through this index. The columns for the loaded copus are all the CONLL columns, plus anything included as metadata.
# get the first sentence using buzz.dataset.sent() first = loaded.sent(0) # using pandas syntax to get first 5 words first.iloc[:5]["w"] # join the wordclasses and words print(" ".join(first.x.str.cat(first.w, sep="/")))
"DET/My NOUN/understanding ADP/from PROPN/Dr. PROPN/Cusamano PUNCT/, DET/your NOUN/family NOUN/physician PUNCT/, VERB/is PRON/you VERB/collapsed PUNCT/?
You don't need to know pandas, however, in order to use buzz, because buzz makes possible some more intuitive measures with linguistics in mind. For example, if you want to slice the corpus some way, you can easily do this using the
skip properties, combined with the column/metadata feature you want to filter by:
tony = loaded.just.speaker.TONY # you can use brackets (i.e. for regular expressions): no_punct = loaded.skip.lemmata("^[^a-zA-Z0-9]") # or you can pass in a list/set/tuple: end_in_s = loaded.just.pos(["NNS", "NNPS", "VBZ"])
Any object created by buzz has a
.view() method, which launches a
tabview interactive space where you can explore corpora, frequencies or concordances.
spaCy is used under the hood for dependency parsing, and a couple of other things. spaCy bring with it a lot of state of the art methods in NLP. You can access the
spaCy representation of your data with:
corpus.to_spacy() # or loaded.to_spacy()
To search the dependency graph generated by spaCy during parsing, you can use the depgrep method.
# search dependencies for nominal subjects with definite articles nsubj = loaded.depgrep('f/nsubj.*/ -> (w"the" & x"DET")')
The search language works by modelling nodes and the links between them. Specifying a node, like
f/nsubj/, is done by specifying the feature you want to match (
function), and a query inside slashes (for regular expressions) or inside quotation marks (for literal matches).
The arrow-like link specifies that the
nsubj must govern the determiner. The
& relation specifies that the two nodes are actually the same node. Brackets may be necessary to contain the query.
When you search a
Dataset, the result is simply another Dataset, representing a subset of the Corpus. Therefore, rather than trying to construct one query string that gets everything you want, it is often easier to perform multiple small searches:
query = 'f/nsubj/ <- f/ROOT/' tony_subjects = loaded.skip.wordclass.PUNCT.just.speaker.TONY.depgrep(query)
Note that for any searches that do not require traversal of the grammatical structure, you should use the
just methods. tgrep and depgrep only need to be used when your search involves the grammar, and not just token features.
Searching constituency trees
Constituency tree searching can be done with the
tgrep method, which provides a Python implementation of the
tgrep2 query syntax:
nps_with_adjectives = loaded.tgrep('NP < JJ')
It also works with nodes and links, though there are numerous differences. In particular, note that arrows appear reversed ---
NP < JJ is an NP that dominates a JJ, while something similar in depgrep would be
f/nsubj/ -> f/amod/, a nominal subject governing an adjective.
Viewing search results
An important principle in buzz is the separation of searching and viewing results. Unlike many other tools, you do not search for a concordance---instead, you search the corpus, and then visualise the output of the data as a concordance.
Concordancing is a nice way of looking at results. The main thing you have to do is tell buzz how you want the match column to look---it can be just the matching words, but also any combination of things. To show words and their parts of speech, you can do:
nsubj = loaded.just.function.nsubj nsubj.conc(show=["w", "p"])
You can turn your dataset into frequency tables, both before or after searching or filtering. Tabling takes a
show argument similar to the
show argument for concordancing, as well as an additional
show represents the how the columns will be formatted, and
subcorpora is used as the index. Below we create a frequency table of
nsubj tokens, in lemma form, organised by speaker.
tab = nsubj.table(show="l", subcorpora=["speaker"])
Possible keyword arguments for the
.table() method are as follows:
||Feature(s) to use as the index of the table. Passing in a list of multiple features will create a multiindex||
||Feature(s) to use as the columns of the table. Passing a list will join the features with slash, so
||How to sort the results. 'total'/'infreq', 'increase/'decrease', 'static/turbulent', 'name'/'inverse'||
||Use relative, rather than absolute frequencies with
||Sorting by increase/decrease/static/turbulent calculates the slope of the frequencies across each subcorpus, and p-values where the null hypothesis is no slope. If you pass in a float, entries with p-values above this float are dropped from the results. Passing in
||If True, keep generated statistics related to the trajectory calculation||
||Keep the original case for
This creates a
Table object, which is also based on DataFrame. You can use its
.view() method to quickly explore results. Pressing enter on a given frequency will bring up a concordance of instances of this entry.
You can also use buzz to create high-quality visualisations of frequency data. This relies completely on pandas' plotting method. A
plot method more tailored to language datasets is still in development.