snlp 0.1.3.post4

Statistical NLP

snlp

Statistical NLP (SNLP) is a practical package with statistical tools for natural language processing. SNLP is based on statistical and distributional attributes of natural language and hence most of its functionalities are unsupervised.

Features

• Text Analysis
• Extraction of Multiword Expressions
• Identification of Statistically Redundant Words
• Text Cleaning

Upcoming Features

• Identification of Non-compositional Expressions (e.g. red tape and brain drain). Non-compositional expressions have application in profanity detection, language understanding, and language generation.

• Entropy for Natural Language Entropy has a wide range of applications in NLP. Recently, researchers have show that it can be used to improve the quality of conversational AI [1] and text summarization [2].

• Label Evaluation Labeled datasets often come with certain level of human error. If not captured early on, this these errors will propagate to downstream machine learning models and hinder their quality rendering an otherwise well-performing model performs poorly. At model building and training time there is no easy way to identify if the error lies in the labels or the model itself. This often leads to spending a large amount of time trying to fix the problem. To avoid this, Label Evaluation features provides a set of functionalities to evaluate the labels and identify problematic ones based on measures of Inter Rater Agreement, and Correlations.

Usage

Install the package:

pip install snlp

See the description of different functionalities with worked examples below.

Text Analysis

snlp provides an easy to use function (text_analysis.generate_report) for analyzing text with an extensive analysis report. text_analysis.generate_report receives as input a dataframe that contains a text column, and an optional number of label columns. Currently, text_analysis.generate_report can generate plots for up to 4 numerical or categorical labels. See the example below for more details.

from snlp.text_analysis import generate_report

generate_report(df=imdb_train,
                out_dir='output_dir',
                text_col='text',
                label_cols=[('label', 'categorical')])

The above script creates an analysis report that includes distribution plots and word clouds for different POS tags, for text, and bar plots and histograms for labels. You can specify up to 4 labels of type categorical or numerical. See the example below for including another label of numerical type. The report is automatically rendered in the browser via plotly default port assignment. But you also have the option of saving the report in an HTML format by setting the save_report argument to True.

import numpy as np
import random

# In addition to the original label, for illustration purpose, let's create two random labels:
imdb_train['numerical_label'] = np.random.randint(1, 500, imdb_train.shape[0])
imdb_train['new_label'] = random.choices(['a', 'b', 'c', 'd'], [0.2, 0.5, 0.8, 0.9], k=imdb_train.shape[0])

generate_report(df=imdb_train,
                out_dir='output_dir',
                text_col='text',
                label_cols=[('label', 'categorical'), ('new_label', 'categorical'), ('numerical_label', 'numerical')])

The above yields a report in HTML, with interactive plotly plots as can be seen in example screenshots below.

You can easily zoom in any part of the plot to a have a closer look:

You can get word clouds for different part of speech tags, as can be seen in the below example where word clouds for nouns, adjectives and verbs are rendered:

Extraction of Multiword Expressions

Multiword Expressions (also known as collocations of fixed expressions) are phrases that function as a single semantic unit E.g. swimming pool and climate change. Multiword Expressions have application in a wide range of NLP tasks ranging from sentiment analysis to topic models and key-phrase extraction.

You can use snlp to identify different types of MWEs in your text leveraging statistical measures such as PMI and NPMI. To do so, first create an instance of MWE class:

from snlp.mwes import MWE
my_mwe_types = ["NC", "JNC"]
mwe = MWE(df=imdb_train, mwe_types=my_mwe_types, text_column='text')

If the text in text_column is un-tokenized or poorly tokenized, MWE recognizes this issue at instantiation time and shows you a warning. If you already know that your text is not tokenized, you can run the same instantiation with flag tokenize=True. Next you need to run the method build_count(). Since creating counts is a time consuming procedure, it was implemented independently from extract_mwes() method that works on top of the output of build_count(). This way, you can get the counts which is a time consuming process once, and then run extract_mwes() several times with different parameters.

mwe.build_counts()
mwe.extract_mwes()

Running the above results in a json file, containing dictionary of mwe types defined in the mwe_types argument of MWE, to their association score (specified by am argument of extract_mwes()). Note that the MWEs in this json file are sorted with respect to their am score. All MWEs and their counts are stored in respective directories inside the output_dir argument of MWE. The default value is tmp.

NOUN-NOUN COMPOUNDS
-------------------
jet li
clint eastwood
monty python
kung fu
blade runner


ADJECTIVE-NOUN COMPOUNDS
------------------------
spinal tap
martial arts
citizen kane
facial expressions
global warming

An important use of extracting MWEs is to treat them as a single token. Research shows that when fixed expressions are treated as a single token rather than the sum of their components, they can improve the performance of downstream applications such as classification and NER. Using the replace_mwes function, you can replace the extracted expressions in the corpus with their hyphenated version (global warming --> global-warming) so that they are considered a single token by downstream applications. A worked example can be seen below:

from snlp.mwes import replace_mwes
new_df = replace_mwes(path_to_mwes='tmp/mwes/mwe_data.json', mwe_types=['NC', 'JNC'], df=imdb_train, text_column='text')
new_df.to_csv('tmp/new_df.csv', sep='\t')

Identification of Statistically Redundant Words

Redundant words carry little value and can exacerbate the results of many NLP tasks. To solve this issue, traditionally, a pre-defined list of words, called stop words was defined and removed from the data. However, creating such a list is not optimal because in addition to being a rule-based and manual approach which does not generalize well, one has to assume that there is a universal list of stop words that represents highly low entropy words for all corpora, which is a very strong assumption and not necessarily a true assumption in many cases.

To solve this issue, one can use a purely statistical solution which is completely automatic and does not make any universal assumption. It focuses only on the corpus at hand. Words can be represented with various statistics. For instance, they can be represented by their term frequency (tf) or inverse document frequency (idf). It can be then interpreted that terms with anomalous (very high or very low) statistics carry little value and can be discarded. SNLP enables you to identify such terms in an automatic fashion. The solution might seem complex behind the scene, as it firsts needs calculate certain statistics, gaussanize the distribution of the specified statistics (i.e. tf or ifd), and then identify the terms with anomalous values on the gaussanized distribution by looking at their z-score. However, the API is easy and convenient to use. The example below shows how you can use this API:

from snlp.preprocessing import RedunWords

imdb_train = pd.read_csv('resources/data/imdb_train_sample.tsv', sep='\t', names=['label', 'text'])
rw = RedunWords(imdb_train["text"], method='idf')

Let the program automatically identify a set of redundant words:

red_words = rw.get_redundant_terms()

Alternatively, you can manually set cut-off threshold for the specified score, by setting the manual Flag to True and specifying lower and upper cut-off thresholds.

red_words = rw.get_redundant_terms(manual=True, manual_thresholds: dict={'lower_threshold':1, 'upper_threshold': 8})

In order to get a better understanding of the distribution of the scores before setting the thresholds, you can run show_plot() method from RedunWords class to see this distribution:

rw.show_plot()

When red_words is ready, you can filter the corpus:

# text must be a list of words
res = " ".join([t for t in text if t not in redundant_terms])

Text Cleaning

snlp implements an easy to use and powerful function for cleaning up the text (clean_text). Using, clean_text, you can choose what pattern to accept via keep_pattern argument, what pattern to drop via drop_patterns argument, and what pattern to replace via replace argument. You can also specify the maximum length of tokens. Let's use Stanford's IMDB Sentiment Dataset as an example. A sample of this data can be found in resources/data/imdb_train_sample.tsv.

from snlp.preprocessing import clean_text

imdb_train = pd.read_csv('data/imdb_train_sample.tsv', sep='\t', names=['label', 'text'])

# Let's only keep alphanumeric tokens as well as important punctuation marks:
keep_pattern='^[a-zA-Z0-9!.,?\';:$/_-]+$'

# In this corpus, one can frequently see HTML tags such as `< br / >`. So let's drop them:
drop_patterns={'< br / >'}

# By skimming throw the text one can frequently see many patterns such as !!! or ???. Let's replace them:
replace={'!!!':'!', '\?\?\?':'?'}

# Finally, let's set the maximum length of a token to 15:
maxlen=15

imdb_train.text = imdb_train.text.apply(clean_text, args=(keep_pattern, drop_patterns, replace, maxlen,))

Note that clean_text returns tokenized text.

References

[1] R. Csaky et al. - Improving Neural Conversational Models with Entropy-Based Data Filtering - In Proceedings of ACL 2019 - Florence, Italy.

[2] Maxime Peyrard - A Simple Theoretical Model of Importance for Summarization - In Proceedings of ACL 2019 - Florence, Italy.