Readability assessment is an important task in automatic text simplification that aims identify the text complexity by computing a set of metrics. In this project, we present the development and assessment of an open source library called Pylinguistics.
The project is still in developing process, but until now, we already have some functional metrics.
- Word count
- Sentence count
- Word_per_sentence
- Syllable count
- Average of syllables per word
- Adjective Incidence
- Noun Incidence
- Verb Incidence
- Adverbs Incidence
- Pronoun Incidence
- Content words Incidence
- Functional word Incidence
- Lexical Diversty
- Content word Diversty
- Logic Negation Incidence
- Logic If Incidence
- Logic Or Incidence
- Logic And Incidence
- Logic Operators Incidence
- Connective Incidence
- Connective Additive Incidence
- Connective Logic Incidence
- Connective Temporal Incidence
- Connective Casual Incidence
.... and more
In order to illustrate the use of our tool, we contextualize a real world problem, the complexity and intelligibility of the scientific journalism. There are still few computational linguistic studies devoted to observe their textual constitution with particular emphasis on the characterization of stylistics elements of this textual genre. A thorough scientific journalism description can be extremely important for many of the core problems that computational linguists are concerned with. For example, parsing accuracy could be increased by taking genre into account, for instance, certain object-less constructions occur only in recipes in English. Similarly for POS-tagging, where the frequency of uses of trend as a verb in the Journal of Commerce is 35 times higher than in Sociological Abstracts. In information retrieval, genre classification could enable users to sort search results according to their immediate interests, for example scholarly articles about supercollider, novels about the French Revolution, and so forth.
The two corpora used in this study are geared towards different groups. Thus, they employ different vocabularies and textual structures that can be classified into different levels of complexity. In this study, we compared two corpora: Pesquisa Fapesp, a Brazilian specialized science magazine; and Folha de Sao Paulo (FSP), a Brazilian newspaper aimed at the general public.
We then assessed the predictive power of our features based on 3 feature selection algorithms commonly used for text categorization: Information Gain, Gain Ratio and Chi-square. Figure 1 shows the performance of the SVM when varying the number of features selected for each method. It shows that with only 2 features we can already predict the genre with over 93% accuracy. Additionaly by 7 metrics it already reaches the best possible result (97%). Finding a small subset of predictors is important to avoid over-fitting.
>>> import Pylinguistics as pl
>>> objpl=pl.text('Ia bem em matemática, porém reprovou em física.')
>>> objpl.setLanguage("pt-br")
>>> # this is a multiline comment
>>> objpl.getFeatures()
{'ConnectiveAdditiveIncidence': 125.0, 'redability': 66.6, 'word_count': 8, 'ConnectiveLogicIncidence': 0.0, 'syllable_count': 17, 'avg_word_per_sentence': 8.0, 'LogicIfIncidence': 0.0, 'LogicAndIncidence': 0.0, 'ContentDiversty': 1.0, 'pronIncidence': 0.0, 'LogicOperatorsIncidence': 0.0, 'verbIncidence': 250.0, 'functionalIncidence': 375.0, 'nounIncidence': 250.0, 'LogicOrIncidence': 0.0, 'adjectiveIncidence': 0.0, 'LogicNegationIncidence': 0.0, 'contentIncidence': 625.0, 'ConnectiveIncidence': 125.0, 'avg_syllables_per_word': 2.125, 'ConnectiveTemporalIncidence': 0.0, 'sentence_count': 1, 'ConnectiveCasualIncidence': 0.0, 'advIncidence': 125.0, 'LexicalDiversty': 0.9}Pylinguistics also requires extra resources: NLTK and nlpnet. Additionally, NLTK needs some extra downloads. After installing it, call
>>> import nltk
>>> nltk.download()You can also test pylinguistics by yourself without any instalation: http://app.mybinder.org/1746087056/notebooks/pylinguistics_test.ipynb
(not working yet - We'll do it ASAP) pip install git+git://github.com/vwoloszyn/pylinguistics.git
http://www.fsma.edu.br/si/edicao18/FSMA_SI_2016_2_Principal_2_en.html https://www.lume.ufrgs.br/bitstream/handle/10183/147640/000999695.pdf?sequence=1
- Article{Castilhos2016,
- author = {Castilhos, S. and Woloszyn, V, and Barno, D. and Wives, L. K.}, title = {Pylinguistics: an open source library for readability assessment of texts written in Portuguese}, journal = {Revista de Sistemas de Informação da FSMA}, year = {2016}, volume = {18}, issn = {1983-5604},
}