A natural language processing based technique for sentiment analysis of college english corpus

Jingjing XuPeerJ Comput. Sci 2023 · DOI

TL;DR

This paper introduces an improved TF-IDF algorithm for sentiment analysis of college English corpora, aiming to enhance learning efficiency. The novel approach incorporates critical information weight and word density processing to better capture emotional nuances, achieving higher accuracy and efficiency compared to traditional TF-IDF.

What Problem Does This Paper Address?

The vast amount of English language material available online presents a challenge for learners to efficiently extract relevant information. While Natural Language Processing (NLP) offers tools for this, research into dynamic language analysis, particularly for college English corpora and emotion analysis, is still in its early stages. Existing methods may not adequately capture the subtle emotional tendencies and values within these texts, potentially hindering effective learning. This paper tackles this by developing a sentiment analysis model specifically tailored for college English corpora to improve information extraction and learning efficiency.

Methods

The core of this paper is an improved Term Frequency-Inverse Document Frequency (TF-IDF) algorithm. The standard TF-IDF algorithm, which relies on word frequency and inverse document frequency, is enhanced by adding two new processing links:

  • Critical Information Weight Processing: This likely aims to give more importance to words appearing in specific, critical parts of a text (though the exact mechanism isn't fully detailed, it's implied to go beyond simple TF).
  • Word Density Weight Processing: This considers how frequently a word appears within a specific document or category, potentially distinguishing between common words and those that are characteristic of a particular sentiment or topic.

The paper also introduces metrics like Outer Class Density (OCD) and Intra Class Density (ICD) to quantify word distribution across different text categories, which are then integrated into the improved TF-IDF weight calculation.

For emotion extraction, a weighted average method is used to determine the emotional value of inspirational words. The model also accounts for polarity reversal phenomena (where a word's sentiment changes based on context) by generating an attribute emotion dictionary. Sentence structural stratification theory is applied, breaking down sentences into clauses for analysis.

Datasets

The primary dataset used is the American Contemporary English Corpus (COCA), described as containing one billion words of text. This corpus is divided into five categories: spoken language, novel, popular magazine, newspaper, and academic journal, with an even distribution. The COCA is noted for its detailed retrieval functions, including frequency information and definitions.

Key Insights and Ideas

  • Beyond Simple Frequency: The paper argues that simply relying on word frequency (TF-IDF) isn't sufficient for sentiment analysis, as words with high frequency might not always be the most indicative of sentiment. The proposed improvements aim to address this by considering word placement and density.
  • Context Matters for Sentiment: The acknowledgment of "polarity reversal phenomena" highlights the understanding that word sentiment isn't absolute and depends heavily on the context and the attribute it's describing. This necessitates a more nuanced approach than a simple lexicon lookup.
  • Attribute-Specific Sentiment: The development of an "attribute emotion dictionary" suggests a move towards understanding sentiment in relation to specific aspects or attributes within the text, which is crucial for detailed analysis.
  • Clustering for Deeper Attributes: The use of clustering algorithms to group similar words and identify deeper, more abstract attributes of the English corpus indicates an effort to move beyond surface-level features.

Key Results

  • Classification Accuracy: The improved TF-IDF algorithm shows improved classification accuracy. When the proportion of feature words is 5% and 50%, accuracy reaches 73.03% and 88.39%, respectively. Accuracy tends to saturate when feature words exceed 15%.
  • Precision: The improved TF-IDF algorithm achieved a precision of 84.88% at maximum data set size, compared to 74.87% for traditional TF-IDF, representing an improvement of 7.38%~23.83%.
  • Recall: At maximum data set size, the improved TF-IDF algorithm achieved a recall of 88.900%, compared to 80.027% for traditional TF-IDF, an increase of 8.19%~26.84%.
  • Running Time: The running time of the model increases with the proportion of feature words. For 5% and 50% feature words, running times are 29.83s and 784.00s, respectively. The running time tends to saturate between 5% and 30% feature words.

Limitations and Future Work

  • Limited Dataset: The authors acknowledge that the data set used in this article is limited, suggesting that further research with larger datasets is needed.
  • Comparison with More Methods: The paper states that future work should involve comparing the proposed method with more advanced techniques beyond TF-IDF, especially those that excel in complex sentiment analysis tasks.
  • Novelty over SoTA: Future research will focus on developing novel approaches that surpass current State-of-the-Art (SoTA) methods, particularly in areas where TF-IDF might be lacking.

Summary updated

BibTeX
@article{xu2023natural,
  author = {Jingjing Xu},
  title = {{A natural language processing based technique for sentiment analysis of college english corpus}},
  journal = {PeerJ Comput. Sci.},
  year = {2023},
  doi = {10.7717/peerj-cs.1235},
}
Figure