Emotion Classification in Short English Texts using Deep Learning Techniques
^††thanks: Identify applicable funding agency here. If none, delete this.

Recent research has highlighted the effectiveness of deep learning methodologies, such as transfer learning and word embedding, in recognizing emotions in languages with limited resources. Transfer learning, particularly in the realm of text embedding, has been leveraged to customize existing models for emotion detection tasks. Various iterations of BERT and similar models have been employed for text embedding purposes. Moreover, transfer learning techniques have been applied to enhance the comprehension of emotional word semantics within deep learning frameworks. Researchers also utilize approaches like combining models from diverse linguistic backgrounds and employing data augmentation techniques to enhance the accuracy of emotion recognition systems. In the domain of English text emotion recognition, early studies primarily relied on shallow learning approaches but gradually transitioned to more sophisticated methods like GRU and BERT-based embedding models such as XLM-R [8]. Prior investigations often failed to consider the impact of text length on classification complexity, despite its acknowledged significance. Hence, further exploration is warranted to meet the challenges associated with detecting emotions in concise English texts.

The goal of this research is to propose a novel training model to the BERT language model with standard statistical text models of character, word, and structure levels for the single-level topical classification for the gathered text dataset. Standard models are based on the letters and punctuation mark occurrences, word and parts-of-speech n-grams.

The study introduces SmallEnglishEmotions, a novel dataset comprising concise English texts labeled with five distinct emotional categories. The choice of short texts for emotion detection is rooted in the scientific rationale that single-emotion presence in each short text simplifies annotation for evaluators, facilitating more precise assessment by deep learning algorithms. Additionally, reducing the number of emotions per text enhances automatic emotion recognition confidence, yielding more robust outcomes. Furthermore, the brevity of short texts poses challenges for both human readers and deep learning models, urging researchers to devise improved classification methods. Thus, this benchmark dataset presents a more demanding task, encouraging researchers to develop and evaluate emotion classification techniques with heightened assurance. This segment provides a detailed description of the SmallEnglishEmotions dataset along with its principal attributes.

The dataset employed in this study, known as SmallEnglishEmotions, comprises of 6372 short English texts sourced from Twitter. Annotation of the dataset adheres to Rachael Jack’s emotional model [31], categorizing texts into five emotional classes: happiness, sadness, anger, fear, and others. Unlike many publicly available datasets, SmallEnglishEmotions focuses specifically on shorter texts, with an average length of 50 words.
Data collection involved utilizing the Snscrape library in Python to gather tweets from Twitter selectively. Tweets were sourced around specific topics such as sports, politics, trending news, and the USA, among others. Each topic was associated with a specific query containing relevant keywords. Data collection spanned the years 2020 to 2023, with texts initially not restricted by length. Subsequently, texts ranging from 20 to 70 words were selected as short texts for inclusion in the dataset. Geographical restrictions were not imposed, but only English tweets were considered during the search process. .

Each entry within the dataset is associated with a single emotional label. Human annotators, numbering five in total, were tasked with assigning emotional labels to each sample. This annotation process was facilitated through a web-based application developed using React and a CSV database. Upon completion of the annotation task, the collected data were stored in the database and subsequently retrieved into a pandas dataframe for analysis. Each annotator was assigned a unique identifier (ID) to track their contributions. Using this ID and the provided user interface, annotators navigated through the dataset, assigning emotions to each sample by selecting from a range of predefined options. Annotators were instructed to exercise caution, avoiding the assignment of multiple labels or labeling ambiguous instances by utilizing the provided ”ignore” button.

The SmallEnglishEmotions dataset contains 6372 instances in five emotion classes happiness, sadness, fear, anger, and other. The number of instances of each class in the SmallEnglishEmotions dataset is shown in table 1

All methodologies proposed in this investigation were implemented utilizing the Python programming language, employing the Scikit Learn and Keras libraries. The experiments were conducted within a Google Colab cloud environment, equipped with 12.7 GB RAM, a T4 GPU, and 15.0 GB GPU RAM. For training the model, the Adam optimizer was employed alongside a batch size of 32. In order to mitigate the risks of underfitting, overfitting, and prolonged training durations, early stopping mechanisms were integrated during the training phase of the model. Specifically, a maximum of 100 epochs with a patience of 10 were specified.The input layer’s maximum sentence length was set to 100 words for the SmallEnglishEmotions dataset.