fastText Model Research Articles

Towards a word similarity gold standard for Akkadian: creation and model optimization

Abstract We present a word similarity gold standard for Akkadian, a language documented in ancient Mesopotamian sources from the 24th century BCE until the first century CE. The gold standard comprises 300 word pairs ranked by their paradigmatic similarity by five independently working Assyriologists. We use the gold standard to tune PMI + SVD and fastText models to improve their performance. We also present a hyper-parametrized PMI + SVD model for building count-based word embeddings, that aims to deal with the data sparsity and repetition issues encountered in Akkadian texts. Our model combines Dirichlet smoothing with context distribution smoothing, and uses context similarity weighting to down-sample distortion caused by formulaic litanies and partially or fully duplicated passages.

AN AUTOMATED NEW APPROACH IN FAST TEXT CLASSIFICATION: A CASE STUDY FOR KURDISH TEXT

With the rapid development of internet technology, text classification has become a vital part of obtaining quick and accurate data. Traditional machine learning methods often suffer from poor performance and high-dimensional feature spaces, which reduce their accuracy. In this paper, the FastText model is proposed as the first-ever classifier on Kurdish text and the results are compared with traditional machine learning methods to show the effects on Kurdish Text. For evaluating the model four datasets Kurdish News Dataset Headlines (KNDH), Medical Kurdish Dataset (MKD), Kurdish-Emotional-Dataset (KMD-77000), and KurdiSent are utilized and compared the results with the traditional machine learning algorithms such as: Random Forest (RF), k-nearest Neighbor (k-NN), Logistic Regression (LR), Multinomial Naïve Bayes (MNB), Support Vector Machine (SVM), Decision Tree (DT), Stochastic Gradient Descent (SGD), as well as the deep learning model Bidirectional Encoder Representations from Transformers (BERT). The outcomes indicate that the FastText model achieved the highest performance with 89% for each precision, recall, F1-score, and 89.10% accuracy for the KNDH dataset. Moreover, when the KMD dataset is utilized the FatText model obtained outperforms all others by approximately 2%. In addition, the comparative analysis showed that FastText is superior when Kurdisent is considered with precision, recall, F1-score, and accuracy by 81.32, 81.83, 81.57, and 81.4 respectively. In addition, when MKD is implemented, the FastText model obtained the highest performance with a precision of 93.32%, recall of 93.36, F1-score of 93.34, and accuracy of 93.1%.

Sentiment analysis model for cryptocurrency tweets using different deep learning techniques

Abstract Bitcoin (BTC) is one of the most important cryptocurrencies widely used in various financial and commercial transactions due to the fluctuations in the price of this currency. Recent research in large data analytics and natural language processing has resulted in the development of automated techniques for assessing the sentiment in online communities, which has emerged as a crucial platform for users to express their thoughts and comments. Twitter, one of the most well-known social media platforms, provides many tweets about the BTC cryptocurrency. With this knowledge, we can apply deep learning (DL) to use these data to predict BTC price variations. The researchers are interested in studying and analyzing the reasons contributing to the BTC price’s erratic movement by analyzing Twitter sentiment. The main problem in this article is that no standard model with high accuracy can be relied upon in analyzing textual emotions, as it represents one of the factors affecting the rise and fall in the price of cryptocurrencies. This article aims to classify the sentiments of an expression into positive, negative, or neutral emotions. The methods that have been used are word embedding FastText model in addition to different DL methods that deal with time series, one-dimensional convolutional neural networks (CONV1D), long-short-term memory networks (LSTMs), recurrent neural networks, gated recurrent units, and a Bi-LSTM + CONV1D The main results revealed that the LSTM method, based on the DL technique, achieved the best results. The performance accuracy of the methods was 95.01, 95.95, 80.59, 95.82, and 95.67%, respectively. Thus, we conclude that the LSTM method achieved better results than other methods in analyzing the textual sentiment of BTC.

Multi-label Classification of Indonesian Al-Quran Translation based CNN, BiLSTM, and FastText

Studying the Qur'an is a pivotal act of worship in Islam, which necessitates a structured understanding of its verses to facilitate learning and referencing. Reflecting this complexity, each Quranic verse is rich with unique thematic elements and can be classified into a range of distinct categories. This study explores the enhancement of a multi-label classification model through the integration of FastText. Employing a CNN+Bi-LSTM architecture, the research undertakes the classification of Quranic translations across categories such as Tauhid, Ibadah, Akhlak, and Sejarah. Based on model evaluation using F1-Score, it shows significant differences between the CNN+Bi-LSTM model without FastText, with the highest result being 68.70% in the 80:20 testing configuration. Conversely, the CNN+Bi-LSTM+FastText model, combining embedding size and epoch parameters, achieves a result of 73.30% with an embedding size of 200, epoch of 100, and a 90:10 testing configuration. These findings underscore the significant impact of FastText on model optimization, with an enhancement margin of 4.6% over the base model.

Design of cultural and creative experience products in scenic spots based on CORE model

Abstract Under the background of “cultural and tourism integration”, scenic cultural and creative experience products are an important means to enhance the communication power of scenic spots and generate income. In this paper, the CORE model is used as the design concept, and the CORE design process for scenic cultural and creative products is constructed. The plain Bayesian classification algorithm is used to extract review feedback from the product, and the vectorization of the extracted reviews is achieved using the FastText model. The vectorized reviews are input into the logistic regression classification model to obtain sentiment classification results. A fused sentiment feature vector is constructed by combining Apirior and a sentiment dictionary aimed at separating sentiment vector classification. Finally, the experience effect of cultural and creative products is explored as an example of X scenic spot, and the optimization strategy of product design is explored. The attribute feature of X Scenic Spot’s cultural and creative experience product, “color” embraces four emotional attributes, and the corresponding adaptations are 4.77, 5.42, 5.40, and 5.38, respectively. The SDs of the emotional attributes of the existing product are The SDs of the emotional attributes of the existing products are 1.141, 1.025, and 0.184, respectively, and the differences between the product descriptions and the users’ opinions on the perceptual attributes are large, which provides the evaluation correlation in the CORE model. Under the guidance of the CORE model, four aspects of cultural and creative marketing, scenic area IP, consumer needs and product experience dimensions are proposed.

Comparison of the accuracy of Japanese synonym identifications using word embeddings in the radiological technology field

The terminology in radiological technology is crucial, encompassing a broad range of principles from radiation to medical imaging, and involving various specialists. This study aimed to evaluate the accuracy of automatic synonym detection considering the characteristics of the Japanese language by Word2vec and fastText in the radiological technology field for the terminology elaboration. We collected around 340 thousand abstracts in Japanese. First, preprocessing of the abstract data was performed. Then, training models were created with Word2vec and fastText with different architectures: continuous bag-of-words (CBOW) and skip-gram, and vector sizes. Baseline synonym sets were curated by two experts, utilizing terminology resources specific to radiological technology. A term in the dataset input into the generated models, and the top-10 synonym candidates which had high cosine similarities were obtained. Subsequently, precision, recall, F1-score, and accuracy for each model were calculated. The fastText model with CBOW at 300 dimensions was most precise in synonym detection, excelling in cases with shared n-grams. Conversely, fastText with skip-gram and Word2vec were favored for synonyms without common n-grams. In radiological technology, where n-grams are prevalent, fastText with CBOW proved advantageous, while in informatics, characterized by abbreviations and transliterations, Word2vec with CBOW was more effective.

A multi granularity information fusion text classification model based on attention mechanism

Existing research on Chinese text classification primarily focuses on classifying data information at different granularities, such as character, word, sentence, and chapter. However, this approach often fails to capture the semantic information embedded in these different levels of granularity. To enhance the extraction of the text’s core content, this study proposes a text classification model that incorporates an attention mechanism to fuse multi-granularity information. The model begins by constructing embedding vectors for characters, words, and sentences. Character and word vectors are generated using the Word2Vec training model, allowing the data to be converted into these respective vectors. To capture contextual semantic features, a bidirectional long and short-term memory network is employed for character and word vectors. Sentence vectors, on the other hand, are processed using the FastText model to extract the features they contain. To extract further important semantic information from the different feature vectors, they are fed into an attention mechanism layer. This layer enables the model to prioritize and emphasize the most significant information within the text. Experimental results demonstrate that the proposed model outperforms both single-granularity classification and combinations of two or more granularities. The model exhibits improved classification accuracy across three publicly available Chinese datasets.

Pars-OFF: A Benchmark for Offensive Language Detection on Farsi Social Media

With the increasing use of social media with its ability for users to share comments immediately, the extent of a system to identify offensive content has become a necessity in all languages. Due to the lack of publicly available resources on offensive language identification for Farsi, which has more than 110 million speakers, we present Pars-OFF, a three-layered annotated corpus for offensive language detection in Farsi to fill the existing gap. The introduced corpus contains 10,563 data samples. The tweets have been collected with a combination of similarity-based and keyword-based data selection techniques to avoid severe unbalancedness. Additionally, as a baseline, this paper reports the performance of the traditional machine learning approaches and Transformer based models over the Pars-OFF dataset. The best performance was obtained by the BERT+fastText model, yielding the F1-Macro score of 89.57.

Lambda Architecture-Based Big Data System for Large-Scale Targeted Social Engineering Email Detection

In this research, we delve deep into the realm of Targeted Social Engineering Email Detection, presenting a novel approach that harnesses the power of Lambda Architecture (LA). Our innovative methodology strategically segments the BERT model into two distinct components: the embedding generator and the classification segment. This segmentation not only optimizes resource consumption but also improves system efficiency, making it a pioneering step in the field. Our empirical findings, derived from a rigorous comparison between the fastText and BERT models, underscore the superior performance of the latter. Specifically, The BERT model has high precision rates for identifying malicious and benign emails, with impressive recall values and F1 scores. Its overall accuracy rate was 0.9988, with a Matthews Correlation Coefficient value of 0.9978. In comparison, the fastText model showed lower precision rates. Leveraging principles reminiscent of the Lambda architecture, our study delves into the performance dynamics of data processing models. The Separated-BERT (Sep-BERT) model emerges as a robust contender, adept at managing both real-time (stream) and large-scale (batch) data processing. Compared to the traditional BERT, Sep-BERT showcased superior efficiency, with reduced memory and CPU consumption across diverse email sizes and ingestion rates. This efficiency, combined with rapid inference times, positions Sep-BERT as a scalable and cost-effective solution, aligning well with the demands of Lambda- inspired architectures. This study marks a significant step forward in the fields of big data and cybersecurity. By introducing a novel methodology and demonstrating its efficacy in detecting targeted social engineering emails, we not only advance the state of knowledge in these domains but also lay a robust foundation for future research endeavors, emphasizing the transformative potential of integrating advanced big data frameworks with machine learning models.

Prediction of American Society of Anesthesiologists Physical Status Classification from preoperative clinical text narratives using natural language processing

BackgroundElectronic health records (EHR) contain large volumes of unstructured free-form text notes that richly describe a patient’s health and medical comorbidities. It is unclear if perioperative risk stratification can be performed directly from these notes without manual data extraction. We conduct a feasibility study using natural language processing (NLP) to predict the American Society of Anesthesiologists Physical Status Classification (ASA-PS) as a surrogate measure for perioperative risk. We explore prediction performance using four different model types and compare the use of different note sections versus the whole note. We use Shapley values to explain model predictions and analyze disagreement between model and human anesthesiologist predictions.MethodsSingle-center retrospective cohort analysis of EHR notes from patients undergoing procedures with anesthesia care spanning all procedural specialties during a 5 year period who were not assigned ASA VI and also had a preoperative evaluation note filed within 90 days prior to the procedure. NLP models were trained for each combination of 4 models and 8 text snippets from notes. Model performance was compared using area under the receiver operating characteristic curve (AUROC) and area under the precision recall curve (AUPRC). Shapley values were used to explain model predictions. Error analysis and model explanation using Shapley values was conducted for the best performing model.ResultsFinal dataset includes 38,566 patients undergoing 61,503 procedures with anesthesia care. Prevalence of ASA-PS was 8.81% for ASA I, 31.4% for ASA II, 43.25% for ASA III, and 16.54% for ASA IV-V. The best performing models were the BioClinicalBERT model on the truncated note task (macro-average AUROC 0.845) and the fastText model on the full note task (macro-average AUROC 0.865). Shapley values reveal human-interpretable model predictions. Error analysis reveals that some original ASA-PS assignments may be incorrect and the model is making a reasonable prediction in these cases.ConclusionsText classification models can accurately predict a patient’s illness severity using only free-form text descriptions of patients without any manual data extraction. They can be an additional patient safety tool in the perioperative setting and reduce manual chart review for medical billing. Shapley feature attributions produce explanations that logically support model predictions and are understandable to clinicians.

Topic Mining and Future Trend Exploration in Digital Economy Research

This work proposes a new literature topic clustering analysis framework, based on which the topics of digital-economy-related studies are condensed. First, we calculated the word vector of keywords using the FastText model, and then the keywords were merged according to semantic similarity. A hierarchical clustering method based on the Jaccard coefficient was employed to cluster the domain documents. Finally, the information gain method was applied to estimate the high-gain feature words for each category of topics. Based on the above framework, 23 categories of research topics were formed. We divided these topics into layers of digital technology, convergence innovation and digital governance, and we constructed a three-level digital economy research framework. Thereafter, the current hot spots and frontier trends were derived based on the number and growth rate of the literature. Our study revealed that the research on digital technology, which is the basic layer of the digital economy, has waned. The field related to the integration and innovation of digital technology and the real economy was the current research focus, among which the results with respect to “New Business Forms in the Digital Age”, “Circular Economy” and “Gig Economy” were abundant. The problems of the unbalanced development of the digital economy and digital monopoly have strengthened research on digital governance. Furthermore, research on “Regional Digital Economy”, “Chinese Digital Economy” and “Data Management” is in its initial stage and is a potential area of future research.

Topic Discovery and Hotspot Analysis of Sentiment Analysis of Chinese Text Using Information-Theoretic Method.

Currently, sentiment analysis is a research hotspot in many fields such as computer science and statistical science. Topic discovery of the literature in the field of text sentiment analysis aims to provide scholars with a quick and effective understanding of its research trends. In this paper, we propose a new model for the topic discovery analysis of literature. Firstly, the FastText model is applied to calculate the word vector of literature keywords, based on which cosine similarity is applied to calculate keyword similarity, to carry out the merging of synonymous keywords. Secondly, the hierarchical clustering method based on the Jaccard coefficient is used to cluster the domain literature and count the literature volume of each topic. Thirdly, the information gain method is applied to extract the high information gain characteristic words of various topics, based on which the connotation of each topic is condensed. Finally, by conducting a time series analysis of the literature, a four-quadrant matrix of topic distribution is constructed to compare the research trends of each topic within different stages. The 1186 articles in the field of text sentiment analysis from 2012 to 2022 can be divided into 12 categories. By comparing and analyzing the topic distribution matrices of the two phases of 2012 to 2016 and 2017 to 2022, it is found that the various categories of topics have obvious research development changes in different phases. The results show that: ① Among the 12 categories, online opinion analysis of social media comments represented by microblogs is one of the current hot topics. ② The integration and application of methods such as sentiment lexicon, traditional machine learning and deep learning should be enhanced. ③ Semantic disambiguation of aspect-level sentiment analysis is one of the current difficult problems this field faces. ④ Research on multimodal sentiment analysis and cross-modal sentiment analysis should be promoted.

Linguistic markers of Alzheimer’s in Spanish speakers: Automated metrics for free speech and verbal fluency tasks

AbstractBackgroundAutomated speech analysis can reveal objective, scalable markers of Alzheimer’s disease (AD). Yet, few studies have targeted Spanish, the most spoken language in Latin America, where AD is quickly escalating. Also, most research presents low interpretability, targeting linguistic features unrelated to the disorder’s core neuropsychological profile. Working with Hispanics, we have validated novel automated metrics capturing critical lexico‐semantic targets for AD: semantic granularity (coarseness of concepts) and ongoing semantic variability (conceptual closeness of successive words) –Figure 1A. I aim to describe and illustrate these metrics via two experiments comparing AD patients and healthy controls (HCs) on free speech and verbal fluency tasks. To test for specificity, replications were conducted in Parkinson’s disease (PD) and behavioral variant frontotemporal dementia cohorts.MethodIn Experiment‐1, 21 AD patients, 16 HCs, and 18 PD patients performed spontaneous and semi‐spontaneous speech tasks. In Experiment‐2, 32 AD patients, 27 HCs, 19 PD patients, and 32 bvFTD patients performed phonological and semantic fluency tasks. In both studies, responses were transcribed and lemmatized. Semantic granularity scores were computed on Python’s NLTK library to access WordNet, a lexical taxonomy leading from hypernyms (top node: ‘entity’) to hyponyms (e.g., ‘dog’). Each word’s granularity is defined as the distance between its node and ‘entity’ (Figure 1B). Ongoing semantic variability was analyzed with a FastText model pre‐trained with Spanish corpora. Each unique word in the vocabulary was assigned a vector, and the distance between words was quantified with the cosine of the angle between their assigned vectors (Figure 1C). Extracted features from both dimensions were analyzed through ANOVAs and XGBoost classifiers.ResultsIn Experiment‐1, relative to HCs, AD (but not PD) patients used more unspecific and fewer specific concepts, alongside more discontinuous conceptual choices. These features robustly classified between AD patients and HCs, but not between PD patients and HCs. In Experiment‐2, only AD patients exhibited reduced semantic granularity, which contributed to robust subject‐level classification. Ongoing semantic variability failed to distinguish among groups.ConclusionThese novel digital metrics can reveal objective, interpretable, and specific markers of AD in Spanish speakers, providing culture‐specific data for a vast but understudied linguistic community.

Empirical Study: How Issue Classification Influences Software Defect Prediction

Software defect prediction aims to identify potentially defective software modules to better allocate limited quality assurance resources. Practitioners often do this by utilizing supervised models trained using historical data. This data is gathered by mining version control and issue tracking systems. Version control commits are linked to issues they address. If the linked issue is classified as a bug report, the change is considered as bug fixing. The problem arises from the fact that issues are often incorrectly classified within issue tracking systems. This introduces noise into the gathered datasets. In this paper, we investigate the influence issue classification has on software defect prediction dataset quality and resulting model performance. To do this, we mine data from 7 popular open-source repositories, create issue classification and software defect prediction datasets for each of them. We investigate issue classification using four different methods; a simple keyword heuristic, an improved keyword heuristic, the FastText model and the RoBERTa model. Our results show that using the RoBERTa model for issue classification produces the best software defect prediction datasets, containing on average 14.3641% of mislabeled instances. SDP models trained on such datasets achieve superior performance, to those trained on SDP datasets created using other issue classification methods, in 65 out of 84 experiments, with 55 of them being statistically relevant. Furthermore, in 17 out of 28 experiments we could not show a statistically relevant performance difference between SDP models trained on RoBERTa derived software defect prediction datasets and those created using manually labeled issues.

Comprehensibility and Automation: Plain Language in the Era of Digitalization

Abstract The current article briefly presents a pilot machine-learning experiment on the classification of official texts addressed to lay readers with the use of support vector machine as a baseline and fastText models. For this purpose, a hand-crafted corpus was used, created by the experts of the National Tax and Customs Administration of Hungary under the office’s Public Accessibility Programme. The corpus contained sentences that were paraphrased or completely rewritten by the experts to make them more readable for lay people, as well their original counter pairs. The aim was to automatically distinguish between these two classes by using supervised machine-learning algorithms. If successful, such a machine-learning-based model could be used to draw the attention of experts involved in making the texts of official bodies more comprehensible to the average reader to the potentially problematic points of a text. Therefore, the process of rephrasing such texts could be sped up drastically. Such a rephrasing (considering, above all, the needs of the average reader) can improve the overall comprehensibility of official (mostly legal) texts, and therefore supports access to justice, the transparency of governmental organizations and, most importantly, improves the rule of law in a given country.

PARTICLE SWARM OPTIMIZATION IN TWIN SUPPORT VECTOR MACHINE TO CLASSIFY FAKE NEWS

Slang word is a complex word, difficult and cannot be ignored. Slang is used by certain circles and is limited so that not everyone knows the meaning of the conversations carried out by group members. Based on previous research that has been done, namely making slang using a manual process that requires quite a lot of time to collect slang words, so that our research aims to collect slang words by applying Deep Learning, namely Natural Language Processing using the word embedding FastText method to speed up the collection process. slang words. The author implements the techniques and algorithms that have been designed in the previous stage. This stage will ensure that the processes carried out in the research can be carried out in accordance with the theories that support the research. From the combined data between YouTube comments and the Indonesian dictionary, itwas found that 421 words are slang words. These slang words are obtained by means of the process of lookingfor word similarities (similarity words) between YouTube comments and Indonesian dictionaries. In building a slang dictionary from the youtube comment dataset with a pre-trained FastText model, a preprocessing process and normalization is carried out. After the normalization process was carried out to get normal words from each slang candidate, the results of the slang dictionary were 278 rows consisting of four columns, namely the lexical column, threshold, slang candidate, and normal words using a threshold of 0.05, 0.1 and 0.2 Keywords: Slang, Pre-trained FastText, NLP, Similarity Word

An Earthquake Emergency Web Data Cleaning and Classification Method Based on Word Frequency and Position Weighting.

The speed of earthquake emergency web document data cleaning is one of the key factors affecting emergency rescue decision-making. Data classification is the core process of data cleaning, and the efficiency of data classification determines the speed of data cleaning. This article is based on earthquake emergency Web document data and HTML structural features, combined with TF-IDF Algorithm and information calculation model, improves the word frequency factor and location factor parameters, and proposes the weighted frequency algorithm P-TF-IDF for earthquake emergency Web documents. To filter out less frequent words and optimize the FastText model, N-gram Feature word vectors effectively improve the efficiency of Web document data classification; for text classification data, use missing data recognition rules, data classification rules, and data repair rules to design an artificial intelligence-based earthquake emergency network information data cleaning framework to detect invalid data sets value, complete data comparison and redundancy judgment, clean up data conflicts and data errors, and generate a complete data set without duplication. The data cleaning framework not only completes the fusion of earthquake emergency network information but also provides a data foundation for the visualization of earthquake emergency data.

Identification of cyber harassment and intention of target users on social media platforms

Due to Coronavirus diseases in 2020, all the countries departed into lockdown to combat the spread of the pandemic situation. Schools and institutions remain closed and students’ screen time surged. The classes for the students are moved to the digital platform which leads to an increase in social media usage. Many children had become sufferers of cyber harassment which includes threatening comments on young students, sexual torture through a digital platform, people insulting one another, and the use of fake accounts to harass others. The rising effort on automated cyber harassment detection utilizes many AI-related components Natural language processing techniques and machine learning approaches. Though machine learning models using different algorithms fail to converge with higher accuracy, it is much more important to use significant natural language processes and efficient classifiers to detect cyberbullying comments on social media. In this proposed work, the lexical meaning of the text is analysed by the conventional scheme and the word order of the text is performed by the Fast Text model to improve the computational efficacy of the model. The intention of the text is analysed by various feature extraction methods. The score for intention detection is calculated using the frequency of words with a bully-victim participation score. Finally, the proposed model’s performance is measured by different evaluation metrics which illustrate that the accuracy of the model is higher than many other existing classification methods. The error rate is lesser for the detection model.

Evaluation of clustering techniques on Urdu News head-lines: a case of short length text

ABSTRACT Document clustering and topic modelling methods are useful in determining groups of semantically similar documents. In literature, various studies have been conducted to evaluate the performance of lexical and semantic features on a clustering task. However, the focus of such studies was on the English language which is considered a high resource language. This paper analyses the effect of various feature extraction techniques combined with clustering and topic modelling algorithms on grouping news headlines written in Urdu, a low-resource language. Various features were first extracted using different approaches such as term-frequency inverse document frequency (TF-IDF), word embeddings learned using Word2vec and fastText models, deep contextualised word embeddings from BERT, and document embeddings extracted from the Doc2Vec model. Three different clustering algorithms: K-Means, Affinity Propagation and Density-based spatial clustering of applications with noise (DBSCAN) were then used to produce clusters for a given set of news headlines. The Latent Dirichlet Allocation topic model was also assessed for comparison with clustering algorithms. The evaluation based on extrinsic measures revealed that higher-quality clusters were produced when the K-Means algorithm was applied on a frequent unigrams-based TF-IDF feature matrix. In addition, several word embedding models have been trained on the domain of Urdu News headlines using state-of-the-art techniques such as Word2Vec, Doc2Vec, and fastText.

Sportsman’s Mental State Evaluation and Early Warning Method Based on Intelligent CNN

Aiming at the intelligent needs of psychological state assessment of university students, the text information-based psychological problem identification approach is investigated in the paper. This approach uses the text of student forums within universities as the database and introduces the convolutional neural network (CNN) model in deep learning, which contains a convolutional layer, a pooling layer, and a fully connected layer. After the convolution is completed, the convolution result is delinearized by the activation function, and then, pooling is performed to improve the fitting ability of the network for nonlinearities. For data processing, behavioral features, attribute features, content features, and social relationship features are extracted from text information as the input of the CNN. The psychological lexicon of expertise (LIWC) is used to enhance the efficiency of text word frequency statistics when performing text content extraction. To evaluate the performance of the proposed method, simulations are performed in the open dataset of CLPsyh2017 ReachOut Forum, and the FastText method is used as a comparison. The results show that the CNN model achieves an accuracy of 0.71 in the full-sample domain, which is significantly higher than that of the FastText model at 0.64. In the early warning evaluation of mental states, the CNN performance is better than that of FastText.