Whether physical or digital, data organization is a critical part of our routine matters. The growing amount of data in modern working environments demands an effective system to manage files and folders. The file organization and management have several aspects, including labeling and naming digital files. Names and labels play a significant part in organizing the digital files stored in thousands inside computer systems. The number of digital files stored in any computer system is aggregating rapidly with the progress in the development of mass secondary storage devices [1]. With this increasing number of files, most users find it difficult to search and access any specific file. Computer users usually spend much time every day interacting with digital files and folders stored in their systems. These interactions consist of several actions, including creating, downloading, labeling, reviewing, navigating, searching for, moving, saving, copying, sharing, and deleting digital files. In addition, files that contain heterogeneous data, such as numbers, images, or sounds, are becoming popular. Users usually tend to demonstrate substantial creativity in file labeling [2]. However, the file labeling patterns are recognizable such as files being named to display the file they represent, their purpose, or a relevant creation date or deadline [3], but may also contain characters to expedite the sorting of the files to reduce clutter. Commonly, it is believed that a balanced and expressive file name may oblige multiple tenacities, including labeling, information organization, decreased searching time, and increased readability [4].

Filename generation is typically an unsupervised machine learning technique through which the hidden semantic information can be extracted from the corpus [5]. The corpus used in file labeling can typically be made up of thousands of files in the form of a data set, CSV file, and text file. Another essential aspect of this is that sometimes the user files require more than one topic or include many words. No hard and fast rule can be applied to get highly accurate and meaningful data in file labeling [6]. Therefore, file labeling relies on linguistic analysis and preprocessing techniques to filter the data to get reasonably meaningful results [1].

One of the significant issues in file labeling is analyzing the type of content and language used in the file. The data provided for file labeling first needs to go through a rigorous preprocessing stage to remove insignificant content that does not play any role in file labeling. Once effective content is extracted, the hidden feature is analyzed to compute the preprocessed file’s dimensionality, structure, and size. The semantic information obtained can be further converted into one hot vector for the unique representation of each word in files because files may show many Labels, and these Labels can also overlap with each other. File labeling is also used to break down the file into different Labels based on the probability distribution. In this research, different techniques have been used, such as Latent Dirichlet Allocation (LDA) [7], Latent Semantic Analysis (LSA) [8], and Non-negative Matrix Factorization (NMF) [9] for automated file labeling. This research aims to incorporate computational linguistic analysis based on machine learning techniques. It is hypothesized that an automated system can be developed to generate contextually meaningful labels for user files to assist users in the overall management and organization of digital files inside any computer system.

Previously, several attempts have been made in topic modeling and labeling short text and articles. Existing topic modeling tasks are mainly done using specific datasets such as fake news, ABC news dataset, New York times dataset, Twitter dataset, etc. Considering the monotony in these datasets, existing labeling and topic modeling approaches perform very well [10]. Daniel et al. [11] studied biological science-related data and analyzed that standard preprocessing approaches are not good enough to apprehend such diversified documents. They [11] developed a standalone toolbox and extracted common daily life words through graphical representation and heat-map and graphical representation based on topic similarity. Rubayyi et al. [12] used different techniques such as LDA, LSA, and Probabilistic Latent Semantic Analysis (PLSA). They identified probability-based topics over a dataset of hundred documents based on frequent keywords.

Pantel et al. [13] proposed using a lexical-semantic pattern for labeling semantic classes, encompassing all class members to learn a different label. A significant limitation of this task was that it only worked well over the semantically homogeneous, fine-grained clusters. Alokaili et al. [14] proposed a neural approach using the sequence-to-sequence technique for document labeling that does not suffer from this limitation. The model was trained using distant supervision over a sizeable synthetic dataset. Human experts evaluated the model by comparing labels to ones generated by the proposed technique. Initially, Seung et al. [9] conducted [15] human scoring of topics. Humans evaluated topics used in a novel and directly scored different topics learned by a topic model based upon pointwise mutual information. Qiaozhu et al. [16] used unsupervised machine learning approaches for automatically labeling topics. They generated expected labels through bigrams and noun chunks and afterward ranked those expected labels based on divergence with the preselected topic. Another approach that many researchers have used is to match topic words to concepts based on knowledge base [17,18]. Jey et al. [19] extracted top N terms to select topics, while some others have used summarization approaches to create labels for topics [20,21].

One common drawback of existing topic modeling approaches is the absence of interpretable topic space [20]. LDA [7,22] and PLSA [8] are traditional topic modeling techniques and envision different documents as a concoction of some discretized topics on some fixed parameter. Angelov [23] presented the Top2Vec approach based on Word2Vec and Doc2Vec models to build a document and topic vector and an interpretable word space. Another similar approach is BERTopic [24] which separates the embedding stage from the topic creation stage in contrast to the Top2Vec approach. Top2Vec considers words adjacent to the centroid of a cluster and creates coherent and interpretable topic representations very nicely.

On the other hand, BERTopic [24] focuses on the cluster and attempts to model the topic representation from the entire cluster. This allows the topic representations to be a bit more diverse and disregards the notion of centroids. One major limitation of document embedding techniques based on training is that they do not ensure quality on heterogeneous datasets. The model can only learn semantic associations like notions and thoughts between words and sentences but cannot comprehend the central idea of the document [25].

The proposed model consists of the following modules: preprocessing the data, Document matrix (DM) and Term analysis module (TAM), Topic modeling module (TMM), and File name generation module (FNGM) as shown in Fig. 1. The dataset used in this research is an amalgam of ABC news data set, different research articles, and books. The reason for collecting various datasets is to increase heterogeneity in the data used for training the system.

Automated file labeling is an exciting and challenging task simultaneously, as it requires dealing with heterogeneous data with varying amounts of words and tokens in different files. After the preprocessing step, the number of most frequent words was observed, as shown in Fig. 5.

Based on this frequency of words, the TF-IDF extracted terms from files are mentioned in Fig. 6. Unfortunately, the TF-IDF and other document vector results were sparse, noisy, and redundant. Therefore, once the vector representation of the significant and frequent terms was completed, different unsupervised machine learning techniques were applied for label modeling. The main objective of the topic modeling module is to find a concealed theme that governs the semantics of a file, as abstract labels characterize these themes. However, realistically these results are not based on any probabilistic model, as mentioned in Tab. 1.

The same information is also displayed through a histogram, as shown in Fig. 7. The labels obtained through NMF are mentioned in Tab. 2. It is observable that NMF label results are not appropriate as, first of all, it repeats most words, and above all, the words extracted by NMF are not much meaningful, neither have they depicted the actual label of the file. For instance, if we consider labels 2 and 3 from Tab. 2, both contain almost similar terms as extracted by NMF, while in Tab. 1, LDA extracted utterly different words for these files.

LSA is also considered one of the most fundamental approaches in label modeling. After LDA and NMF, we applied LSA to extract label words from the given corpus mentioned in Tab. 3. LSA results are much better than NMF, but LDA still got better results than LSA and NMF.

It is clear from Tabs. 1–3 that each LDA, NMF, and LSA extracted different terms for file labeling. Although NMF did not perform well for most of the files in some cases, such as labels 5 and 7, it extracted more meaningful terms. Furthermore, it was also observed that all of these three topic modeling approaches identified some non-significant words such as “one,” “go,” “like,” etc. Fig. 9 depicts the inter-topic distance map for labels of each file extracted using LDA, NMF, and LSA.

The dominant labels obtained through LDA and the coherence score of these labels are shown in Fig. 8.

One prime objective of this research was to generate an automatic system independent of any human interaction. Therefore coherence and relevance metrics are used to identify common and significant terms and better understand the semantics of extracted words for the final label. One particular problem with statistical methods in textual analysis is that the selection of the words is generally based on the probabilities extracted based on word frequency. These approaches cannot analyze whether any word contributes to specific semantics or is just a most frequent but semantically less significant word. Carson et al. [32] established a relevance metric, which can rearrange the order of frequent words in a label by considering their relevancy to the file. This relevancy is measured through a weighting parameter θ that can range from 0 to 1 and ascribes the corpus frequencies. If the value of θ is close to 1, then the order of top words will be considered equal to the order of standard conditional probabilities, while θ closer to 0 reorder the most specific word to the top of the list. All the label words extracted from different topic modeling approaches were used to calculate the relevance, as mentioned in Fig. 10a.

On the other hand, the coherence metric was developed by David et al. [33] and is being used for model selection and resolutions. However, it can help in guiding intuition when applied to single labels while identifying accurate labels in which a coherent concept might not be observed at first glance. Fig. 10b shows the coherence among the most relevant labels extracted through LDA, NMF, and LSA techniques. Finally, the relevant and most coherent topic words were given to BERT for file label generation. BERT was implemented with a fixed set of 3 words per label, as shown in Tab. 4.

It is clear from these results that the final label is much more meaningful than the label words extracted by LDA, NMF, and LSA individually. Moreover, these labels also give the impression of traditional file names given by a human. For instance, file 9 was a book on the history of the USA and is labeled as “Political history USA.” The label for file 4, an article on intelligent agent architectures, is “AI architecture design.” On the contrary, file 5 was an introductory brusher of different company products, and file 10 was an article on AI’s informed and uninformed search techniques. Both of these files did not get any meaningful results. One primary reason for these non-meaningful labels is that these files were not textually rich and had multiple redundant terms instead of a complete textual theme.

The increasing size and capacity of secondary storage devices and fast operating systems assist in aggregating the files users can store in their systems. However, this increase in the number of digital files might confuse users and cause them to spend additional time and effort finding and managing required files efficiently. Conversely, the need to automatically analyze, manage and label digital files has become significantly relevant. The particular challenge with file labeling is that it contains relatively heterogeneous and noisy data that might infer an inaccurate label. The proposed methodology can reasonably overcome this problem and can be used to label any textual file, including academic papers, user files, PowerPoint presentations, and PDFs. Despite individual label modeling techniques, qualitative evaluation of the sentence level BERT embedding reveals that these embeddings effectively organize a diverse range of digital files.