This project aims to develop a state-of-the-art decision support platform for retail management. The project, called SIGAMER, integrates Natural Language Processing technologies and big data capabilities. The main objective is to use historical data from companies' ERPs and extract insights from diferent sources such as structured data, open data, social media and news. The platform, accessible through a centralized interface and various REST APIs, follows a Software as a Service approach for seamless integration with existing Enterprise Resource Planning systems. The platform's technologies include semantic extraction, knowledge graphs, aspect-based multimodal sentiment analysis, topic modeling, author profiling, and regression and classification models for demand forecasting. The final system is accessible through a web technology-based dashboard composed of configurable key performance indicators to provide decision support to retailers who can use this insight to improve asset management, optimize deployment and identify lucrative market sectors.
• (OB1) Development of an intelligent information crawler for news for oficial reports and social media for subjective information and news sources for oficial reports. The crawlers will compile multimodal data from text, images or PDF documents. • (OB2) Implementation of a semantic extraction and representation system based on ontologies. The goal is to associate information extracted from indexed documents with semantic structures based on knowledge graphs for data reasoning. • (OB3) Create a subjective text analysis system that uses state-of-the-art language models to build an aspect-based multimodal emotion analysis based on aspects extracted from ontologies. • (OB4) Develop regression and classification models for demand forecasting based on product types and sales forecasts. • (OB5) Create an intelligent retail demand management platform based on intelligent analysis of structured data and social networks. This includes integration of the previous modules, validation processes and testing to obtain the final system. The final modular system will consist of a configuration section, a dashboard view, and an alert system to notify users of potential problems.
collect information from websites and social networks. The crawlers can adapt to diferent domains, although they share a common interface. In the case of social networks, specific crawlers have been developed separately, such as X (formerly Twitter) or YouTube. Each smart crawler is configurable and diferent strategies can be defined, such as the frequency with which each crawler starts monitoring and the criteria it must have to decide whether a page is linked or not based on the content and specific keywords.
The smart crawlers can extract metadata from the web pages and multimedia content. On the one hand, metadata in in formats such as JSON-LD1 (JSON for LinkedData) to try to give interoperability to the data and have more confidence in the subject of linked data. Metadata about published and updated data is also tracked. On the other hand, multimodal content such as audio and
images. For audio data extraction, we use Whisper [1], an automatic speech recognition (ASR) model based on the Transformers architecture. Whisper has demonstrated human-level performance and robustness, making it a valuable tool for accurately transcribing audio data. For image data extraction, we use Vision Transformer (ViT) [2], which enables the extraction of visual features and image descriptions. ViT has shown remarkable capabilities in handling image data by transforming images into sequences of tokens, allowing eficient processing and extraction of relevant information. bedding using Sentence Transformers [9] models such as paraphrase-spanish-distilroberta. This representation allows us to encode documents as dense vectors that store the relationships between semantically similar entities. Second, we apply a dimensional reduction process using UMAP [10]. Third, we use HDBSCAN [11], which is a density-based clustering algorithm. Fourth, we extract the topics from the clusters. using a customized variant of TF–IDF.
2.2. Extraction and Semantic The subjetive data analysis module consists of two main components: (1) an aspect-based multimodal emotion representation module analysis system and (2) and an author profiling model. The goal of this module is to extract information from The aspect-based multimodal emotion analysis compreviously compiled and indexed documents using an ponent focuses on identifying and analyzing emotions ontology that includes concepts related to retail man- expressed within specific topics. Aspect-based emotion agement. This ontology was created by merging and analysis is a valuable asset in product reviews, as it faadapting existing ontologies and applies to diferent sub- cilitates the identification of emotions expressed toward domains related to retail management. The retail domain diferent features or attributes of the product, such as includes product concepts and types, distribution areas its performance, design or usability among others. We and their relationship with external agents. This ontol- first build a new multimodal dataset by merging existogy has been enriched with the information about prod- ing textual datasets such as EmoEvalEs [12] with custom ucts and services that each organization has in its own multimodal data extracted from social networks such as database. For example, when a supermarket wants to use YouTube, where audio and its transcription have been the system, its entire product catalog is loaded into the extracted using Whisper [1]. Next, we train a multimodal ontology to identify these concepts as possible aspects emotion analysis model using textual features extracted to be discovered. from MarIA [13] and acoustic features from Wav2Vec
Relevant concepts and their relationships are iden- [14]. Finally, we use the extraction and semantic
reptified from the documents using Stanza’s modules for resentation module to determine the aspects. We use
Dependency Parsing, Named Entity Recognition (NER), this approach because ontologies have been shown to be
Part-of-Speech Tagger (PoS) [3]. In addition to ontology efective in aspect-based sentiment analysis in the past
enrichment, KnowGL [4] is used and the dataset is ex- [
Another stage if topic extraction. For this, we devel- datasets are developed for conducting AP in Spanish: oped a Topic Modeling model based on KeyBERT2 and the (1) Spanish IncluCorpus 2023 and the (2) Spanish BERTopic [8]. To do this, we first build a dataset from CCAACorpus 2023. Both datasets are compiled from the collected evidence, including news and social net- Twitter using the UMUCorpusClassifier tool [ 17]. From works related to retail management. These documents the first list of users, we selected those whose accounts are stored in markdown format to preserve information were public and had published at least 100 tweets. Next, about the structure of the document, such as titles and di- we manually checked the users to ensure that only those visions by content sections. We preprocess the dataset by users were included whose location in their profile could removing the markdown tags as external links. We then be recognized as a place in Spain, or whose gender could apply a stemming process and use Stanza’s PosTagger to be recognized by analyzing their profile name or descripremove non-relevant grammatical information. Once the tion. In the case of gender, we first checked to see if the dataset is assembled, we train the topic modeling model user was gender non-binary. If they could not be classias follows. First, we encode each document as an em- fied as such, we tried to classify them as male or female. For the former, a list of words and phrases associated with non-binary gender was created. For example, it is common for people of this gender to use the pronoun 2https://towardsdatascience.com/ keyword-extraction-with-bert-724efca412ea “elle” in their description. If any keyword from the list of reading and processing data periodically and whose appeared in the user’s description, they were classified output can also be configured, including graphs or tables, as non-binary gender. For the second case, a dictionary among others. In addition, options will be implemented of male and female names was created by consulting var- to allow users to export the data obtained. Figure 2 disious websites that suggest names for babies according plays a screen capture of the dashboard. to gender. Names that were considered unisex were dis- The system also allows personalized alerts to be sent carded and, analogous to the previous case, if the user’s through multiple channels, such as email, when the data name matched one in the dictionary, it was classified exceeds a configurable threshold. as the appropriate gender. To train the author profiling modules, we first merge several existing datasets based on authorship analysis and then extract user-level fea- 3. Future work tures based on sentence embeddings from MarIA [13] and linguistic features [18].
In this module, several machine learning models focus on demand forecasting. These models include classification and clustering models to classify trends and group similar products based on a set of patterns and characteristics. On the other hand, regression models are used to estimate the optimal demand for a given product in a given situation. Explainable Artificial Intelligence (XAI) techniques [19] will help to interpret the results of the machine learning models.
On the one hand, classification models are used to predict the class or category to which the data belongs. For example, they can be used to predict whether or not a product belongs to a particular category. On the other hand, clustering models are used to group similar data into clusters or groups, i.e., products that share similar characteristics or exhibit similar patterns of behavior on various dimensions, such as sales characteristics, prices, or similar product characteristics. The models evaluated include decision trees, support vector machines and logistic regressions, as they are more interpretable than deep neural networks. For regression models, on the other hand, various time series-based models were evaluated to model a dependent variable over time and other independent variables. In addition, external variables such as weather, current economic conditions, commodity prices, or other exogenous variables were included to improve the prediction of future demand trends. The models evaluated include LSTM, ARIMA, and Prophet.
This dashboard will allow the creation and configuration of campaigns in which managers and stakeholders will be able to define the products, websites and social networks to be monitored, as well as other spatio-temporal criteria. Once configured, all the data will be displayed on a dashboard composed of several semantic KPIs, which are independent and autonomous components capable
On the one hand, future work for this project includes several key areas aimed at enhancing the retail management decision support platform. First, there is a focus on expanding data sources to include emerging social media platforms, additional open data repositories, and real-time market data feeds. This broader data integration will provide retailers with a more comprehensive view of market trends and consumer sentiment. Second, the project aims to refine and enhance sentiment analysis capabilities, including multilingual support. These advances will enable the platform to extract more valuable insights from textual data.
Meanwhile, the platform’s user interface and user experience are being improved based on user feedback and usability testing. This includes providing customizable dashboards to meet the diverse needs of retail professionals. Integration with third-party services and APIs is also being explored to extend the platform’s usefulness, such as integration with e-commerce platforms, marketing automation tools, and customer relationship management (CRM) systems. Finally, scalability and performance optimizations will be critical as the volume of data processed increases, involving distributed computing techniques, database query optimizations, and the use of cloud infrastructure to ensure eficiency and responsiveness.
Ministerio para la Transformación Digital y la Función Pública through the "Recovery, Transformation and Resilience Plan" and also funded by the European Union NextGenerationEU/PRTR through the research project 2021/C005/00149877.