Greece

Word Embeddings are used widely in multiple Natural Language Processing (NLP) applications. They are coordinates associated with each word in a dictionary, inferred from statistical properties of these words in a large corpus. In this paper we introduce the notion of ``concept&#39;&#39; as a list of words that have shared semantic content. We use this notion to analyse the learnability of certain concepts, defined as the capability of a classifier to recognise unseen members of a concept after training on a random subset of it. We first use this method to measure the learnability of concepts on pretrained word embeddings. We then develop a statistical analysis of concept learnability, based on hypothesis testing and ROC curves, in order to compare the relative merits of various embedding algorithms using a fixed corpora and hyper parameters. We find that all embedding methods capture the semantic content of those word lists, but fastText performs better than the others.

AIAI 2020

On the Learnability of Concepts With Applications to Comparing Word Embedding Algorithms

nlp evaluation

word embeddings

classification

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>
<br><br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

Registration Required

ACL 2026

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

Word Embeddings are used widely in multiple Natural Language Processing (NLP) applications. They are coordinates associated with each word in a dictionary, inferred from statistical properties of these words in a large corpus. In this paper we introduce the notion of ``concept'' as a list of words that have shared semantic content. We use this notion to analyse the learnability of certain concepts, defined as the capability of a classifier to recognise unseen members of a concept after training on a random subset of it. We first use this method to measure the learnability of concepts on pretrained word embeddings. We then develop a statistical analysis of concept learnability, based on hypothesis testing and ROC curves, in order to compare the relative merits of various embedding algorithms using a fixed corpora and hyper parameters. We find that all embedding methods capture the semantic content of those word lists, but fastText performs better than the others.

technical paper

AIAI 2020 (16th International Conference on Artificial Intelligence Applications and Innovations) should be held in Halkidiki, Greece. Due to current Covid-19 situation the conference will be held virtually. Underline will host a repository of video talks and preserve them for the future. Videos will be released on Underline one day after the conference.

The automation of fake news detection is the focus of a great deal of scientific research. With the rise of social media over the years, there has been a strong preference for users to be informed using their social media account, leading to a proliferation of fake news through them. This presentation of the paper entitled "Hong Kong Protests: Using Natural Language Processing for Fake News Detection on Twitter" evaluates the veracity of politically-oriented news and in particular the tweets about the recent event of Hong Kong protests, with the aid of a dataset recently published by Twitter. From this dataset, Chinese tweets are translated into English, which are kept along with originally English tweets. By utilizing a language-independent filtering process, relevant tweets are identified. To complete the dataset, tweets originating from valid sources are used as the real portion, with journalists rather than news agencies being considered, which constitutes a novel aspect of the methodology. Well-known Machine Learning algorithms are used to classify tweets, which are represented by a feature value vector that is extracted, selected and preprocessed from the datasets and mainly revolves around language use, with word entropy being a novel feature. The results derived from these algorithms highlight morphological, lexical and vocabulary differences between tweets spreading fake and real news, which are for the most part in accordance with past related work.

Hong Kong Protests: Using Natural Language Processing for Fake News Detection on Twitter

Nowadays, Intelligent Personal Agents include Natural Language Understanding (NLU) modules, that utilize Machine Learning(ML), which can be included in different kind of applications in order to enable the translation of users’ input into different kinds of actions, as well as ML modules that handle dialogue. This translation is attained by the matching of a user’s sentence with an intent contained in an Agent.This paper introduces the first generation of the CERTH Intelligent Per-sonal Agent (CIPA) which is based on the RASA1framework and utilizes two machine learning models for NLU and dialogue flow classification.Besides the architecture of CIPA—Generation A, a novel dialogue-story generator that is based on the idea of adjacency pairs is introduced. By utilizing on this novel-generator, the agent is able to create all the possible dialog trees in order to handle conversations without training on existing data in contrast with the majority of the current alternative solutions. CIPA supports multiple intents and it is capable of classifying complex sentences consisting of two user’s intents into two automatic operations from the part of the agent. The introduced CIPA—GenerationA has been deployed and tested in a real-world scenario at Centre’s ofResearch & Technology Hellas (CERTH) nZEB Smart Home2in two different domains, energy and health domain.

Applying an Intelligent Personal Agent on a Smart Home using a Novel Dialogue Generator

Nowadays, the fashion industry is moving towards fast fashion, offering a large selection of garment products in a quicker and cheaper manner. To this end, the fashion designers are required to come up with a wide and diverse amount of fashion products in a short time frame. At the same time, the fashion retailers are oriented towards using technology, in order to design and provide products tailored to their consumers’ needs, in sync with the newest fashion trends. In this work, we propose an artificial intelligence system which operates as a personal assistant to a fashion product designer. The system’s architecture and all its components are presented, with emphasis on the data collection and data clustering subsystems. In our use case scenario, datasets of garment products are retrieved from two different sources and are transformed into a specific format by making use of Natural Language Processes. The two datasets are clustered separately using different mixed-type clustering algorithms and comparative results are provided, highlighting the usefulness of the clustering procedure in the clothing product recommendation problem.

Towards fashion recommendation: An AI system for clothing data retrieval and analysis

This presentation documents the efforts in implementing lyric generation machine learning models in the Greek language for the genre of Éntekhno music. To accomplish this, we used three different Long Short-Term Memory Recurrent Neural Network approaches. The first method utilizes word-level bi-directional network models, the second method expands on the first by learning the word embeddings on the initial layer of the network, while the last method is based on a char-level network model. Our experimental procedure, which utilized a high sample of human judges, shows that texts of lyrics generated by our models are of high quality and are not that easily distinguishable from actual lyrics.

Greek Lyrics Generation

Question Answering (QA) and Question Generation (QG) have been subjects of an intensive study in recent years and much progresshas been made in both areas. However, works on combining these twotopics mainly focus on how QG can be used to improve QA results. Through existing Natural Language Processing (NLP) techniques, we have implemented a tool that addresses these two topics separately. We further use them jointly in a pipeline. Thus, our goal is to understand how these modules can help each other. For QG, our methodology employs a detailed analysis of the relevant content of a sentence through Part-of-speech (POS) tagging and Named Entity Recognition (NER). Ensuring loose coupling with the QA task, in the latter we use Information Retrieval to rank sentences that might contain relevant information regarding a certain question, together with Open Information Retrievalto analyse the sentences. In its current version, the QG tool takes a sentence to formulate a simple question. By connecting QG with the QA component, we provide a means to effortlessly generate a test set for QA. While our current QA approach shows promising results, when enhancing the QG component we will, in the future, provide questions for which a more elaborated QA will be needed. The generated QA datasets contribute to QA evaluation, while QA proves to be an important technique for assessing the ambiguity of the questions.

Exploring NLP and Information Extraction to jointly address Question Generation and Answering

With image processing, robots acquired visual perception skills; enabling them to become autonomous. Since the emergence of Artificial Intelligence (AI), sophis-ticated tasks such as object identification have become possible through inferenc-ing Artificial Neural Networks (ANN). Be that as it may, Autonomous Mobile Robots (AMR) are Embedded Systems (ESs) with limited on-board resources. Thus, efficient techniques in ANN inferencing are required for real-time perfor-mance. This paper presents the process of optimizing ANNs inferencing using tensor-based optimization on embedded Graphical Processing Unit (GPU) with Computer Unified Device Architecture (CUDA) platform for parallel acceleration on ES. This research evaluates renowned network, namely, You-Only-Look-Once (YOLO), on NVIDIA Jetson TX2 System-On-Module (SOM). The find-ings of this paper display a significant improvement in inferencing speed in terms of Frames-Per-Second (FPS) up to 3.5 times the non-optimized inferencing speed. Furthermore, the current CUDA model and TensorRT optimization tech-niques are studied, comments are made on its implementation for inferencing, and improvements are proposed based on the results acquired. These findings will contribute to ES developers and industries will benefit from real-time perfor-mance inferencing for AMR automation solutions

Tensor-based CUDA Optimization for ANN Inferencing using Parallel Acceleration on Embedded GPU

Authorship attribution attempts to identify the authors behind texts and has important applications mainly in cyber-security, digital humanities and social media analytics. An especially challenging but very realistic scenario is cross-domain attribution where texts of known authorship (training set) differ from texts of disputed authorship (test set) in topic or genre. In this paper, we modify a successful authorship verification approach based on a multi-headed neural network language model and combine it with pre-trained language models. Based on experiments on a controlled corpus covering several text genres where topic and genre is specifically controlled, we demonstrate that the proposed approach achieves very promising results. We also demonstrate the crucial effect of the normalization corpus in cross-domain attribution.

Cross-domain Authorship Attribution Using Pre-trained Language Models

Everybody likes to make a good prediction, in particular,
when some sort of personal investment is involved in terms of finance, energy or time. The difficulty is to make a prediction that optimises the reward obtained from the original contribution; this is even more important when investments are the core service offered by a business or pension fund generated by monthly contributions. The complexity of finance is that the human predictor may have other interests or bias than the human investor, the trust between predictor and investor will never
be completely established as the investor will never know if the predictor has generated, intentionally or unintentionally, the optimum possible reward. This paper presents the Random Neural Network in recurrent configuration that makes predictions on time series data, specifically, prices. The biological model inspired by the brain structure and neural interconnections makes predictions entirely on previous data from the time series rather than predictions based on several uncorrelated inputs. The model is validated against the property, stock and Fintech market: 1) UK property prices, 2) stock markets indice prices, 3) cryptocurrency
prices. Experimental results show that the proposed method makes accurate predictions on different investment portfolios.

The Random Neural Network in Price Predictions

We investigated age related synaptic plasticity in thalamic reticular nucleus (TRN) as a part of visual information processing system in the brain. Simulation experiments were performed using a hierarchical spike timing neural network model in NEST simulator. The model consists of multiple layers starting with retinal photoreceptors through thalamic relay, primary visual cortex layers up to the lateral intraparietal cortex (LIP) responsible for decision making and preparation of motor response. All synaptic inter- and intra-layer connections of our model are structured according to the literature information. The present work extends the model with spike timing dependent plastic (STDP) synapses within TRN as well as from visual cortex to LIP area. Synaptic strength changes were forced by teaching signal typical for three different age groups (young, middle and elderly) determined experimentally from eye movement data collected by eye tracking device from human subjects preforming a simplified simulated visual navigation task.

STDP Plasticity in TRN within Hierarchical Spike Timing Model of Visual Information Processing

Over the last years, Indoor Localization Systems (ILS) evolved, due to the inability of Global Positioning Systems (GPS) to localize in indoor environments. A variety of studies tackle indoor localization with technologies such as Bluetooth Beacons and RFID that require costly installation, or techniques such as Google Wi-Fi/Cell DB and fingerprinting that leverage from the already existing Wi-FI and telecommunication infrastructure. Additionally, recent studies attempt to solve the same problem using Bio-Inspired techniques, such as Artificial Neural Networks (ANNs) and Deep Neural Networks (DNN). In this paper, we introduce a Multi-Objective Optimization Radiomap Modelling (MOO-RM) based ILS. The MOO-RM ILS divides the dataset into clusters using a K-Means algorithm and trains ANN models on the data of each cluster. The resulting models are fed into a Multi-
Objective Evolutionary Algorithm based on Decomposition (MOEA/D), which minimizes the required storage space and the localization error, simultaneously. Our experimental studies demonstrate the superiority of athe proposed approach on real datasets of Wi-Fi traces with respect to various existing techniques.

Downloads

Next from AIAI 2020

Hong Kong Protests: Using Natural Language Processing for Fake News Detection on Twitter

Similar lecture

Vocabulary Shapes Cross-Lingual Variation of Word-Order Learnability in Language Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from AIAI 2020

Hong Kong Protests: Using Natural Language Processing for Fake News Detection on Twitter

Similar lecture

Vocabulary Shapes Cross-Lingual Variation of Word-Order Learnability in Language Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads