Claudiu Musat

Riccardo Brioschi, Aleksandr Alekseev, Emanuele Nevali et al.

Graphic layout generation is a growing research area focusing on generating aesthetically pleasing layouts ranging from poster designs to documents. While recent research has explored ways to incorporate user constraints to guide the layout generation, these constraints often require complex specifications which reduce usability. We introduce an innovative approach exploiting user-provided sketches as intuitive constraints and we demonstrate empirically the effectiveness of this new guidance method, establishing the sketch-to-layout problem as a promising research direction, which is currently under-explored. To tackle the sketch-to-layout problem, we propose a multimodal transformer-based solution using the sketch and the content assets as inputs to produce high quality layouts. Since collecting sketch training data from human annotators to train our model is very costly, we introduce a novel and efficient method to synthetically generate training sketches at scale. We train and evaluate our model on three publicly available datasets: PubLayNet, DocLayNet and SlidesVQA, demonstrating that it outperforms state-of-the-art constraint-based methods, while offering a more intuitive design experience. In order to facilitate future sketch-to-layout research, we release O(200k) synthetically-generated sketches for the public datasets above. The datasets are available at https://github.com/google-deepmind/sketch_to_layout.

Madhuri Shanbhogue, Zhe Li, Shanfeng Zhang et al.

We introduce Gemini Embedding 2, a native multimodal embedding model that allows embedding video, audio, image, and text modalities in a unified representation space. We leverage the multimodal capabilities of Gemini to produce embeddings for arbitrary combinations of interleaved inputs across all these modalities that generalize well across a wide variety of tasks. Applying large-scale contrastive learning in a multi-task multi-stage training setup, we achieve state-of-the-art performance on key embedding benchmarks including unimodal, cross-modal, and multimodal retrieval spanning a diverse set of tasks. We show that our embedding model demonstrates strong performance (with a score of 62.9 R@1 on MSCOCO, 68.8 NDCG@10 on Vatex, 69.9 on MTEB multilingual and 84.0 on MTEB Code) across a variety of tasks surpassing the performance of specialized models. These unified capabilities make Gemini Embedding 2 a promising candidate for downstream use cases such as RAG, recommendation and search. Furthermore, its robust zero-shot performance across distinct fields - from astronomy and bioscience to fine arts and the culinary arts - establishes it as a highly reliable, out-of-the-box representation even for specialized domains.

Jonathan Cook, Diego Antognini, Martin Klissarov et al.

Large language models (LLMs) often struggle to learn from corrective feedback within a conversational context. They are rarely proactive in soliciting this feedback, even when faced with ambiguity, which can make their dialogues feel static, one-sided, and lacking the adaptive qualities of human conversation. To address these limitations, we draw inspiration from social meta-learning (SML) in humans - the process of learning how to learn from others. We formulate SML as a finetuning methodology, training LLMs to solicit and learn from language feedback in simulated pedagogical dialogues, where static tasks are converted into interactive social learning problems. SML effectively teaches models to use conversation to solve problems they are unable to solve in a single turn. This capability generalises across domains; SML on math problems produces models that better use feedback to solve coding problems and vice versa. Furthermore, despite being trained only on fully-specified problems, these models are better able to solve underspecified tasks where critical information is revealed over multiple turns. When faced with this ambiguity, SML-trained models make fewer premature answer attempts and are more likely to ask for the information they need. This work presents a scalable approach to developing AI systems that effectively learn from language feedback.

Michael Jungo, Beat Wolf, Andrii Maksai et al.

On-line handwritten character segmentation is often associated with handwriting recognition and even though recognition models include mechanisms to locate relevant positions during the recognition process, it is typically insufficient to produce a precise segmentation. Decoupling the segmentation from the recognition unlocks the potential to further utilize the result of the recognition. We specifically focus on the scenario where the transcription is known beforehand, in which case the character segmentation becomes an assignment problem between sampling points of the stylus trajectory and characters in the text. Inspired by the $k$-means clustering algorithm, we view it from the perspective of cluster assignment and present a Transformer-based architecture where each cluster is formed based on a learned character query in the Transformer decoder block. In order to assess the quality of our approach, we create character segmentation ground truths for two popular on-line handwriting datasets, IAM-OnDB and HANDS-VNOnDB, and evaluate multiple methods on them, demonstrating that our approach achieves the overall best results.

Aleksandr Timofeev, Anastasiia Fadeeva, Andrei Afonin et al.

As text generative models can give increasingly long answers, we tackle the problem of synthesizing long text in digital ink. We show that the commonly used models for this task fail to generalize to long-form data and how this problem can be solved by augmenting the training data, changing the model architecture and the inference procedure. These methods use contrastive learning technique and are tailored specifically for the handwriting domain. They can be applied to any encoder-decoder model that works with digital ink. We demonstrate that our method reduces the character error rate on long-form English data by half compared to baseline RNN and by 16% compared to the previous approach that aims at addressing the same problem. We show that all three parts of the method improve recognizability of generated inks. In addition, we evaluate synthesized data in a human study and find that people perceive most of generated data as real.

Andrei Afonin, Andrii Maksai, Aleksandr Timofeev et al.

Digital ink (online handwriting) generation has a number of potential applications for creating user-visible content, such as handwriting autocompletion, spelling correction, and beautification. Writing is personal and usually the processing is done on-device. Ink generative models thus need to produce high quality content quickly, in a resource constrained environment. In this work, we study ways to maximize the quality of the output of a trained digital ink generative model, while staying within an inference time budget. We use and compare the effect of multiple sampling and ranking techniques, in the first ablation study of its kind in the digital ink domain. We confirm our findings on multiple datasets - writing in English and Vietnamese, as well as mathematical formulas - using two model types and two common ink data representations. In all combinations, we report a meaningful improvement in the recognizability of the synthetic inks, in some cases more than halving the character error rate metric, and describe a way to select the optimal combination of sampling and ranking techniques for any given computational budget.

Martin Klissarov, Jonathan Cook, Diego Antognini et al.

Adapting one's thought process based on corrective feedback is an essential ability in human learning, particularly in collaborative settings. In contrast, the current large language model training paradigm relies heavily on modeling vast, static corpora. While effective for knowledge acquisition, it overlooks the interactive feedback loops essential for models to adapt dynamically to their context. In this work, we propose a framework that treats this interactive in-context learning ability not as an emergent property, but as a distinct, trainable skill. We introduce a scalable method that transforms single-turn verifiable tasks into multi-turn didactic interactions driven by information asymmetry. We first show that current flagship models struggle to integrate corrective feedback on hard reasoning tasks. We then demonstrate that models trained with our approach dramatically improve the ability to interactively learn from language feedback. More specifically, the multi-turn performance of a smaller model nearly reaches that of a model an order of magnitude larger. We also observe robust out-of-distribution generalization: interactive training on math problems transfers to diverse domains like coding, puzzles and maze navigation. Our qualitative analysis suggests that this improvement is due to an enhanced in-context plasticity. Finally, we show that this paradigm offers a unified path to self-improvement. By training the model to predict the teacher's critiques, effectively modeling the feedback environment, we convert this external signal into an internal capability, allowing the model to self-correct even without a teacher.

Arthur Gassner, Claudiu Musat, Alexandru Rusu et al.

Many applications require accurate indoor localization. Fingerprint-based localization methods propose a solution to this problem, but rely on a radio map that is effort-intensive to acquire. We automate the radio map acquisition phase using a software-defined radio (SDR) and a wheeled robot. Furthermore, we open-source a radio map acquired with our automated tool for a 3GPP Long-Term Evolution (LTE) wireless link. To the best of our knowledge, this is the first publicly available radio map containing channel state information (CSI). Finally, we describe first localization experiments on this radio map using a convolutional neural network to regress for location coordinates.

David Dao, Dan Alistarh, Claudiu Musat et al.

It is safe to assume that, for the foreseeable future, machine learning, especially deep learning will remain both data- and computation-hungry. In this paper, we ask: Can we build a global exchange where everyone can contribute computation and data to train the next generation of machine learning applications? We present an early, but running prototype of DataBright, a system that turns the creation of training examples and the sharing of computation into an investment mechanism. Unlike most crowdsourcing platforms, where the contributor gets paid when they submit their data, DataBright pays dividends whenever a contributor's data or hardware is used by someone to train a machine learning model. The contributor becomes a shareholder in the dataset they created. To enable the measurement of usage, a computation platform that contributors can trust is also necessary. DataBright thus merges both a data market and a trusted computation market. We illustrate that trusted computation can enable the creation of an AI market, where each data point has an exact value that should be paid to its creator. DataBright allows data creators to retain ownership of their contribution and attaches to it a measurable value. The value of the data is given by its utility in subsequent distributed computation done on the DataBright computation market. The computation market allocates tasks and subsequent payments to pooled hardware. This leads to the creation of a decentralized AI cloud. Our experiments show that trusted hardware such as Intel SGX can be added to the usual ML pipeline with no additional costs. We use this setting to orchestrate distributed computation that enables the creation of a computation market. DataBright is available for download at https://github.com/ds3lab/databright.

Anastasiia Fadeeva, Philippe Schlattner, Andrii Maksai et al.

The adoption of tablets with touchscreens and styluses is increasing, and a key feature is converting handwriting to text, enabling search, indexing, and AI assistance. Meanwhile, vision-language models (VLMs) are now the go-to solution for image understanding, thanks to both their state-of-the-art performance across a variety of tasks and the simplicity of a unified approach to training, fine-tuning, and inference. While VLMs obtain high performance on image-based tasks, they perform poorly on handwriting recognition when applied naively, i.e., by rendering handwriting as an image and performing optical character recognition (OCR). In this paper, we study online handwriting recognition with VLMs, going beyond naive OCR. We propose a novel tokenized representation of digital ink (online handwriting) that includes both a time-ordered sequence of strokes as text, and as image. We show that this representation yields results comparable to or better than state-of-the-art online handwriting recognizers. Wide applicability is shown through results with two different VLM families, on multiple public datasets. Our approach can be applied to off-the-shelf VLMs, does not require any changes in their architecture, and can be used in both fine-tuning and parameter-efficient tuning. We perform a detailed ablation study to identify the key elements of the proposed representation.

Blagoj Mitrevski, Arina Rak, Julian Schnitzler et al.

Digital note-taking is gaining popularity, offering a durable, editable, and easily indexable way of storing notes in a vectorized form, known as digital ink. However, a substantial gap remains between this way of note-taking and traditional pen-and-paper note-taking, a practice that is still favored by a vast majority. Our work InkSight, aims to bridge the gap by empowering physical note-takers to effortlessly convert their work (offline handwriting) to digital ink (online handwriting), a process we refer to as derendering. Prior research on the topic has focused on the geometric properties of images, resulting in limited generalization beyond their training domains. Our approach combines reading and writing priors, allowing training a model in the absence of large amounts of paired samples, which are difficult to obtain. To our knowledge, this is the first work that effectively derenders handwritten text in arbitrary photos with diverse visual characteristics and backgrounds. Furthermore, it generalizes beyond its training domain into simple sketches. Our human evaluation reveals that 87% of the samples produced by our model on the challenging HierText dataset are considered as a valid tracing of the input image and 67% look like a pen trajectory traced by a human.

Anastasiia Fadeeva, Vincent Coriou, Diego Antognini et al.

Tablets and styluses are increasingly popular for taking notes. To optimize this experience and ensure a smooth and efficient workflow, it's important to develop methods for accurately interpreting and understanding the content of handwritten digital notes. We introduce a foundational model called InkFM for analyzing full pages of handwritten content. Trained on a diverse mixture of tasks, this model offers a unique combination of capabilities: recognizing text in 28 different scripts, mathematical expressions recognition, and segmenting pages into distinct elements like text and drawings. Our results demonstrate that these tasks can be effectively unified within a single model, achieving SoTA text line segmentation out-of-the-box quality surpassing public baselines like docTR. Fine- or LoRA-tuning our base model on public datasets further improves the quality of page segmentation, achieves state-of the art text recognition (DeepWriting, CASIA, SCUT, and Mathwriting datasets) and sketch classification (QuickDraw). This adaptability of InkFM provides a powerful starting point for developing applications with handwritten input.

Andrii Maksai, Henry Rowley, Jesse Berent et al.

We introduce Inkorrect, a data- and label-efficient approach for online handwriting (Digital Ink) spelling correction - DISC. Unlike previous work, the proposed method does not require multiple samples from the same writer, or access to character level segmentation. We show that existing automatic evaluation metrics do not fully capture and are not correlated with the human perception of the quality of the spelling correction, and propose new ones that correlate with human perception. We additionally surface an interesting phenomenon: a trade-off between the similarity and recognizability of the spell-corrected inks. We further create a family of models corresponding to different points on the Pareto frontier between those two axes. We show that Inkorrect's Pareto frontier dominates the points that correspond to prior work.

Martin Milenkoski, Diego Antognini, Claudiu Musat

In this paper, we describe a method to tackle data sparsity and create recommendations in domains with limited knowledge about user preferences. We expand the variational autoencoder collaborative filtering from a single-domain to a multi-domain setting. The intuition is that user-item interactions in a source domain can augment the recommendation quality in a target domain. The intuition can be taken to its extreme, where, in a cross-domain setup, the user history in a source domain is enough to generate high-quality recommendations in a target one. We thus create a Product-of-Experts (POE) architecture for recommendations that jointly models user-item interactions across multiple domains. The method is resilient to missing data for one or more of the domains, which is a situation often found in real life. We present results on two widely-used datasets - Amazon and Yelp, which support the claim that holistic user preference knowledge leads to better recommendations. Surprisingly, we find that in some cases, a POE recommender that does not access the target domain user representation can surpass a strong VAE recommender baseline trained on the target domain.

Milena Filipovic, Blagoj Mitrevski, Diego Antognini et al.

Recommender systems research tends to evaluate model performance offline and on randomly sampled targets, yet the same systems are later used to predict user behavior sequentially from a fixed point in time. Simulating online recommender system performance is notoriously difficult and the discrepancy between online and offline behaviors is typically not accounted for in offline evaluations. This disparity permits weaknesses to go unnoticed until the model is deployed in a production setting. In this paper, we first demonstrate how omitting temporal context when evaluating recommender system performance leads to false confidence. To overcome this, we postulate that offline evaluation protocols can only model real-life use-cases if they account for temporal context. Next, we propose a training procedure to further embed the temporal context in existing models. We use a multi-objective approach to introduce temporal context into traditionally time-unaware recommender systems and confirm its advantage via the proposed evaluation protocol. Finally, we validate that the Pareto Fronts obtained with the added objective dominate those produced by state-of-the-art models that are only optimized for accuracy on three real-world publicly available datasets. The results show that including our temporal objective can improve recall@20 by up to 20%.

Blagoj Mitrevski, Milena Filipovic, Diego Antognini et al.

Multi-objective gradient methods are becoming the standard for solving multi-objective problems. Among others, they show promising results in developing multi-objective recommender systems with both correlated and conflicting objectives. Classic multi-gradient~descent usually relies on the combination of the gradients, not including the computation of first and second moments of the gradients. This leads to a brittle behavior and misses important areas in the solution space. In this work, we create a multi-objective model-agnostic Adamize method that leverages the benefits of the Adam optimizer in single-objective problems. This corrects and stabilizes~the~gradients of every objective before calculating a common gradient descent vector that optimizes all the objectives simultaneously. We evaluate the benefits of Multi-objective Adamize on two multi-objective recommender systems and for three different objective combinations, both correlated or conflicting. We report significant improvements, measured with three different Pareto front metrics: hypervolume, coverage, and spacing. Finally, we show that the \textit{Adamized} Pareto front strictly dominates the previous one on multiple objective pairs.

Kirtan Padh, Diego Antognini, Emma Lejal Glaude et al.

The goal of fairness in classification is to learn a classifier that does not discriminate against groups of individuals based on sensitive attributes, such as race and gender. One approach to designing fair algorithms is to use relaxations of fairness notions as regularization terms or in a constrained optimization problem. We observe that the hyperbolic tangent function can approximate the indicator function. We leverage this property to define a differentiable relaxation that approximates fairness notions provably better than existing relaxations. In addition, we propose a model-agnostic multi-objective architecture that can simultaneously optimize for multiple fairness notions and multiple sensitive attributes and supports all statistical parity-based notions of fairness. We use our relaxation with the multi-objective architecture to learn fair classifiers. Experiments on public datasets show that our method suffers a significantly lower loss of accuracy than current debiasing algorithms relative to the unconstrained model.

Mohammad Saeed Rad, Thomas Yu, Claudiu Musat et al.

Super-resolution (SR) has traditionally been based on pairs of high-resolution images (HR) and their low-resolution (LR) counterparts obtained artificially with bicubic downsampling. However, in real-world SR, there is a large variety of realistic image degradations and analytically modeling these realistic degradations can prove quite difficult. In this work, we propose to handle real-world SR by splitting this ill-posed problem into two comparatively more well-posed steps. First, we train a network to transform real LR images to the space of bicubically downsampled images in a supervised manner, by using both real LR/HR pairs and synthetic pairs. Second, we take a generic SR network trained on bicubically downsampled images to super-resolve the transformed LR image. The first step of the pipeline addresses the problem by registering the large variety of degraded images to a common, well understood space of images. The second step then leverages the already impressive performance of SR on bicubically downsampled images, sidestepping the issues of end-to-end training on datasets with many different image degradations. We demonstrate the effectiveness of our proposed method by comparing it to recent methods in real-world SR and show that our proposed approach outperforms the state-of-the-art works in terms of both qualitative and quantitative results, as well as results of an extensive user study conducted on several real image datasets.

Diego Antognini, Claudiu Musat, Boi Faltings

Using personalized explanations to support recommendations has been shown to increase trust and perceived quality. However, to actually obtain better recommendations, there needs to be a means for users to modify the recommendation criteria by interacting with the explanation. We present a novel technique using aspect markers that learns to generate personalized explanations of recommendations from review texts, and we show that human users significantly prefer these explanations over those produced by state-of-the-art techniques. Our work's most important innovation is that it allows users to react to a recommendation by critiquing the textual explanation: removing (symmetrically adding) certain aspects they dislike or that are no longer relevant (symmetrically that are of interest). The system updates its user model and the resulting recommendations according to the critique. This is based on a novel unsupervised critiquing method for single- and multi-step critiquing with textual explanations. Experiments on two real-world datasets show that our system is the first to achieve good performance in adapting to the preferences expressed in multi-step critiquing.

Giuseppe Russo, Nora Hollenstein, Claudiu Musat et al.

We introduce CGA, a conditional VAE architecture, to control, generate, and augment text. CGA is able to generate natural English sentences controlling multiple semantic and syntactic attributes by combining adversarial learning with a context-aware loss and a cyclical word dropout routine. We demonstrate the value of the individual model components in an ablation study. The scalability of our approach is ensured through a single discriminator, independently of the number of attributes. We show high quality, diversity and attribute control in the generated sentences through a series of automatic and human assessments. As the main application of our work, we test the potential of this new NLG model in a data augmentation scenario. In a downstream NLP task, the sentences generated by our CGA model show significant improvements over a strong baseline, and a classification performance often comparable to adding same amount of additional real data.

Larissa Schmidt, Lucy Linder, Sandra Djambazovska et al.

We introduce a dictionary containing forms of common words in various Swiss German dialects normalized into High German. As Swiss German is, for now, a predominantly spoken language, there is a significant variation in the written forms, even between speakers of the same dialect. To alleviate the uncertainty associated with this diversity, we complement the pairs of Swiss German - High German words with the Swiss German phonetic transcriptions (SAMPA). This dictionary becomes thus the first resource to combine large-scale spontaneous translation with phonetic transcriptions. Moreover, we control for the regional distribution and insure the equal representation of the major Swiss dialects. The coupling of the phonetic and written Swiss German forms is powerful. We show that they are sufficient to train a Transformer-based phoneme to grapheme model that generates credible novel Swiss German writings. In addition, we show that the inverse mapping - from graphemes to phonemes - can be modeled with a transformer trained with the novel dictionary. This generation of pronunciations for previously unknown words is key in training extensible automated speech recognition (ASR) systems, which are key beneficiaries of this dictionary.

Guillaume Raille, Sandra Djambazovska, Claudiu Musat

Data augmentation promises to alleviate data scarcity. This is most important in cases where the initial data is in short supply. This is, for existing methods, also where augmenting is the most difficult, as learning the full data distribution is impossible. For natural language, sentence editing offers a solution - relying on small but meaningful changes to the original ones. Learning which changes are meaningful also requires large amounts of training data. We thus aim to learn this in a source domain where data is abundant and apply it in a different, target domain, where data is scarce - cross-domain augmentation. We create the Edit-transformer, a Transformer-based sentence editor that is significantly faster than the state of the art and also works cross-domain. We argue that, due to its structure, the Edit-transformer is better suited for cross-domain environments than its edit-based predecessors. We show this performance gap on the Yelp-Wikipedia domain pairs. Finally, we show that due to this cross-domain performance advantage, the Edit-transformer leads to meaningful performance gains in several downstream tasks.

Nikola Milojkovic, Diego Antognini, Giancarlo Bergamin et al.

Recommender systems need to mirror the complexity of the environment they are applied in. The more we know about what might benefit the user, the more objectives the recommender system has. In addition there may be multiple stakeholders - sellers, buyers, shareholders - in addition to legal and ethical constraints. Simultaneously optimizing for a multitude of objectives, correlated and not correlated, having the same scale or not, has proven difficult so far. We introduce a stochastic multi-gradient descent approach to recommender systems (MGDRec) to solve this problem. We show that this exceeds state-of-the-art methods in traditional objective mixtures, like revenue and recall. Not only that, but through gradient normalization we can combine fundamentally different objectives, having diverse scales, into a single coherent framework. We show that uncorrelated objectives, like the proportion of quality products, can be improved alongside accuracy. Through the use of stochasticity, we avoid the pitfalls of calculating full gradients and provide a clear setting for its applicability.

Lucy Linder, Michael Jungo, Jean Hennebert et al.

This paper presents SwissCrawl, the largest Swiss German text corpus to date. Composed of more than half a million sentences, it was generated using a customized web scraping tool that could be applied to other low-resource languages as well. The approach demonstrates how freely available web pages can be used to construct comprehensive text corpora, which are of fundamental importance for natural language processing. In an experimental evaluation, we show that using the new corpus leads to significant improvements for the task of language modeling. To capture new content, our approach will run continuously to keep increasing the corpus over time.

Diego Antognini, Claudiu Musat, Boi Faltings

Automated predictions require explanations to be interpretable by humans. Past work used attention and rationale mechanisms to find words that predict the target variable of a document. Often though, they result in a tradeoff between noisy explanations or a drop in accuracy. Furthermore, rationale methods cannot capture the multi-faceted nature of justifications for multiple targets, because of the non-probabilistic nature of the mask. In this paper, we propose the Multi-Target Masker (MTM) to address these shortcomings. The novelty lies in the soft multi-dimensional mask that models a relevance probability distribution over the set of target variables to handle ambiguities. Additionally, two regularizers guide MTM to induce long, meaningful explanations. We evaluate MTM on two datasets and show, using standard metrics and human annotations, that the resulting masks are more accurate and coherent than those generated by the state-of-the-art methods. Moreover, MTM is the first to also achieve the highest F1 scores for all the target variables simultaneously.

Noémien Kocher, Christian Scuito, Lorenzo Tarantino et al.

In sequence modeling tasks the token order matters, but this information can be partially lost due to the discretization of the sequence into data points. In this paper, we study the imbalance between the way certain token pairs are included in data points and others are not. We denote this a token order imbalance (TOI) and we link the partial sequence information loss to a diminished performance of the system as a whole, both in text and speech processing tasks. We then provide a mechanism to leverage the full token order information -Alleviated TOI- by iteratively overlapping the token composition of data points. For recurrent networks, we use prime numbers for the batch size to avoid redundancies when building batches from overlapped data points. The proposed method achieved state of the art performance in both text and speech related tasks.

Mohammad Saeed Rad, Behzad Bozorgtabar, Claudiu Musat et al.

Despite significant progress toward super resolving more realistic images by deeper convolutional neural networks (CNNs), reconstructing fine and natural textures still remains a challenging problem. Recent works on single image super resolution (SISR) are mostly based on optimizing pixel and content wise similarity between recovered and high-resolution (HR) images and do not benefit from recognizability of semantic classes. In this paper, we introduce a novel approach using categorical information to tackle the SISR problem; we present a decoder architecture able to extract and use semantic information to super-resolve a given image by using multitask learning, simultaneously for image super-resolution and semantic segmentation. To explore categorical information during training, the proposed decoder only employs one shared deep network for two task-specific output layers. At run-time only layers resulting HR image are used and no segmentation label is required. Extensive perceptual experiments and a user study on images randomly selected from COCO-Stuff dataset demonstrate the effectiveness of our proposed method and it outperforms the state-of-the-art methods.

Athanasios Giannakopoulos, Maxime Coriou, Andreea Hossmann et al.

State-of-the-art methods for text classification include several distinct steps of pre-processing, feature extraction and post-processing. In this work, we focus on end-to-end neural architectures and show that the best performance in text classification is obtained by combining information from different neural modules. Concretely, we combine convolution, recurrent and attention modules with ensemble methods and show that they are complementary. We introduce ECGA, an end-to-end go-to architecture for novel text classification tasks. We prove that it is efficient and robust, as it attains or surpasses the state-of-the-art on varied datasets, including both low and high data regimes.

Meryem M'hamdi, Robert West, Andreea Hossmann et al.

Most work in text classification and Natural Language Processing (NLP) focuses on English or a handful of other languages that have text corpora of hundreds of millions of words. This is creating a new version of the digital divide: the artificial intelligence (AI) divide. Transfer-based approaches, such as Cross-Lingual Text Classification (CLTC) - the task of categorizing texts written in different languages into a common taxonomy, are a promising solution to the emerging AI divide. Recent work on CLTC has focused on demonstrating the benefits of using bilingual word embeddings as features, relegating the CLTC problem to a mere benchmark based on a simple averaged perceptron. In this paper, we explore more extensively and systematically two flavors of the CLTC problem: news topic classification and textual churn intent detection (TCID) in social media. In particular, we test the hypothesis that embeddings with context are more effective, by multi-tasking the learning of multilingual word embeddings and text classification; we explore neural architectures for CLTC; and we move from bi- to multi-lingual word embeddings. For all architectures, types of word embeddings and datasets, we notice a consistent gain trend in favor of multilingual joint training, especially for low-resourced languages.

Khalil Mrini, Claudiu Musat, Michael Baeriswyl et al.

We propose a method to create document representations that reflect their internal structure. We modify Tree-LSTMs to hierarchically merge basic elements such as words and sentences into blocks of increasing complexity. Our Structure Tree-LSTM implements a hierarchical attention mechanism over individual components and combinations thereof. We thus emphasize the usefulness of Tree-LSTMs for texts larger than a sentence. We show that structure-aware encoders can be used to improve the performance of document classification. We demonstrate that our method is resilient to changes to the basic building blocks, as it performs well with both sentence and word embeddings. The Structure Tree-LSTM outperforms all the baselines on two datasets by leveraging structural clues. We show our model's interpretability by visualizing how our model distributes attention inside a document. On a third dataset from the medical domain, our model achieves competitive performance with the state of the art. This result shows the Structure Tree-LSTM can leverage dependency relations other than text structure, such as a set of reports on the same patient.

Yassine Benyahia, Kaicheng Yu, Kamil Bennani-Smires et al.

We identify a phenomenon, which we refer to as multi-model forgetting, that occurs when sequentially training multiple deep networks with partially-shared parameters; the performance of previously-trained models degrades as one optimizes a subsequent one, due to the overwriting of shared parameters. To overcome this, we introduce a statistically-justified weight plasticity loss that regularizes the learning of a model's shared parameters according to their importance for the previous models, and demonstrate its effectiveness when training two models sequentially and for neural architecture search. Adding weight plasticity in neural architecture search preserves the best models to the end of the search and yields improved results in both natural language processing and computer vision tasks.

Kaicheng Yu, Christian Sciuto, Martin Jaggi et al.

Neural Architecture Search (NAS) aims to facilitate the design of deep networks for new tasks. Existing techniques rely on two stages: searching over the architecture space and validating the best architecture. NAS algorithms are currently compared solely based on their results on the downstream task. While intuitive, this fails to explicitly evaluate the effectiveness of their search strategies. In this paper, we propose to evaluate the NAS search phase. To this end, we compare the quality of the solutions obtained by NAS search policies with that of random architecture selection. We find that: (i) On average, the state-of-the-art NAS algorithms perform similarly to the random policy; (ii) the widely-used weight sharing strategy degrades the ranking of the NAS candidates to the point of not reflecting their true performance, thus reducing the effectiveness of the search process. We believe that our evaluation framework will be key to designing NAS strategies that consistently discover architectures superior to random ones.

Bojan Petrovski, Ignacio Aguado, Andreea Hossmann et al.

Most of the world's data is stored in relational databases. Accessing these requires specialized knowledge of the Structured Query Language (SQL), putting them out of the reach of many people. A recent research thread in Natural Language Processing (NLP) aims to alleviate this problem by automatically translating natural language questions into SQL queries. While the proposed solutions are a great start, they lack robustness and do not easily generalize: the methods require high quality descriptions of the database table columns, and the most widely used training dataset, WikiSQL, is heavily biased towards using those descriptions as part of the questions. In this work, we propose solutions to both problems: we entirely eliminate the need for column descriptions, by relying solely on their contents, and we augment the WikiSQL dataset by paraphrasing column names to reduce bias. We show that the accuracy of existing methods drops when trained on our augmented, column-agnostic dataset, and that our own method reaches state of the art accuracy, while relying on column contents only.

Christian Abbet, Meryem M'hamdi, Athanasios Giannakopoulos et al.

We propose a new method to detect when users express the intent to leave a service, also known as churn. While previous work focuses solely on social media, we show that this intent can be detected in chatbot conversations. As companies increasingly rely on chatbots they need an overview of potentially churny users. To this end, we crowdsource and publish a dataset of churn intent expressions in chatbot interactions in German and English. We show that classifiers trained on social media data can detect the same intent in the context of chatbots. We introduce a classification architecture that outperforms existing work on churn intent detection in social media. Moreover, we show that, using bilingual word embeddings, a system trained on combined English and German data outperforms monolingual approaches. As the only existing dataset is in English, we crowdsource and publish a novel dataset of German tweets. We thus underline the universal aspect of the problem, as examples of churn intent in English help us identify churn in German tweets and chatbot conversations.

André Cibils, Claudiu Musat, Andreea Hossman et al.

Text summarization condenses a text to a shorter version while retaining the important informations. Abstractive summarization is a recent development that generates new phrases, rather than simply copying or rephrasing sentences within the original text. Recently neural sequence-to-sequence models have achieved good results in the field of abstractive summarization, which opens new possibilities and applications for industrial purposes. However, most practitioners observe that these models still use large parts of the original text in the output summaries, making them often similar to extractive frameworks. To address this drawback, we first introduce a new metric to measure how much of a summary is extracted from the input text. Secondly, we present a novel method, that relies on a diversity factor in computing the neural network loss, to improve the diversity of the summaries generated by any neural abstractive model implementing beam search. Finally, we show that this method not only makes the system less extractive, but also improves the overall rouge score of state-of-the-art methods by at least 2 points.

Mladen Dimovski, Claudiu Musat, Vladimir Ilievski et al.

Spoken language understanding (SLU) systems, such as goal-oriented chatbots or personal assistants, rely on an initial natural language understanding (NLU) module to determine the intent and to extract the relevant information from the user queries they take as input. SLU systems usually help users to solve problems in relatively narrow domains and require a large amount of in-domain training data. This leads to significant data availability issues that inhibit the development of successful systems. To alleviate this problem, we propose a technique of data selection in the low-data regime that enables us to train with fewer labeled sentences, thus smaller labelling costs. We propose a submodularity-inspired data ranking function, the ratio-penalty marginal gain, for selecting data points to label based only on the information extracted from the textual embedding space. We show that the distances in the embedding space are a viable source of information that can be used for data selection. Our method outperforms two known active learning techniques and enables cost-efficient training of the NLU unit. Moreover, our proposed selection technique does not need the model to be retrained in between the selection steps, making it time efficient as well.

Vladimir Ilievski, Claudiu Musat, Andreea Hossmann et al.

Goal-Oriented (GO) Dialogue Systems, colloquially known as goal oriented chatbots, help users achieve a predefined goal (e.g. book a movie ticket) within a closed domain. A first step is to understand the user's goal by using natural language understanding techniques. Once the goal is known, the bot must manage a dialogue to achieve that goal, which is conducted with respect to a learnt policy. The success of the dialogue system depends on the quality of the policy, which is in turn reliant on the availability of high-quality training data for the policy learning method, for instance Deep Reinforcement Learning. Due to the domain specificity, the amount of available data is typically too low to allow the training of good dialogue policies. In this paper we introduce a transfer learning method to mitigate the effects of the low in-domain data availability. Our transfer learning based approach improves the bot's success rate by 20% in relative terms for distant domains and we more than double it for close domains, compared to the model without transfer learning. Moreover, the transfer learning chatbots learn the policy up to 5 to 10 times faster. Finally, as the transfer learning approach is complementary to additional processing such as warm-starting, we show that their joint application gives the best outcomes.

Kamil Bennani-Smires, Claudiu Musat, Andreea Hossmann et al.

The dramatic success of deep neural networks across multiple application areas often relies on experts painstakingly designing a network architecture specific to each task. To simplify this process and make it more accessible, an emerging research effort seeks to automate the design of neural network architectures, using e.g. evolutionary algorithms or reinforcement learning or simple search in a constrained space of neural modules. Considering the typical size of the search space (e.g. $10^{10}$ candidates for a $10$-layer network) and the cost of evaluating a single candidate, current architecture search methods are very restricted. They either rely on static pre-built modules to be recombined for the task at hand, or they define a static hand-crafted framework within which they can generate new architectures from the simplest possible operations. In this paper, we relax these restrictions, by capitalizing on the collective wisdom contained in the plethora of neural networks published in online code repositories. Concretely, we (a) extract and publish GitGraph, a corpus of neural architectures and their descriptions; (b) we create problem-specific neural architecture search spaces, implemented as a textual search mechanism over GitGraph; (c) we propose a method of identifying unique common subgraphs within the architectures solving each problem (e.g., image processing, reinforcement learning), that can then serve as modules in the newly created problem specific neural search space.

Kamil Bennani-Smires, Claudiu Musat, Andreea Hossmann et al.

Keyphrase extraction is the task of automatically selecting a small set of phrases that best describe a given free text document. Supervised keyphrase extraction requires large amounts of labeled training data and generalizes very poorly outside the domain of the training data. At the same time, unsupervised systems have poor accuracy, and often do not generalize well, as they require the input document to belong to a larger corpus also given as input. Addressing these drawbacks, in this paper, we tackle keyphrase extraction from single documents with EmbedRank: a novel unsupervised method, that leverages sentence embeddings. EmbedRank achieves higher F-scores than graph-based state of the art systems on standard datasets and is suitable for real-time processing of large amounts of Web data. With EmbedRank, we also explicitly increase coverage and diversity among the selected keyphrases by introducing an embedding-based maximal marginal relevance (MMR) for new phrases. A user study including over 200 votes showed that, although reducing the phrases' semantic overlap leads to no gains in F-score, our high diversity selection is preferred by humans.

Pierre-Edouard Honnet, Andrei Popescu-Belis, Claudiu Musat et al.

The goal of this work is to design a machine translation (MT) system for a low-resource family of dialects, collectively known as Swiss German, which are widely spoken in Switzerland but seldom written. We collected a significant number of parallel written resources to start with, up to a total of about 60k words. Moreover, we identified several other promising data sources for Swiss German. Then, we designed and compared three strategies for normalizing Swiss German input in order to address the regional diversity. We found that character-based neural MT was the best solution for text normalization. In combination with phrase-based statistical MT, our solution reached 36% BLEU score when translating from the Bernese dialect. This value, however, decreases as the testing data becomes more remote from the training one, geographically and topically. These resources and normalization techniques are a first step towards full MT of Swiss German dialects.

Athanasios Giannakopoulos, Diego Antognini, Claudiu Musat et al.

Aspect Term Extraction (ATE) detects opinionated aspect terms in sentences or text spans, with the end goal of performing aspect-based sentiment analysis. The small amount of available datasets for supervised ATE and the fact that they cover only a few domains raise the need for exploiting other data sources in new and creative ways. Publicly available review corpora contain a plethora of opinionated aspect terms and cover a larger domain spectrum. In this paper, we first propose a method for using such review corpora for creating a new dataset for ATE. Our method relies on an attention mechanism to select sentences that have a high likelihood of containing actual opinionated aspects. We thus improve the quality of the extracted aspects. We then use the constructed dataset to train a model and perform ATE with distant supervision. By evaluating on human annotated datasets, we prove that our method achieves a significantly improved performance over various unsupervised and supervised baselines. Finally, we prove that sentence selection matters when it comes to creating new datasets for ATE. Specifically, we show that, using a set of selected sentences leads to higher ATE performance compared to using the whole sentence set.

Athanasios Giannakopoulos, Claudiu Musat, Andreea Hossmann et al.

Aspect Term Extraction (ATE) identifies opinionated aspect terms in texts and is one of the tasks in the SemEval Aspect Based Sentiment Analysis (ABSA) contest. The small amount of available datasets for supervised ATE and the costly human annotation for aspect term labelling give rise to the need for unsupervised ATE. In this paper, we introduce an architecture that achieves top-ranking performance for supervised ATE. Moreover, it can be used efficiently as feature extractor and classifier for unsupervised ATE. Our second contribution is a method to automatically construct datasets for ATE. We train a classifier on our automatically labelled datasets and evaluate it on the human annotated SemEval ABSA test sets. Compared to a strong rule-based baseline, we obtain a dramatically higher F-score and attain precision values above 80%. Our unsupervised method beats the supervised ABSA baseline from SemEval, while preserving high precision scores.