Probing Tasks Research Articles

Topic aware probing: From sentence length prediction to idiom identification how reliant are neural language models on topic?

Abstract Transformer-based neural language models achieve state-of-the-art performance on various natural language processing tasks. However, an open question is the extent to which these models rely on word-order/syntactic or word co-occurrence/topic-based information when processing natural language. This work contributes to this debate by addressing the question of whether these models primarily use topic as a signal, by exploring the relationship between Transformer-based models’ (BERT and RoBERTa’s) performance on a range of probing tasks in English, from simple lexical tasks such as sentence length prediction to complex semantic tasks such as idiom token identification, and the sensitivity of these tasks to the topic information. To this end, we propose a novel probing method which we call topic-aware probing. Our initial results indicate that Transformer-based models encode both topic and non-topic information in their intermediate layers, but also that the facility of these models to distinguish idiomatic usage is primarily based on their ability to identify and encode topic. Furthermore, our analysis of these models’ performance on other standard probing tasks suggests that tasks that are relatively insensitive to the topic information are also tasks that are relatively difficult for these models.

Unveiling Code Pre-Trained Models: Investigating Syntax and Semantics Capacities

Code models have made significant advancements in code intelligence by encoding knowledge about programming languages. While previous studies have explored the capabilities of these models in learning code syntax, there has been limited investigation on their ability to understand code semantics. Additionally, existing analyses assume the number of edges between nodes at the abstract syntax tree (AST) is related to syntax distance, and also often require transforming the high-dimensional space of deep learning models to a low-dimensional one, which may introduce inaccuracies. To study how code models represent code syntax and semantics, we conduct a comprehensive analysis of 7 code models, including four representative code pre-trained models (CodeBERT, GraphCodeBERT, CodeT5, and UnixCoder) and three large language models (StarCoder, CodeLlama and CodeT5+). We design four probing tasks to assess the models’ capacities in learning both code syntax and semantics. These probing tasks reconstruct code syntax and semantics structures (AST, CDG, DDG and CFG) in the representation space. These structures are core concepts for code understanding. We also investigate the syntax token role in each token representation and the long dependency between the code tokens. Additionally, we analyze the distribution of attention weights related to code semantic structures. Through extensive analysis, our findings highlight the strengths and limitations of different code models in learning code syntax and semantics. The results demonstrate that these models excel in learning code syntax, successfully capturing the syntax relationships between tokens and the syntax roles of individual tokens. However, their performance in encoding code semantics varies. CodeT5 and CodeBERT demonstrate proficiency in capturing control and data dependencies, while UnixCoder shows weaker performance in this aspect. We do not observe LLMs generally performing much better than pre-trained models. The shallow layers of LLMs perform better than their deep layers. The investigation of attention weights reveals that different attention heads play distinct roles in encoding code semantics. Our research findings emphasize the need for further enhancements in code models to better learn code semantics. This study contributes to the understanding of code models’ abilities in syntax and semantics analysis. Our findings provide guidance for future improvements in code models, facilitating their effective application in various code-related tasks.

Rethinking the Exploitation of Monolingual Data for Low-Resource Neural Machine Translation

Abstract The utilization of monolingual data has been shown to be a promising strategy for addressing low-resource machine translation problems. Previous studies have demonstrated the effectiveness of techniques such as back-translation and self-supervised objectives, including masked language modeling, causal language modeling, and denoise autoencoding, in improving the performance of machine translation models. However, the manner in which these methods contribute to the success of machine translation tasks and how they can be effectively combined remains an under-researched area. In this study, we carry out a systematic investigation of the effects of these techniques on linguistic properties through the use of probing tasks, including source language comprehension, bilingual word alignment, and translation fluency. We further evaluate the impact of pre-training, back-translation, and multi-task learning on bitexts of varying sizes. Our findings inform the design of more effective pipelines for leveraging monolingual data in extremely low-resource and low-resource machine translation tasks. Experiment results show consistent performance gains in seven translation directions, which provide further support for our conclusions and understanding of the role of monolingual data in machine translation.

Morphosyntactic probing of multilingual BERT models

Abstract We introduce an extensive dataset for multilingual probing of morphological information in language models (247 tasks across 42 languages from 10 families), each consisting of a sentence with a target word and a morphological tag as the desired label, derived from the Universal Dependencies treebanks. We find that pre-trained Transformer models (mBERT and XLM-RoBERTa) learn features that attain strong performance across these tasks. We then apply two methods to locate, for each probing task, where the disambiguating information resides in the input. The first is a new perturbation method that “masks” various parts of context; the second is the classical method of Shapley values. The most intriguing finding that emerges is a strong tendency for the preceding context to hold more information relevant to the prediction than the following context.

Testing the Effectiveness of the Diagnostic Probing Paradigm on Italian Treebanks

The outstanding performance recently reached by neural language models (NLMs) across many natural language processing (NLP) tasks has steered the debate towards understanding whether NLMs implicitly learn linguistic competence. Probes, i.e., supervised models trained using NLM representations to predict linguistic properties, are frequently adopted to investigate this issue. However, it is still questioned if probing classification tasks really enable such investigation or if they simply hint at surface patterns in the data. This work contributes to this debate by presenting an approach to assessing the effectiveness of a suite of probing tasks aimed at testing the linguistic knowledge implicitly encoded by one of the most prominent NLMs, BERT. To this aim, we compared the performance of probes when predicting gold and automatically altered values of a set of linguistic features. Our experiments were performed on Italian and were evaluated across BERT’s layers and for sentences with different lengths. As a general result, we observed higher performance in the prediction of gold values, thus suggesting that the probing model is sensitive to the distortion of feature values. However, our experiments also showed that the length of a sentence is a highly influential factor that is able to confound the probing model’s predictions.

Visual Probing: Cognitive Framework for Explaining Self-Supervised Image Representations

Recently introduced self-supervised methods for image representation learning provide on par or superior results to their fully supervised competitors, yet the corresponding efforts to explain the self-supervised approaches lag behind. Motivated by this observation, we introduce a novel visual probing framework for explaining the self-supervised models by leveraging probing tasks employed previously in natural language processing. The probing tasks require knowledge about semantic relationships between image parts. Hence, we propose a systematic approach to obtain analogs of natural language in vision, such as visual words, context, and taxonomy. Our proposal is grounded in Marr's computational theory of vision and concerns features like textures, shapes, and lines. We show the effectiveness and applicability of those analogs in the context of explaining self-supervised representations. Our key findings emphasize that relations between language and vision can serve as an effective yet intuitive tool for discovering how machine learning models work, independently of data modality. Our work opens a plethora of research pathways towards more explainable and transparent AI.

On Robustness and Sensitivity of a Neural Language Model: A Case Study on Italian L1 Learner Errors

In this paper, we propose a comprehensive linguistic study aimed at assessing the implicit behavior of one of the most prominent Neural Language Models (NLM) based on Transformer architectures, BERT Devlin et al., when dealing with a particular source of noisy data, namely essays written by L1 Italian learners containing a variety of errors targeting grammar, orthography and lexicon. Differently from previous works, we focus on the pre-training stage and we devise two complementary evaluation tasks aimed at assessing the impact of errors on sentence-level inner representations in terms of semantic robustness and linguistic sensitivity. While the first evaluation perspective is meant to probe the model's ability to encode the semantic similarity between sentences also in the presence of errors, the second type of probing task evaluates the influence of errors on BERT's implicit knowledge of a set of raw and morpho-syntactic properties of a sentence. Our experiments show that BERT's ability to compute sentence similarity and to correctly encode multi-leveled linguistic information of a sentence are differently modulated by the category of errors and that the error hierarchies in terms of robustness and sensitivity change across layer-wise representations.

BiVaSE: A bilingual variational sentence encoder with randomly initialized Transformer layers

AbstractTransformer-based NLP models have achieved state-of-the-art results in many NLP tasks including text classification and text generation. However, the layers of these models do not output any explicit representations for texts units larger than tokens (e.g. sentences), although such representations are required to perform text classification. Sentence encodings are usually obtained by applying a pooling technique during fine-tuning on a specific task. In this paper, a new sentence encoder is introduced. Relying on an autoencoder architecture, it was trained to learn sentence representations from the very beginning of its training. The model was trained on bilingual data with variational Bayesian inference. Sentence representations were evaluated in downstream and linguistic probing tasks. Although the newly introduced encoder generally performs worse than well-known Transformer-based encoders, the experiments show that it was able to learn to incorporate linguistic information in the sentence representations.

Are Vision-Language Transformers Learning Multimodal Representations? A Probing Perspective

In recent years, joint text-image embeddings have significantly improved thanks to the development of transformer-based Vision-Language models. Despite these advances, we still need to better understand the representations produced by those models. In this paper, we compare pre-trained and fine-tuned representations at a vision, language and multimodal level. To that end, we use a set of probing tasks to evaluate the performance of state-of-the-art Vision-Language models and introduce new datasets specifically for multimodal probing. These datasets are carefully designed to address a range of multimodal capabilities while minimizing the potential for models to rely on bias. Although the results confirm the ability of Vision-Language models to understand color at a multimodal level, the models seem to prefer relying on bias in text data for object position and size. On semantically adversarial examples, we find that those models are able to pinpoint fine-grained multimodal differences. Finally, we also notice that fine-tuning a Vision-Language model on multimodal tasks does not necessarily improve its multimodal ability. We make all datasets and code available to replicate experiments.

DKPLM: Decomposable Knowledge-Enhanced Pre-trained Language Model for Natural Language Understanding

Knowledge-Enhanced Pre-trained Language Models (KEPLMs) are pre-trained models with relation triples injecting from knowledge graphs to improve language understanding abilities.Experiments show that our model outperforms other KEPLMs significantly over zero-shot knowledge probing tasks and multiple knowledge-aware language understanding tasks. To guarantee effective knowledge injection, previous studies integrate models with knowledge encoders for representing knowledge retrieved from knowledge graphs. The operations for knowledge retrieval and encoding bring significant computational burdens, restricting the usage of such models in real-world applications that require high inference speed. In this paper, we propose a novel KEPLM named DKPLM that decomposes knowledge injection process of the pre-trained language models in pre-training, fine-tuning and inference stages, which facilitates the applications of KEPLMs in real-world scenarios. Specifically, we first detect knowledge-aware long-tail entities as the target for knowledge injection, enhancing the KEPLMs' semantic understanding abilities and avoiding injecting redundant information. The embeddings of long-tail entities are replaced by ``pseudo token representations'' formed by relevant knowledge triples. We further design the relational knowledge decoding task for pre-training to force the models to truly understand the injected knowledge by relation triple reconstruction. Experiments show that our model outperforms other KEPLMs significantly over zero-shot knowledge probing tasks and multiple knowledge-aware language understanding tasks. We further show that DKPLM has a higher inference speed than other competing models due to the decomposing mechanism.

Multi-view self-attention networks

Self-attention networks (SANs) have attracted an amount of research attention for their outstanding performance under the machine translation community. Recent studies proved that SANs can be further improved by exploiting different inductive biases, each of which guides SANs to learn a specific view of the input sentence, e.g., short-term dependencies, forward and backward views, as well as phrasal patterns. However, less studies investigate how these inductive techniques complementarily improve the capability of SANs and this would be an interesting question to be answered. In this paper we selected five inductive biases which are simple and not over parameterized to investigate their complementarily. We further propose multi-view self-attention networks, which jointly learn different linguistic aspects of the input sentence under a unified framework. Specifically, we propose and exploit a variety of inductive biases to regularize the conventional attention distribution. Different views are then aggregated by a hybrid attention mechanism to quantify and leverage the specific views and their associated representation conveniently. Experiments on various translation tasks demonstrate that different views are able to progressively improve the performance of SANs, and the proposed approach outperforms both the strong Transformer baseline and related models on Transformer-base and Transformer-big settings. Extensive analyses on 10 linguistic probing tasks verify that different views indeed tend to extract distinct linguistic features and our method gives highly effective improvements in their integration.11We release the code of this work in https://github.com/NLP2CT/MV-Attn.

A Self-Supervised Representation Learning of Sentence Structure for Authorship Attribution

The syntactic structure of sentences in a document substantially informs about its authorial writing style. Sentence representation learning has been widely explored in recent years and it has been shown that it improves the generalization of different downstream tasks across many domains. Even though utilizing probing methods in several studies suggests that these learned contextual representations implicitly encode some amount of syntax, explicit syntactic information further improves the performance of deep neural models in the domain of authorship attribution. These observations have motivated us to investigate the explicit representation learning of syntactic structure of sentences. In this article, we propose a self-supervised framework for learning structural representations of sentences. The self-supervised network contains two components; a lexical sub-network and a syntactic sub-network which take the sequence of words and their corresponding structural labels as the input, respectively. Due to the n -to-1 mapping of words to their structural labels, each word will be embedded into a vector representation which mainly carries structural information. We evaluate the learned structural representations of sentences using different probing tasks, and subsequently utilize them in the authorship attribution task. Our experimental results indicate that the structural embeddings significantly improve the classification tasks when concatenated with the existing pre-trained word embeddings.

Decoding Word Embeddings with Brain-Based Semantic Features

Word embeddings are vectorial semantic representations built with either counting or predicting techniques aimed at capturing shades of meaning from word co-occurrences. Since their introduction, these representations have been criticized for lacking interpretable dimensions. This property of word embeddings limits our understanding of the semantic features they actually encode. Moreover, it contributes to the “black box” nature of the tasks in which they are used, since the reasons for word embedding performance often remain opaque to humans. In this contribution, we explore the semantic properties encoded in word embeddings by mapping them onto interpretable vectors, consisting of explicit and neurobiologically motivated semantic features (Binder et al. 2016). Our exploration takes into account different types of embeddings, including factorized count vectors and predict models (Skip-Gram, GloVe, etc.), as well as the most recent contextualized representations (i.e., ELMo and BERT).In our analysis, we first evaluate the quality of the mapping in a retrieval task, then we shed light on the semantic features that are better encoded in each embedding type. A large number of probing tasks is finally set to assess how the original and the mapped embeddings perform in discriminating semantic categories. For each probing task, we identify the most relevant semantic features and we show that there is a correlation between the embedding performance and how they encode those features. This study sets itself as a step forward in understanding which aspects of meaning are captured by vector spaces, by proposing a new and simple method to carve human-interpretable semantic representations from distributional vectors.

Deep Transformer modeling via grouping skip connection for neural machine translation

Most of the deep neural machine translation (NMT) models are based on a bottom-up feedforward fashion, in which representations in low layers construct or modulate high layers representations. We conjecture that this unidirectional encoding fashion could be a potential issue in building a deep NMT model. In this paper, we propose to build a deeper Transformer encoder by properly organizing encoder layers into multiple groups, which are connected via a grouping skip connection mechanism. Here, each group is further appropriately fed into subsequent groups to build a deep Transformer encoder. In this way, we successfully build a deep Transformer encoder with up to 48 layers. Moreover, we can share the parameters among groups to extend the encoder (virtual) depth even without introducing additional parameters. Detailed experimentation on the large-scale WMT (workshop on machine translation) 2014 English-to-German, English-to-French translation, WMT 2016 English-to-German, and WMT 2017 Chinese-to-English tasks demonstrates that our proposed deep Transformer model significantly outperforms the strong Transformer baseline. Furthermore, we carry out linguistic probing tasks to analyze the problems existing in the original Transformer model and explain how our deep Transformer encoder improves the translation quality. One particularly nice property of our approach is that it is incredibly easy to implement. We make our code available on Github https://github.com/liyc7711/deep-nmt.

From General Language Understanding to Noisy Text Comprehension

Obtaining meaning-rich representations of social media inputs, such as Tweets (unstructured and noisy text), from general-purpose pre-trained language models has become challenging, as these inputs typically deviate from mainstream English usage. The proposed research establishes effective methods for improving the comprehension of noisy texts. For this, we propose a new generic methodology to derive a diverse set of sentence vectors combining and extracting various linguistic characteristics from latent representations of multi-layer, pre-trained language models. Further, we clearly establish how BERT, a state-of-the-art pre-trained language model, comprehends the linguistic attributes of Tweets to identify appropriate sentence representations. Five new probing tasks are developed for Tweets, which can serve as benchmark probing tasks to study noisy text comprehension. Experiments are carried out for classification accuracy by deriving the sentence vectors from GloVe-based pre-trained models and Sentence-BERT, and by using different hidden layers from the BERT model. We show that the initial and middle layers of BERT have better capability for capturing the key linguistic characteristics of noisy texts than its latter layers. With complex predictive models, we further show that the sentence vector length has lesser importance to capture linguistic information, and the proposed sentence vectors for noisy texts perform better than the existing state-of-the-art sentence vectors.

Revisiting Negation in Neural Machine Translation

In this paper, we evaluate the translation of negation both automatically and manually, in English–German (EN–DE) and English– Chinese (EN–ZH). We show that the ability of neural machine translation (NMT) models to translate negation has improved with deeper and more advanced networks, although the performance varies between language pairs and translation directions. The accuracy of manual evaluation in EN→DE, DE→EN, EN→ZH, and ZH→EN is 95.7%, 94.8%, 93.4%, and 91.7%, respectively. In addition, we show that under-translation is the most significant error type in NMT, which contrasts with the more diverse error profile previously observed for statistical machine translation. To better understand the root of the under-translation of negation, we study the model’s information flow and training data. While our information flow analysis does not reveal any deficiencies that could be used to detect or fix the under-translation of negation, we find that negation is often rephrased during training, which could make it more difficult for the model to learn a reliable link between source and target negation. We finally conduct intrinsic analysis and extrinsic probing tasks on negation, showing that NMT models can distinguish negation and non-negation tokens very well and encode a lot of information about negation in hidden states but nevertheless leave room for improvement.

An Experimental Comparison of the Geometry of Models Trained on Natural Language and Synthetic Data

Deep learning models have been used successfully to solve many Natural Language Processing problems, but less is known about the mechanisms that make them work. Unlike simpler models that can be understood mathematically, experimental compar- isons and probing tasks are being used on deep neural net model attributes to try and determine how the syntactic and semantic information is being represented. In this work, we focus on models that produce contextual word vectors and sentence vectors. To determine how the information in these embeddings is being represented, we use a class of feed-forward neural networks to design a set of experiments that are run on the normal downstream tasks of two common Natural Language Inference benchmarks and one Sentiment Analysis benchmark, together with the classification tasks of a set of synthetic datasets that we have created. A quantitative analysis of these experiments shows that the behaviour of the models trained on the natural language benchmark data match those of the synthetic data in many cases. We suggest that this similar behaviour implies that the decomposition of semantic and syntactic data in the models trained on the language data are similar to the known structures of the synthetic data.

Sensorless Optimal Interaction Control Exploiting Environment Stiffness Estimation

Industrial robots are increasingly used to perform tasks that require an interaction with the surrounding environment (e.g., assembly tasks). Such environments are usually (partially) unknown to the robot (in terms of dynamic characteristics), demanding the implemented controllers to suitably react to the established interaction. Standard controllers require force/torque measurements to close the loop, making it, if possible, to adapt the robot behavior to the specific environment. However, most of the industrial manipulators do not have embedded force/torque sensor(s), which entails additional effort in terms of costs and implementation for their integration in the robotic setup. To extend the use of sensorless compliant controllers to force control, a robot–environment interaction dynamics model-based methodology is presented in this article. Relying on the sensorless Cartesian impedance control, an extended Kalman filter (EKF) is proposed to estimate the stiffness of an interaction environment. Exploiting the provided estimation, the robot–environment coupled dynamic modeling is used to design an optimal LQR interaction controller to close the force loop. The control gains can be analytically computed by solving the related Riccati equation, as such gains are a function of the impedance control and environment parameters. In addition, the interaction force can be estimated to close the force loop, making the sensorless robot able to perform the target interaction task. The described approach has been validated with experiments by analyzing two scenarios: a probing task and a closing of a plastic (i.e., compliant) box with a snap-fit closure mechanism. The performance of the proposed control framework has been evaluated, highlighting the capabilities of the EKF and the optimal LQR interaction controller. Finally, the proposed control schema is enhanced by the adaptation of the EKF for the estimation of the external wrench. Two additional experiments are provided to show the improvements on the control schema (a polishing-like task and an assembly task). The Franka EMIKA panda robot has been used as the reference robotic platform for experimental validation.

Sensorless environment stiffness and interaction force estimation for impedance control tuning in robotized interaction tasks

Industrial robots are increasingly used to perform tasks requiring an interaction with the surrounding environment (e.g., assembly tasks). Such environments are usually (partially) unknown to the robot, requiring the implemented controllers to suitably react to the established interaction. Standard controllers require force/torque measurements to close the loop. However, most of the industrial manipulators do not have embedded force/torque sensor(s) and such integration results in additional costs and implementation effort. To extend the use of compliant controllers to sensorless interaction control, a model-based methodology is presented in this paper. Relying on sensorless Cartesian impedance control, two Extended Kalman Filters (EKF) are proposed: an EKF for interaction force estimation and an EKF for environment stiffness estimation. Exploiting such estimations, a control architecture is proposed to implement a sensorless force loop (exploiting the provided estimated force) with adaptive Cartesian impedance control and coupling dynamics compensation (exploiting the provided estimated environment stiffness). The described approach has been validated in both simulations and experiments. A Franka EMIKA panda robot has been used. A probing task involving different materials (i.e., with different - unknown - stiffness properties) has been considered to show the capabilities of the developed EKFs (able to converge with limited errors) and control tuning (preserving stability). Additionally, a polishing-like task and an assembly task have been implemented to show the achieved performance of the proposed methodology.

Bi-manual Haptic-based Periodontal Simulation with Finger Support and Vibrotactile Feedback

The rise of virtual reality and haptic technologies has created exciting new applications in medical training and education. In a dental simulation, haptic technology can create the illusion of substances (teeth, gingiva, bone, etc.) by providing interaction forces within a simulated virtual world of the mouth. In this article, a haptic periodontal training simulation system, named Haptodont, is developed and evaluated for simulating periodontal probing. Thirty-two faculty members from New York University College of Dentistry were recruited and divided into three groups to evaluate three fundamental functionalities: Group 1 evaluated bi-manual 3 Degrees of Freedome (DoF) haptic interaction, Group 2 evaluated bi-manual 3 DoF haptic interaction with a finger support mechanism, and Group 3 evaluated bi-manual 3 DoF haptic interaction with finger support mechanism and vibrotactile feedback. The probe and mirror interactions were simulated with the Geomagic Touch haptic device whereas the finger support was implemented using the Novint Falcon device. The three groups conducted two probing tasks: healthy gingiva scenario with no pockets (2- to 3-mm depth) and periodontitis scenario with deep pockets (4- to 8-mm depth). Results demonstrated that experts performed comparably to clinical settings in terms of probing depth error (within 0.3 to 0.6 mm) and probing forces (less than 0.5 N). Furthermore, the finger support mechanism significantly improved the probing accuracy for periodontitis condition in the lingual region. The argument that probing the lingual region is more difficult than the buccal region is supported by quantitative evidence (significantly higher probing depth error and probing force). Further research is planned to improve the usability of the finger support, integrate the Haptodont system into the pre-clinical curriculum, and evaluate the Haptodont system with dental students as a learning tool.