Analyses of transformer-based models have shown that they encode a variety of linguistic information from their textual input. While these analyses have shed a light on the relation between linguistic information on one side, and internal architecture and parameters on the other, a question remains unanswered: how is this linguistic information reflected in sentence embeddings? Using datasets consisting of sentences with known structure, we test to what degree information about chunks (in particular noun, verb or prepositional phrases), such as grammatical number, or semantic role, can be localized in sentence embeddings. Our results show that such information is not distributed over the entire sentence embedding, but rather it is encoded in specific regions. Understanding how the information from an input text is compressed into sentence embeddings helps understand current transformer models and help build future explainable neural models.

How do we evaluate Large Language Models (LLMs) and determine the aspects and limits of their intelligent behaviour? It is currently conjectured that shortcomings of LLMs in multi-linguality and reasoning are due to a lack of ability to generalize. It has been argued that, instead, humans are better at generalization because they have a tendency at extracting rules from complex data. We propose a method to evaluate LLMs ability to rule-based generalization. When exposed to tests of analytic intelligence, for example the visual RAVEN IQ test, human problem-solvers identify the relevant objects in the picture and their relevant attributes and reason based on rules applied to them. Based on the induced rules, they are able to provide a generalisation and a solution to the test. An analogous language task has recently been proposed (called BLM) for LLM. In this paper, we argue that we can use this task to investigate what linguistic reasoning LLM develop, by asking them to solve some simple variants of the BLM task. We find that current state-of-the-art generative models, such as ChatGPT, can handle the task in the sense that they easily understand the instructions and can provide step-by-step reasoning that shows that it can solve two of the main cognitive hurdles: correspondence finding (object and attribute identification) and item novelty. However, overall they cannot find the correct answer, even with considerable help. In particular, they never identify the structure of the problem, exhibiting, we hypothesize, a lack of goal and subgoal management abilities, an ability that has been argued to measure differential abilities in humans. We argue that this finding supports the usefulness of the task as a method to test the limits and specific properties of generalisation ability in Large Language Models, providing an intrinsic evaluation method inspired by tests of human intelligence.

Current NLP models appear to be achieving performance comparable to human capabilities on well-established benchmarks. New benchmarks are now necessary to test deeper layers of understanding of natural languages by these models. Blackbird’s Language Matrices are a recently developed framework that draws inspiration from tests of human analytic intelligence. The BLM task has revealed that successful performances in previously studied linguistic problems do not yet stem from a deep understanding of the generative factors that define these problems. In this study, we define a new BLM task for predicate-argument structure, and develop a structured dataset for its investigation, concentrating on the spray-load verb alternations in English, as a case study. The context sentences include one alternant from the spray-load alternation and the target sentence is the other alternant, to be chosen among a minimally contrastive and adversarial set of answers. We describe the generation process of the dataset and the reasoning behind the generating rules. The dataset aims to facilitate investigations into how verb information is encoded in sentence embeddings and how models generalize to the complex properties of argument structures. Benchmarking experiments conducted on the dataset and qualitative error analysis on the answer set reveal the inherent challenges associated with the problem even for current high-performing representations.

Successful machine learning systems currently rely on massive amounts of data, which are very effective in hiding some of the shallowness of the learned models. To help train models with more complex and compositional skills, we need challenging data, on which a system is successful only if it detects structure and regularities, that will allow it to generalize. In this paper, we describe a French dataset (BLM-AgrF) for learning the underlying rules of subject-verb agreement in sentences, developed in the BLM framework, a new task inspired by visual IQ tests known as Raven’s Progressive Matrices. In this task, an instance consists of sequences of sentences with specific attributes. To predict the correct answer as the next element of the sequence, a model must correctly detect the generative model used to produce the dataset. We provide details and share a dataset built following this methodology. Two exploratory baselines based on commonly used architectures show that despite the simplicity of the phenomenon, it is a complex problem for deep learning systems.

Sentence embeddings induced with various transformer architectures encode much semantic and syntactic information in a distributed manner in a one-dimensional array. We investigate whether specific grammatical information can be accessed in these distributed representations. Using data from a task developed to test rule-like generalizations, our experiments on detecting subject-verb agreement yield several promising results. First, we show that while the usual sentence representations encoded as one-dimensional arrays do not easily support extraction of rule-like regularities, a two-dimensional reshaping of these vectors allows various learning architectures to access such information. Next, we show that various architectures can detect patterns in these two-dimensional reshaped sentence embeddings and successfully learn a model based on smaller amounts of simpler training data, which performs well on more complex test data. This indicates that current sentence embeddings contain information that is regularly distributed, and which can be captured when the embeddings are reshaped into higher dimensional arrays. Our results cast light on representations produced by language models and help move towards developing few-shot learning approaches.

To develop a system with near-human language capabilities, we need to understand current systems’ generalisation and compositional abilities. We approach this by generating compositional, structured data, inspired from visual intelligence tests, that depend on the problem-solvers being able to disentangle objects and their absolute and relative properties in a sequence of images. We design an analogous task and develop the corresponding datasets that capture specific linguistic phenomena and their properties. Solving each problem instance depends on detecting the relevant linguistic objects and generative rules of the problem. We propose two datasets modelling two linguistic phenomena – subject-verb agreement in French, and verb alternations in English. The datasets can be used to investigate how LLMs encode linguistic objects, such as phrases, their grammatical and semantic properties, such as number or semantic role, and how such information is combined to correctly solve each problem. Specifically generated error types help investigate the behaviour of the system, which important information it is able to detect, and which structures mislead it.

Generative adversarial networks (GANs) have succeeded in inducing cross-lingual word embeddings - maps of matching words across languages - without supervision. Despite these successes, GANs’ performance for the difficult case of distant languages is still not satisfactory. These limitations have been explained by GANs’ incorrect assumption that source and target embedding spaces are related by a single linear mapping and are approximately isomorphic. We assume instead that, especially across distant languages, the mapping is only piece-wise linear, and propose a multi-adversarial learning method. This novel method induces the seed cross-lingual dictionary through multiple mappings, each induced to fit the mapping for one subspace. Our experiments on unsupervised bilingual lexicon induction and cross-lingual document classification show that this method improves performance over previous single-mapping methods, especially for distant languages.

What do people know when they know the meaning of words? Word associations have been widely used to tap into lexical repre- sentations and their structure, as a way of probing semantic knowledge in humans. We investigate whether current word embedding spaces (contextualized and uncontextualized) can be considered good models of human lexi- cal knowledge by studying whether they have comparable characteristics to human associa- tion spaces. We study the three properties of association rank, asymmetry of similarity and triangle inequality. We find that word embeddings are good mod- els of some word associations properties. They replicate well human associations between words, and, like humans, their context-aware variants show violations of the triangle in- equality. While they do show asymmetry of similarities, their asymmetries do not map those of human association norms.

To process the syntactic structures of a language in ways that are compatible with human expectations, we need computational representations of lexical and syntactic properties that form the basis of human knowledge of words and sentences. Recent neural-network-based and distributed semantics techniques have developed systems of considerable practical success and impressive performance. As has been advocated by many, however, such systems still lack human-like properties. In particular, linguistic, psycholinguistic and neuroscientific investigations have shown that human processing of sentences is sensitive to structure and unbounded relations. In the spirit of better understanding the structure building and long-distance properties of neural networks, I will present an overview of recent results on agreement and island effects in syntax in several languages. While certain sets of results in the literature indicate that neural language models exhibit long-distance agreement abilities, other finer-grained investigation of how these effects are calculated indicates that that the similarity spaces they define do not correlate with human experimental results on intervention similarity in long-distance dependencies. This opens the way to reflections on how to better match the syntactic properties of natural languages in the representations of neural models.

Research on the bilingual lexicon has uncovered fascinating interactions between the lexicons of the native language and of the second language in bilingual speakers. In particular, it has been found that the lexicon of the underlying native language affects the organisation of the second language. In the spirit of interpreting current distributed representations, this paper investigates two models of cross-lingual word embeddings, comparing them to the shared-translation effect and the cross-lingual coactivation effects of false and true friends (cognates) found in humans. We find that the similarity structure of the cross-lingual word embeddings space yields the same effects as the human bilingual lexicon.

Distributed representations of words which map each word to a continuous vector have proven useful in capturing important linguistic information not only in a single language but also across different languages. Current unsupervised adversarial approaches show that it is possible to build a mapping matrix that aligns two sets of monolingual word embeddings without high quality parallel data, such as a dictionary or a sentence-aligned corpus. However, without an additional step of refinement, the preliminary mapping learnt by these methods is unsatisfactory, leading to poor performance for typologically distant languages. In this paper, we propose a weakly-supervised adversarial training method to overcome this limitation, based on the intuition that mapping across languages is better done at the concept level than at the word level. We propose a concept-based adversarial training method which improves the performance of previous unsupervised adversarial methods for most languages, and especially for typologically distant language pairs.

The recent wide-spread and strong interest in RNNs has spurred detailed investigations of the distributed representations they generate and specifically if they exhibit properties similar to those characterising human languages. Results are at present inconclusive. In this paper, we extend previous work on long-distance dependencies in three ways. We manipulate word embeddings to translate them in a space that is attuned to the linguistic properties under study. We extend the work to sentence embeddings and to new languages. We confirm previous negative results: word embeddings and sentence embeddings do not unequivocally encode fine-grained linguistic properties of long-distance dependencies.

Despite their practical success and impressive performances, neural-network-based and distributed semantics techniques have often been criticized as they remain fundamentally opaque and difficult to interpret. In a vein similar to recent pieces of work investigating the linguistic abilities of these representations, we study another core, defining property of language: the property of long-distance dependencies. Human languages exhibit the ability to interpret discontinuous elements distant from each other in the string as if they were adjacent. This ability is blocked if a similar, but extraneous, element intervenes between the discontinuous components. We present results that show, under exhaustive and precise conditions, that one kind of word embeddings and the similarity spaces they define do not encode the properties of intervention similarity in long-distance dependencies, and that therefore they fail to represent this core linguistic notion.

This paper describes the University of Geneva’s submission to the CoNLL 2017 shared task Multilingual Parsing from Raw Text to Universal Dependencies (listed as the CLCL (Geneva) entry). Our submitted parsing system is the grandchild of the first transition-based neural network dependency parser, which was the University of Geneva’s entry in the CoNLL 2007 multilingual dependency parsing shared task, with some improvements to speed and portability. These results provide a baseline for investigating how far we have come in the past ten years of work on neural network dependency parsing.

The growing work in multi-lingual parsing faces the challenge of fair comparative evaluation and performance analysis across languages and their treebanks. The difficulty lies in teasing apart the properties of treebanks, such as their size or average sentence length, from those of the annotation scheme, and from the linguistic properties of languages. We propose a method to evaluate the effects of word order of a language on dependency parsing performance, while controlling for confounding treebank properties. The method uses artificially-generated treebanks that are minimal permutations of actual treebanks with respect to two word order properties: word order variation and dependency lengths. Based on these artificial data on twelve languages, we show that longer dependencies and higher word order variability degrade parsing performance. Our method also extends to minimal pairs of individual sentences, leading to a finer-grained understanding of parsing errors.

Traditional machine translation evaluation metrics such as BLEU and WER have been widely used, but these metrics have poor correlations with human judgements because they badly represent word similarity and impose strict identity matching. In this paper, we propose some modifications to the traditional measures based on word embeddings for these two metrics. The evaluation results show that our modifications significantly improve their correlation with human judgements.

In this paper we discuss the design and implementation of a parser for German and Arabic, which is currently being used in a tutoring system for foreign language training. Computer-aided language tutoring is a good application for testing the robustness and flexibility of a parsing system, since the input is usually ungrammatical in some way. Efficiency is also a concern, as tutoring applications typically run on personal computers, with the parser sharing memory with other components of the system. Our system is principle-based, which ensures a compact representation, and improves portability, needed in order to extend the initial design from German to Arabic and (eventually) Spanish. Currently, the parser diagnoses agreement errors, case errors, selection errors, and some word order errors. The parser can handle simple and complex declaratives and questions, topicalisations, verb movement, relative clauses — broad enough coverage to be useful in the design of real exercises and dialogues.