Evaluating Compositionality in Sentence Embeddings

By Ishita Dasgupta, Demi Guo, Andreas Stuhlmüller, Samuel J. Gershman, Noah D. Goodman. 2018

129 citations recorded by the source corpus.

An important frontier in the quest for human-like AI is compositional semantics: how do we design systems that understand an infinite number of expressions built from a finite vocabulary? Recent research has attempted to solve this problem by using deep neural networks to learn vector space embeddings of sentences, which then serve as input to supervised learning problems like paraphrase detection and sentiment analysis. Here we focus on 'natural language inference' (NLI) as a critical test of a system's capacity for semantic compositionality. In the NLI task, sentence pairs are assigned one of three categories: entailment, contradiction, or neutral. We present a new set of NLI sentence pairs that cannot be solved using only word-level knowledge and instead require some degree of compositionality. We use state of the art sentence embeddings trained on NLI (InferSent, Conneau et al. (2017)), and find that performance on our new dataset is poor, indicating that the representations learned by this model fail to capture the needed compositionality. We analyze some of the decision rules learned by InferSent and find that they are largely driven by simple heuristics at the word level that are ecologically valid in the SNLI dataset on which InferSent is trained. Further, we find that augmenting the training dataset with our new dataset improves performance on a held-out test set without loss of performance on the SNLI test set. This highlights the importance of structured datasets in better understanding, as well as improving the performance of, AI systems.

Subfields: Artificial Intelligence, General Social Sciences, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In General Social Sciences

In Representation & Self-Supervised Learning

Citations in the corpus

Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference (2019, methodology): “Similar in spirit to our lexical overlap heuristic, Dasgupta et al. (2018), Nie et al. (2018), and Kim et al. (2018) also studied the performance of NLI models on specific phenomena where word order matters; however, we use a larger set of phenomena with the aim of studying lexical overlap in…”
Annotation Artifacts in Natural Language Inference Data (2018, methodology): “In parallel to this work, Dasgupta et al. (2018) found that the InferSent model (Conneau et al., 2017) relies on word-level heuristics to achieve state-of-the-art performance on SNLI.”
e-SNLI: Natural Language Inference with Natural Language Explanations (2018, background): “For example, Dasgupta et al. [9] assembled a dataset to test whether inference models actually capture compositionality beyond word level.”
Tailor: Generating and Perturbing Text with Semantic Controls (2021, neutral): “heuristic on which NLI models have been shown to rely (Dasgupta et al., 2018; Naik et al., 2018).”
Analysis Methods in Neural Language Processing: A Survey (2018, background): “(Dasgupta et al., 2018) NLI Compositionality English 44010 Automatic (Sanchez et al.”
What do you learn from context? Probing for sentence structure in contextualized word representations (2019, background): “Often, probing tasks take the form of “challenge sets”, or test sets which are generated using templates and/or perturbations of existing test sets in order to isolate particular linguistic phenomena, e.g. compositional reasoning (Dasgupta et al., 2018; Ettinger et al., 2018).”
What BERT Is Not: Lessons from a New Suite of Psycholinguistic Diagnostics for Language Models (2019, neutral): “, 2016), and additional work has been done to design curated versions of these tasks to test for specific linguistic capabilities (Dasgupta et al., 2018; Poliak et al., 2018; McCoy et al., 2019).”
Negated and Misprimed Probes for Pretrained Language Models: Birds Can Talk, But Cannot Fly (2019, background): “A wide range of literature analyzes linguistic knowledge stored in pretrained embeddings (Jumelet and Hupkes, 2018; Gulordava et al., 2018; Giulianelli et al., 2018; McCoy et al., 2019; Dasgupta et al., 2018; Marvin and Linzen, 2018; Warstadt and Bowman, 2019; Kann et al., 2019).”
Adversarial Filters of Dataset Biases (2020, neutral): “Dasgupta et al. (2018) introduce an NLI dataset that cannot be solved using only word-level knowledge and requires some compositionality.”
Investigating BERT’s Knowledge of Language: Five Analysis Methods with NPIs (2019, methodology): “Recent work evaluating grammatical knowledge of sentence encoders like BERT (Devlin et al., 2018) has employed a variety of methods. For example, Shi et al. (2016), Ettinger et al.”
Inherent Disagreements in Human Textual Inferences (2019, methodology): “…be useful in applications, and currently, NLI is increasingly used as a means for ‘‘probing’’ neural models to assess their knowledge of arbitrary linguistic phenomena (Dasgupta et al., 2018; Ettinger et al., 2018; Poliak et al., 2018b; White et al., 2017; Poliak et al., 2018a; McCoy et al., 2019).”
Mining Discourse Markers for Unsupervised Sentence Representation Learning (2019, methodology): “Inspired by recent work (Dasgupta et al., 2018; Poliak et al., 2018; Levy et al., 2018; Glockner et al., 2018) on the unexpected properties of recent manually labelled datasets (e.g. SNLI), we will also analyze our dataset to check whether labels are easy to guess, and whether the proposed model…”
An Empirical Study on Robustness to Spurious Correlations using Pre-trained Language Models (2020, background): “For example, in natu-ral language inference (NLI) tasks, previous work has found that models learned on notable benchmarks achieve high accuracy by associating high word overlap between the premise and the hypothesis with entailment (Dasgupta et al., 2018; McCoy et al., 2019).”
Learning to Learn Words from Visual Scenes (2019, background): “Compositions have also been studied in natural language processing [9,12].”
Compositional generalization through meta sequence-to-sequence learning (2019, background): “Recent work has revisited these classic critiques through studies of modern neural architectures [10, 16, 3, 21, 23, 2, 6], with a focus on the sequence-to-sequence (seq2seq) models used successfully in machine translation and other NLP tasks [34, 4, 38].”
OCNLI: Original Chinese Natural Language Inference (2020, background): “There has also been a large body of work using probing datasets/tasks to stress-test NLI models trained on datasets such as SNLI and MNLI, in order to expose the weaknesses and biases in either the models or the data (Dasgupta et al., 2018; Naik et al., 2018; McCoy et al., 2019).”
Analyzing Compositionality-Sensitivity of NLI Models (2018, methodology): “Evans et al. (2018) introduce a dataset of logical entailments for measuring models’ ability to capture the structure of logical expressions against an entailment prediction task.”
RNNs Implicitly Implement Tensor Product Representations (2018, unknown): “One approach is to infer the information encoded in the representations from the system’s behavior on examples targeting specific representational components, such as semantics (Pavlick, 2017; Dasgupta et al., 2018; Poliak et al., 2018) or syntax (Linzen et al.”