Can RNNs learn Recursive Nested Subject-Verb Agreements?

Abstract

One of the fundamental principles of contemporary linguistics states that language processing requires the ability to extract recursively nested tree structures. However, it remains unclear whether and how this code could be implemented in neural circuits. Recent advances in Recurrent Neural Networks (RNNs), which achieve near-human performance in some language tasks, provide a compelling model to address such questions. Here, we present a new framework to study recursive processing in RNNs, using subject-verb agreement as a probe into the representations of the neural network. We trained six distinct types of RNNs on a simplified probabilistic context-free grammar designed to independently manipulate the length of a sentence and the depth of its syntactic tree. All RNNs generalized to subject-verb dependencies longer than those seen during training. However, none systematically generalized to deeper tree structures, even those with a structural bias towards learning nested tree (i.e., stack-RNNs). In addition, our analyses revealed primacy and recency effects in the generalization patterns of LSTM-based models, showing that these models tend to perform well on the outer-and innermost parts of a center-embedded tree structure, but poorly on its middle levels. Finally, probing the internal states of the model during the processing of sentences with nested tree structures, we found a complex encoding of grammatical agreement information (e.g. grammatical number), in which all the information for multiple words nouns was carried by a single unit. Taken together, these results indicate how neural networks may extract bounded nested tree structures, without learning a systematic recursive rule.

Deep language algorithms predict semantic comprehension from brain activity

“…However, what these features actually represent remains largely unknown. Previous studies have shown that language transformers explicitly represent syntactic 14,52 and semantic features 14 . Similarly, Manning et al showed that syntactic trees appear to be encoded by the distances between contextualized word embedding 52 .…”

“…This limit is expected: several studies demonstrate that current deep language models fail to capture several aspects critical to comprehension 16,19 : they (i) often fail to generalize beyond the training distribution 56 , (ii) do not perfectly capture deep syntactic structures 14,52 and (iii) remain relatively poor at summarizing texts, generating stories and answering questions [20][21][22] . Furthermore, GPT-2 is only trained with textual data and does not situate objects in a grounded environment that would capture their real-world interactions 18,57 .…”

“…Recent studies suggest that they partially do: the hidden representations of various deep neural networks have shown to linearly predict single-sample fMRI [4][5][6][7][8][9][10][11] MEG 5,7 , and intracranial responses to spoken and written texts 6,12 .However, whether these models encode, retrieve and pay attention to information that specifically relates to behavior in general, and to comprehension in particular remains controversial [13][14][15][16][17][18][19] . This issue is all-the-more relevant that the behavior of deep language models remains challenged by complex questions, including subjectverb agreement 14,15,17 , causal reasoning 16,19 , story generation, text summarization as well as dialogue and question answering [20][21][22][23][24] .To explore the relationship between comprehension and the representations of GPT-2, we compare GPT-2's activations to the functional Magnetic Resonance Imaging of 101 subjects listening to 70min of seven short stories. We first quantify this similarity with a "brain score" (M) 25,26 .…”

“…However, whether these models encode, retrieve and pay attention to information that specifically relates to behavior in general, and to comprehension in particular remains controversial [13][14][15][16][17][18][19] . This issue is all-the-more relevant that the behavior of deep language models remains challenged by complex questions, including subjectverb agreement 14,15,17 , causal reasoning 16,19 , story generation, text summarization as well as dialogue and question answering [20][21][22][23][24] .…”

Long-range and hierarchical language predictions in brains and algorithms

“…Yet, a major gap remains between humans and these algorithms: current language models are still poor at story generation and summarization as well as dialogue and question answering (10)(11)(12)(13)(14); they fail to capture many syntactic constructs and semantics properties (15)(16)(17)(18)(19), and their linguistic understanding is often superficial (16,(18)(19)(20).…”