Singapore

In this work-in-progress, we investigate the multi-step reasoning mechanisms of
transformer-based language models by adopting a reverse-engineering approach, inspired by work on mechanistic interpretability. To this end, we study a 6-layer transformer trained on a simplified tree traversal task. We find that the model forms backtracking heads which allow the model to backtrack a certain number of steps from the goal node. In addition, we find that the MLPs in the final layers are cru-
cial for finding paths when the goal is further away, implying that the model combines different internal mechanisms to solve even simple algorithmic tasks.

EMNLP 2023

Planning in Transformers: Evidence for Backtracking Circuits in Models Trained on Tree Traversal

mechanistic interpretability

reasoning

language models

workshop paper

**Welcome to EMNLP 2023!**

On behalf of the EMNLP 2023 Organizing Committee, I extend a warm and heartfelt welcome to all of you to the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). It is with immense pleasure and excitement that we gather here in Singapore, a vibrant hub of innovation and technological advancement.

The conference program is packed with insightful presentations, thought-provoking workshops, and engaging networking opportunities. In addition to the technical sessions, we have also planned several social events that will provide you with the opportunity to connect with your colleagues.

The conference is held in person at the Resorts World Convention Centre in Singapore, and available on-line with the help of Underline.

For those in Singapore, I hope that you will find time to explore the exciting Sentosa Island, Gardens by the Bay, Singapore Botanic Gardens and the many other attractions unique to Singapore.

Of course, we are grateful to our sponsors and partners for their generous support of EMNLP 2023, whose contributions make it possible for us to host this world-class event.

Yuji Matsumoto (RIKEN AIP) 
EMNLP 2023 General Chair

To access the **EMNLP 2023** event page on Underline, you need to register for the Conference. 
Please follow **[this link](https://2023.emnlp.org/registration/)** for more details.

EMNLP 2023 took place in Singapore from Dec 6th to Dec 10th, 2023.

Probing contextualized representations typically involves comparing task-specific model predictions against ground truth linguistic labels. Although this methodology shows what information can be recovered by a classifier, it does not reveal how a classifier uses the representation to make its decision. To address the latter problem, we ask: Do task-classifiers rely on representation- and task-independent geometric patterns in the embedding space? We explore this question by developing METAPROBE, an approach that uses geometric properties of representations to predict the behavior of task-specific classifiers (i.e., their predictions as opposed to the ground truth). Our experiments reveal the existence of universal geometric patterns across representations that can predict classifier predictions. Consequently, this allows us to posit a geometric explanation for the impressive performance of contextualized representations.

METAPROBE: A Representation- and Task-Agnostic Probe

Measuring Moral Inconsistencies of Large Language Models

Contrast set consistency is a robustness measurement that evaluates the rate at which a model correctly responds to all instances in a bundle of minimally different examples relying on the same knowledge. To draw additional insights, we propose to complement consistency with relative consistency -- the probability that an equally accurate model would surpass the consistency of the proposed model, given a distribution over possible consistencies. Models with 100% relative consistency have reached a consistency peak for their accuracy. We reflect on prior work that reports consistency in contrast sets and observe that relative consistency can alter the assessment of a model's consistency compared to another. We anticipate that our proposed measurement and insights will influence future studies aiming to promote consistent behavior in models.

How Much Consistency Is Your Accuracy Worth?

Humans exhibit strong preferences in syntactic alternations based on availability effects, where more 'available' constituents are ordered earlier in a sentence. Here, I show that autoregressive LMs learn these preferences both from training data and from architectural considerations on memory. I demonstrate that constraining context window size in both LSTM RNNs and n-gram models induces a preference towards producing short constituents before long ones. In particular, this has implications for incremental models of human language production.

The architecture gave the language model the constituent length preferences

We examine the licensing of negative polarity items (NPIs) in large language models (LLMs) to enrich the picture of how models acquire NPIs as linguistic phenomena at the syntax-semantics interface. NPIs are a class of words which have a restricted distribution, appearing only in certain licensing contexts, prototypically negation. Unlike much of previous work which assumes NPIs and their licensing environments constitute unified classes, we consider NPI distribution in its full complexity: different NPIs are possible in different licensing environments. By studying this phenomenon across a broad range of models, we are able to explore which features of the model architecture, properties of the training data, and linguistic characteristics of the NPI phenomenon itself drive performance.

NPIs Aren't Exactly Easy: Variation in Licensing across Large Language Models

Pretrained language models (PLMs) form the basis of most state-of-the-art NLP technologies. Nevertheless, they are essentially black boxes: Humans do not have a clear understanding of what knowledge is encoded in different parts of the models, especially in individual neurons. The situation is different in computer vision, where feature visualization provides a decompositional interpretability technique for
neurons of vision models. Activation maximization is used to synthesize inherently interpretable visual representations of the information encoded in individual neurons.
Our work is inspired by this but presents a cautionary tale on the interpretability of single neurons, based on the first large-scale attempt to adapt activation maximization to NLP, and, more specifically, large PLMs. We propose feature textualization, a technique to produce dense representations of neurons in the PLM
word embedding space. We apply feature textualization to the BERT model (Devlin et al., 2019) to investigate whether the knowledge encoded in individual neurons can be interpreted and symbolized. We find that the produced representations can provide insights about the knowledge encoded in individual neurons, but that individual neurons do not represent clearcut symbolic units of language such as words. Additionally, we use feature textualization to investigate how many neurons are needed to encode words in BERT.

Investigating the Encoding of Words in BERT's Neurons using Feature Textualization

Despite the fact that Transformers perform well in NLP tasks, recent studies suggest that self-attention is theoretically limited in learning even some regular and context-free languages. These findings motivated us to think about their implications in modeling natural language, which is hypothesized to be mildly context-sensitive. We test the Transformer's ability to learn mildly context-sensitive languages of varying complexities, and find that they generalize well to unseen in-distribution data, but their ability to extrapolate to longer strings is worse than that of LSTMs. Our analyses show that the learned self-attention patterns and representations modeled dependency relations and demonstrated counting behavior, which may have helped the models solve the languages.

Evaluating Transformer's Ability to Learn Mildly Context-Sensitive Languages

Large language models (LLMs) like GPT and
PALM have excelled in numerous natural language
processing (NLP) tasks such as text generation,
question answering, and translation.
However, they are also found to have inherent
social biases. To address this, recent studies
have proposed debiasing techniques like iterative
nullspace projection (INLP) and Counterfactual
Data Augmentation (CDA). Additionally,
there’s growing interest in understanding
the intricacies of these models. Some
researchers focus on individual neural units,
while others examine specific layers. In our
study, we benchmark newly released models,
assess the impact of debiasing methods, and
investigate how biases are linked to different
transformer layers using a method called Logit
Lens. Specifically, we evaluate three modern
LLMs: OPT, LLaMA, and LLaMA2, and their
debiased versions. Our experiments are based
on two popular bias evaluation datasets, StereoSet
and CrowS-Pairs, and we perform a layerby-
layer analysis using the Logit Lens.

Layered Bias: Interpreting Bias in Pretrained Large Language Models

Constituents are groups of words that behave as a syntactic unit. Many linguistic phenomena (e.g., question formation, diathesis alternations) require the manipulation and rearrangement of constituents in a sentence.
In this paper, we investigate how different finetuning setups affect the ability of pretrained sequence-to-sequence language models such as BART and T5 to replicate constituency tests --- transformations that involve manipulating constituents in a sentence. 
We design multiple evaluation settings by varying the combinations of constituency tests and sentence types that a model is exposed to during finetuning.
We show that models can replicate a linguistic transformation on a specific type of sentence that they saw during finetuning, but performance degrades substantially in other settings, showing a lack of systematic generalization. These results suggest that models often learn to manipulate sentences at a surface level unrelated to the constituent-level syntactic structure, for example by copying the first word of a sentence. These results may partially explain the brittleness of pretrained language models in downstream tasks.

Systematic Generalization by Finetuning? Analyzing Pretrained Language Models Using Constituency Tests

Language models (LMs) exhibit and amplify many types of undesirable biases learned from the training data, including gender bias. However, we lack tools for effectively and efficiently changing this behavior without hurting general language modeling performance. In this paper, we study three methods for identifying causal relations between LM components and particular output: causal mediation analysis, automated circuit discovery and our novel, efficient method called DiffMask+ based on differential masking. We apply the methods to GPT-2 small and the problem of gender bias, and use the discovered sets of components to perform parameter-efficient fine-tuning for bias mitigation. However, our work also underscores the difficulty of defining and measuring bias, and the sensitivity of causal discovery procedures to dataset choice. We hope our work can contribute to more attention for dataset development, and lead to more effective mitigation strategies for other types of bias.


Next from EMNLP 2023

METAPROBE: A Representation- and Task-Agnostic Probe