United Arab Emirates

Understanding Transformer-based models has attracted significant attention, as they lie at the heart of recent technological advances across machine learning.
While most interpretability methods rely on running models over inputs, recent work has shown that a zero-pass approach, where parameters are interpreted directly without a forward/backward pass is feasible for \emph{some} Transformer parameters, and for two-layer attention networks.
In this work, we present a theoretical analysis where \emph{all} parameters of a trained Transformer are interpreted by projecting them into the \emph{embedding space}, that is, the space of vocabulary items they operate on.
We derive a simple theoretical framework to support our arguments and provide ample evidence for its validity. First, an empirical analysis showing that parameters of both pretrained and fine-tuned models can be interpreted in embedding space. Second, we present two applications of our framework: (a) aligning the parameters of different models that share a vocabulary, and (b) constructing a classifier \emph{without training} by
``translating&#39;&#39; the parameters of a fine-tuned classifier to parameters of a different model that was only pretrained. Overall, our findings open the door 
to interpretation methods that, at least in part, abstract away from model specifics and operate in the embedding space only.

EMNLP 2022

Analyzing Transformers in Embedding Space

embedding space

explainability

interpretability

Understanding Transformer-based models has attracted significant attention, as they lie at the heart of recent technological advances across machine learning.
While most interpretability methods rely on running models over inputs, recent work has shown that a zero-pass approach, where parameters are interpreted directly without a forward/backward pass is feasible for \emph{some} Transformer parameters, and for two-layer attention networks.
In this work, we present a theoretical analysis where \emph{all} parameters of a trained Transformer are interpreted by projecting them into the \emph{embedding space}, that is, the space of vocabulary items they operate on.
We derive a simple theoretical framework to support our arguments and provide ample evidence for its validity. First, an empirical analysis showing that parameters of both pretrained and fine-tuned models can be interpreted in embedding space. Second, we present two applications of our framework: (a) aligning the parameters of different models that share a vocabulary, and (b) constructing a classifier \emph{without training} by
``translating'' the parameters of a fine-tuned classifier to parameters of a different model that was only pretrained. Overall, our findings open the door 
to interpretation methods that, at least in part, abstract away from model specifics and operate in the embedding space only.

workshop paper

## Welcome to EMNLP 2022!
I am delighted to welcome you to EMNLP 2022! I believe this conference has been complicated beyond any precedent. Over the past year, it’s been thrilling to see the organization team approach each new puzzle with creativity and enthusiasm. We hope that those participating in Abu Dhabi as well as those joining remotely will leave the conference feeling newly inspired by the program and newly connected to our ever-growing community. Following EMNLP 2021 and major NLP conferences since, EMNLP 2022 is “hybrid,” serving both virtual and in-person participants.

Our key innovations for EMNLP 2022 include:

* EMNLP 2022 is “hybrid” in a second sense, as well: we allowed both direct and rolling review paper submissions, building on the pilot experiment of EMNLP 2021, which considered a small number of ARR submissions. 
* Familiar from NAACL but new to EMNLP, we’ve added an industry track.
* During the conference, “portals” will link virtual poster sessions to in-person conference participants during poster sessions each day.
* The first ACL-family conference in the United Arab Emirates.

 *Message from Noah A. Smith, University of Washington and Allen Institute for AI, Seattle, Washington, USA* 
***EMNLP 2022 General Chair***
 
[![](https://assets.underline.io/uploads/markdown_image/1/image/9eec7d4a287ee18c278b08229290aa83.png)](https://drive.google.com/file/d/1OlPv6QBeo62VVTughj2jkiLeyHd1WnUt/view)
 
[![](https://assets.underline.io/uploads/markdown_image/1/image/a3db7a768409f05192210d98601edb25.png)](https://emnlp2022.rocket.chat/)

To access this site you need to register. Please register [here](https://2022.emnlp.org/registration/).

Register here

Welcome!
EMNLP 2022 will take place in Abu Dhabi from December 7th to December 11th, 2022. And it will be held in hybrid mode, both online and offline.

We study feature interactions in the context of feature attribution methods for post-hoc interpretability. 
Feature interactions are a mixed blessing: on the one hand, the ability to deal with interacting features is  seen as key to the success of neural networks in modern AI, allowing a model to build up hierarchical representations for its input. 
On the other hand, uncovering what exact role these interactions play in the decision making process of a model is challenging, leading to a diverse range of interaction attribution methods being proposed. 
In this paper, we introduce a unified framework that allows for direct comparison of these varied methods. Our framework is based on the `Explaining by Removing' framework of Covert et al. (2021) for feature attributions, and generalises it to feature interactions. To evaluate the faithfulness of methods for detecting feature interactions, we work out a grey box methodology, where we train deep learning models to perfection on a formal language classification task, using PCFGs. We show that under specific configurations, some interaction attribution methods are indeed able to uncover the grammatical rules acquired by a model. Our framework, and empirical results, thus provide a sound basis for comparing and extending methods for detecting feature interactions.

FIDAM-Eval: A Framework for Evaluating Feature Interaction Detection and Attribution Methods

The gold standard for model explanation should be analysis of the causal effects of model components on model behavior. We argue that causal abstraction provides a powerful toolkit for achieving such explanations.

Faithful, Interpretable Model Explanations via Causal Abstraction

There are numerous applications for the prediction of new links within knowledge graphs. The best models for this task, knowledge graph embedding (KGE) models, are typically evaluated by computing some average performance metric on a test set. This approach makes it difficult to identify where the models might systematically fail. To combat this issue, we propose a new evaluation framework that builds upon the idea of behavioral testing, allowing us to target critical system capabilities and evaluate the performance of KGE models on them.

Behavioral Testing of Knowledge Graph Embedding Models for Link Prediction

LSTMs trained on next word prediction can accurately perform linguistic tasks that require tracking long-distance syntactic dependencies. Notably, model accuracy approaches human performance on number agreement tasks (Gulordava et al., 2018). However, we do not have a mechanistic understanding of how LSTMs perform such linguistic tasks. Do LSTMs learn abstract grammatical rules, or do they rely on simple heuristics? Here, we test gender agreement in French which requires tracking both hierarchical syntactic structures and the inherent gender of lexical units. Our model is able to reliably predict long-distance gender agreement in two subject-predicate contexts: noun-adjective and noun-passive-verb agreement. The model showed more inaccuracies on plural noun phrases with gender attractors compared to singular cases, suggesting a reliance on clues from gendered articles for agreement. Overall, our study highlights key ways in which LSTMs deviate from human behaviour and questions whether LSTMs genuinely learn abstract syntactic rules and categories. We propose using gender agreement as a useful probe to investigate the underlying mechanisms, internal representations, and linguistic capabilities of LSTM language models.

Do LSTMs See Gender? Probing the Ability of LSTMs to Learn Abstract Syntactic Rules

Following how humans communicate, free-text rationales aim to use natural language to explain neural language model (LM) behavior. However, free-text rationales' unconstrained nature makes them prone to hallucination, so it is important to have metrics for free-text rationale quality. Existing free-text rationale metrics measure how consistent the rationale is with the LM's predicted label, but there is no protocol for assessing such metrics' reliability. Thus, we propose FRAME, a framework for evaluating rationale-label consistency (RLC) metrics for free-text rationales. FRAME is based on three axioms: (1) good metrics should yield highest scores for reference rationales, which maximize RLC by construction; (2) good metrics should be appropriately sensitive to semantic perturbation of rationales; and (3) good metrics should be robust to variation in the LM's task performance. Across three text classification datasets, we show that existing RLC metrics cannot satisfy all three FRAME axioms, since they are implemented via model pretraining which muddles the metric's signal. Then, we introduce a non-pretraining RLC metric that greatly outperforms baselines on (1) and (3), while performing competitively on (2). Finally, we discuss the limitations of using RLC to evaluate free-text rationales.

FRAME: Evaluating Rationale-Label Consistency Metrics for Free-Text Rationales

How language models process complex input that requires multiple steps of inference is not well understood. Previous research has shown that information about the intermediate results of their inputs can be extracted from the internal state of the models, but it is unclear where that information is encoded and whether that information is indeed used during inference. We analyze a BERT-based model by focusing on equations and their intermediate results. To identify where information about intermediate results is encoded, we measure the correlation between intermediate results and internal state using principal component analysis (PCA). Furthermore, by manipulating the features, we perform causal inference of the model. As a result, we show that the parameters in which the intermediate results information is encoded are not only correlated, but are indeed used in the prediction of the model. Our findings imply that there is locality in the model, and this is useful for increasing the interpretability of the model.

Tracing and Manipulating Intermediate Results in Neural Math Problem Solvers

Recently, pattern-based fine-tuning was introduced as an alternative approach for task adaption, i.e. adapting a pre-trained language model to a downstream tasks,. An interesting property of pattern-based fine-tuning that has received considerable attention recently, is its improved generalization over vanilla fine-tuning when applied in the few-shot setting. However, despite its popularity, pattern-based fine-tuning remains poorly understood. In this work, we show that the improved generalization of pattern-based fine-tuning is not consistent across models and datasets of the same task. Moreover, we show that previous works considerably underestimates the generalization of vanilla fine-tuning and provide a new baseline approach for fine-tuning. Lastly, we argue that it is critical to study the differences between vanilla and pattern-based fine-tuning more thoroughly, as well as to investigate when and why pattern-based fine-tuning works, in order to provide a better understanding of the generalization of fine-tuned language models in general.

A Critical Look at Fine-tuning Generalization: Are Patterns All We Need?

With the rise of AI systems in real-world applications comes the need for reliable and trustworthy AI\@. An essential aspect of this are explainable AI systems. However, there is no agreed standard on how such systems should be assessed. Inspired by the Turing test, we introduce a human-centric assessment framework where a leading domain expert accepts or rejects the solutions of an AI system and another domain expert. By comparing the acceptance rates of provided solutions, we can assess how the AI system performs compared to the domain expert, and whether the AI system's explanations are human-understandable. This setup---comparable to the Turing test---can serve as a framework for a wide range of human-centric AI system assessments. 

A Human-Centric Assessment Framework for AI

Recent success of pre-trained language models (PLMs) has stimulated interest in their ability to understand and work with numbers. Yet, the numerical reasoning over measurements has not been formally studied despite their importance. In this study, we show that PLMs lack the capability required for reasoning over measurements. Furthermore, we find that a language model trained on a measurement-rich corpus shows better performance on understanding measurements. We propose a simple embedding strategy to better distinguish between numbers and units, which leads to a significant improvement in the probing tasks.

Do Language Models Understand Measurements?

While Transformer-based language models are generally very robust to pruning, there is the recently discovered outlier phenomenon: disabling only 48 out of 110M parameters in BERT-base drops its performance by nearly 30% on MNLI. We replicate the original evidence for the outlier phenomenon and we link it to the geometry of the embedding space. We find that in both BERT and RoBERTa the magnitude of hidden state coefficients corresponding to outlier dimensions correlate with the frequencies of encoded tokens in pre-training data, and they also contribute to the "vertical" self-attention pattern enabling the model to focus on the special tokens. This explains the drop in performance from disabling the outliers, and it suggests that to decrease anisotopicity in future models we need pre-training schemas that would better take into account the skewed token distributions.

Analyzing Transformers in Embedding Space

Next from EMNLP 2022

FIDAM-Eval: A Framework for Evaluating Feature Interaction Detection and Attribution Methods

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES