Dominican Republic

Feed-forward layers constitute two-thirds of a transformer model&#39;s parameters, yet their role in the network remains under-explored. We show that feed-forward layers in transformer-based language models operate as key-value memories, where each key correlates with textual patterns in the training examples, and each value induces a distribution over the output vocabulary. Our experiments show that the learned patterns are human-interpretable, and that lower layers tend to capture shallow patterns, while upper layers learn more semantic ones. The values complement the keys&#39; input patterns by inducing output distributions that concentrate probability mass on tokens likely to appear immediately after each pattern, particularly in the upper layers. Finally, we demonstrate that the output of a feed-forward layer is a composition of its memories, which is subsequently refined throughout the model&#39;s layers via residual connections to produce the final output distribution.

EMNLP 2021

Transformer Feed-Forward Layers Are Key-Value Memories

EMNLP 2021 is planned to be a hybrid event in Punta Cana, Dominican Republic, with both on-site and fully virtual participation possible. The experience for on-site participants would closely approximate a normal pre-COVID *ACL conference, with 5-6 thematically organized parallel sessions and live Q/A and interactive discussion immediately after the talks. Presentations by virtual participants will be equitably interleaved with those of on-site participants, projected on the auditorium screens as if on-site, and also followed immediately by live Q/A and interactive discussion at a time during reasonable waking hours for the virtual presenter. For all participants, on-site and virtual, who are unable to attend a session due to either time-zone issues or because they are participating in another session live, talk recordings and slides will be available online at a minimum after the live presentation (and in many cases before as well), and questions may be submitted in advance on session-specific discussion boards and answered live in session with the usual visual aids if desired.

<iframe style="width:700px;height:400px" src="https://online.fliphtml5.com/ebtyf/ceby/"  seamless="seamless" scrolling="no" frameborder="0" allowtransparency="true" allowfullscreen="true" ></iframe>

Please Note: The EMNLP registration system is not currently connected to the underline site as we are still in the process of building out EMNLP 2021. You will receive access instructions from underline the week of November 1st. 

Access is given only to EMNLP upon registration, if you have not registered please do so [here](https://2021.emnlp.org/registration).

Registered attendees will receive access the week of November 1st.

EMNLP 2021 is planned to be a hybrid event in Punta Cana, Dominican Republic, with both on-site and fully virtual participation possible.

Feed-forward layers constitute two-thirds of a transformer model's parameters, yet their role in the network remains under-explored. We show that feed-forward layers in transformer-based language models operate as key-value memories, where each key correlates with textual patterns in the training examples, and each value induces a distribution over the output vocabulary. Our experiments show that the learned patterns are human-interpretable, and that lower layers tend to capture shallow patterns, while upper layers learn more semantic ones. The values complement the keys' input patterns by inducing output distributions that concentrate probability mass on tokens likely to appear immediately after each pattern, particularly in the upper layers. Finally, we demonstrate that the output of a feed-forward layer is a composition of its memories, which is subsequently refined throughout the model's layers via residual connections to produce the final output distribution.

poster

The goal of question answering (QA) is to answer any question. However, major QA datasets have skewed distributions over gender, profession, and nationality. Despite that skew, model accuracy analysis reveals little evidence that accuracy is lower for people based on gender or nationality; instead, there is more variation on professions (question topic). But QA's lack of representation could itself hide evidence of bias, necessitating QA datasets that better represent global diversity.

Toward Deconfounding the Influence of Subject's Demographic Characteristics in Question Answering

Commonsense reasoning benchmarks have been largely solved by fine-tuning language models. The downside is that fine-tuning may cause models to overfit to task-specific data and thereby forget their knowledge gained during pre-training. Recent works only propose lightweight model updates as models may already possess useful knowledge from past experience, but a challenge remains in understanding what parts and to what extent models should be refined for a given task. In this paper, we investigate what models learn from commonsense reasoning datasets. We measure the impact of three different adaptation methods on the generalization and accuracy of models. Our experiments with two models show that fine-tuning performs best, by learning both the content and the structure of the task, but suffers from overfitting and limited generalization to novel answers. We observe that alternative adaptation methods like prefix-tuning have comparable accuracy, but generalize better to unseen answers and are more robust to adversarial splits.

Exploring Strategies for Generalizable Commonsense Reasoning with Pre-trained Models

Pre-trained Transformer language models~(LM) have become go-to text representation encoders. Prior research fine-tunes deep LMs to encode text sequences such as sentences and passages into single dense vector representations for efficient text comparison and retrieval. However, dense encoders require a lot of data and sophisticated techniques to effectively train and suffer in low data situations. This paper finds a key reason is that standard LMs' internal attention structure is not ready-to-use for dense encoders, which needs to aggregate text information into the dense representation. We propose to pre-train towards dense encoder with a novel Transformer architecture, Condenser, where LM prediction CONditions on DENSE Representation. Our experiments show Condenser improves over standard LM by large margins on various text retrieval and similarity tasks.

Condenser: a Pre-training Architecture for Dense Retrieval

Slow emerging topic detection is a task between event detection, where we aggregate behaviors of different words on short period of time, and language evolution, where we monitor their long term evolution. In this work, we tackle the problem of early detection of slowly emerging new topics. To this end, we gather evidence of weak signals at the word level. We propose to monitor the behavior of words representation in an embedding space and use one of its geometrical properties to characterize the emergence of topics. As evaluation is typically hard for this kind of task, we present a framework for quantitative evaluation. We show positive results that outperform state-of-the-art methods on two public datasets of press and scientific articles.

Monitoring geometrical properties of word embeddings for detecting the emergence of new topics.

This paper describes a compact and effective model for low-latency passage retrieval in con- versational search based on learned dense rep- resentations. Prior to our work, the state-of- the-art approach uses a multi-stage pipeline comprising conversational query reformula- tion and information retrieval modules. De- spite its effectiveness, such a pipeline often includes multiple neural models that require long inference times. In addition, indepen- dently optimizing each module ignores de- pendencies among them. To address these shortcomings, we propose to integrate conver- sational query reformulation directly into a dense retrieval model. To aid in this goal, we create a dataset with pseudo-relevance labels for conversational search to overcome the lack of training data and to explore different train- ing strategies. We demonstrate that our model effectively rewrites conversational queries as dense representations in conversational search and open-domain question answering datasets. Finally, after observing that our model learns to adjust the L2 norm of query token embed- dings, we leverage this property for hybrid re- trieval and to support error analysis.

Contextualized Query Embeddings for Conversational Search

The semantic matching capabilities of neural information retrieval can ameliorate synonymy and polysemy problems of symbolic approaches. However, neural models' dense representations are more suitable for re-ranking, due to their inefficiency. Sparse representations, either in symbolic or latent form, are more efficient with an inverted index. Taking the merits of the sparse and dense representations, we propose an ultra-high dimensional (UHD) representation scheme equipped with directly controllable sparsity. UHD's large capacity and minimal noise and interference among the dimensions allow for binarized representations, which are highly efficient for storage and search. Also proposed is a bucketing method, where the embeddings from multiple layers of BERT are selected/merged to represent diverse linguistic aspects. We test our models with MS MARCO and TREC CAR, showing that our model outperforms other sparse models.

Ultra-High Dimensional Sparse Representations with Binarization for Efficient Text Retrieval

In interpretable NLP, we require faithful rationales that reflect the model’s decision-making process for an explained instance. While prior work focuses on extractive rationales (a subset of the input words), we investigate their less-studied counterpart: free-text natural language rationales. We demonstrate that pipelines, models for faithful rationalization on information-extraction style tasks, do not work as well on “reasoning” tasks requiring free-text rationales. We turn to models that jointly predict and rationalize, a class of widely used high-performance models for free-text rationalization. We investigate the extent to which the labels and rationales predicted by these models are associated, a necessary property of faithful explanation. Via two tests, robustness equivalence and feature importance agreement, we find that state-of-the-art T5-based joint models exhibit desirable properties for explaining commonsense question-answering and natural language inference, indicating their potential for producing faithful free-text rationales.

Measuring Association Between Labels and Free-Text Rationales

In the current study, we analyzed 15297 texts from 39 cancer survivors who posted or commented on Reddit in order to detect the language particularities of cancer survivors from online discourse. We performed a computational linguistic analysis (part-of-speech analysis, emoji detection, sentiment analysis) on submissions around the time of the cancer diagnosis and around the time of remission. We found several significant differences in the texts posted around the time of remission compared to those around the time of diagnosis. Though our results need to be backed up by a higher corpus of data, they do cue to the fact that cancer survivors, around the time of remission, focus more on others, are more active on social media, and do not see the glass as half empty as suggested by the valence of the emojis.

Natural language processing as a tool to identify the Reddit particularities of cancer survivors around the time of diagnosis and remission: A pilot study

Finding the lineage of a research topic is crucial for understanding the prior state of the art and advancing scientific displacement. The deluge of scholarly articles makes it difficult to locate the most relevant prior work and causes researchers to spend a considerable amount of time building up their literature list. Citations play a significant role in discovering relevant literature. However, not all citations are created equal. A majority of the citations that a paper receives are for providing contextual, and background information to the citing papers and are not central to the theme of those papers. However, some papers are pivotal to the citing paper and inspire or stem up the research in the citing paper. Hence the nature of citation the former receives from the later is significant. In this work in progress paper, we discuss our preliminary idea towards establishing a lineage for a given research via identifying significant citations. We hypothesize that such an automated system can facilitate relevant literature discovery and help identify knowledge flow for at least a certain category of papers. The distal goal of this work is to identify the real impact of research work or a facility beyond direct citation counts.

Identifying Significant Citations via Mining Paper Full-Text

The exponential growth in the scientific literature in recent years has to lead to a serious scholarly information overload on researchers. Automatically identifying the specific contribution made in a research paper would help quicker comprehension of the work, faster literature survey, comparison with the related works, etc. Here in this work, we propose a method to automatically extract the contribution statements from research articles. We design a model (ContriSci) multi-task deep neural network leveraging section identification and citance classification of scientific statements to predict whether a given scientific sentence specifies a contribution or not. Our long-term goal, we create a knowledge graph of scientific contributions for machine comprehension and more straightforward navigation of research contributions for a particular domain. Our proposed model achieves the best performance over earlier methods (a relative improvement of 8.08% in terms ofF1score) for contributing sentence identification over a dataset of NaturalLanguage Processing (NLP) papers.

Transformer Feed-Forward Layers Are Key-Value Memories

Downloads

Next from EMNLP 2021

Toward Deconfounding the Influence of Subject's Demographic Characteristics in Question Answering

Similar lecture

Mutual-Learning Improves End-to-End Speech Translation

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES