Rethinking the Construction of Effective Metrics for Understanding the Mechanisms of Pretrained Language Models
Extracted source text ยท arXiv 2310.12454
% This must be in the first 5 lines to tell arXiv to use pdfLaTeX, which is strongly recommended.
\pdfoutput=1
% In particular, the hyperref package requires pdfLaTeX in order to break URLs across lines.
\documentclass[11pt]{article}
% Remove the "review" option to generate the final version.
\usepackage[]{EMNLP2023}
% Standard package includes
\usepackage{times}
\usepackage{latexsym}
\usepackage{graphicx}
\usepackage{amsthm}
\usepackage{subcaption}
\usepackage{adjustbox}
\usepackage{caption}
\usepackage{bm}
\usepackage{enumitem}
\newcommand*{\dif}{\mathop{}\!d}
% For proper rendering and hyphenation of words containing Latin characters (including in bib files)
\usepackage[T1]{fontenc}
% For Vietnamese characters
% \usepackage[T5]{fontenc}
% See https://www.latex-project.org/help/documentation/encguide.pdf for other character sets
% This assumes your files are encoded as UTF8
\usepackage[utf8]{inputenc}
% This is not strictly necessary and may be commented out.
% However, it will improve the layout of the manuscript,
% and will typically save some space.
\usepackage{microtype}
% This is also not strictly necessary and may be commented out.
% However, it will improve the aesthetics of text in
% the typewriter font.
\usepackage{inconsolata}
% If the title and author information does not fit in the area allocated, uncomment the following
%
%\setlength\titlebox{<dim>}
%
% and set <dim> to something 5cm or larger.
\title{Rethinking the Construction of Effective Metrics for Understanding the Mechanisms of Pretrained Language Models}
% Author information can be set in various styles:
% For several authors from the same institution:
% \author{Author 1 \and ... \and Author n \\
% Address line \\ ... \\ Address line}
% if the names do not fit well on one line use
% Author 1 \\ {\bf Author 2} \\ ... \\ {\bf Author n} \\
% For authors from different institutions:
% \author{Author 1 \\ Address line \\ ... \\ Address line
% \And ... \And
% Author n \\ Address line \\ ... \\ Address line}
% To start a separate ``row'' of authors use \AND, as in
% \author{Author 1 \\ Address line \\ ... \\ Address line
% \AND
% Author 2 \\ Address line \\ ... \\ Address line \And
% Author 3 \\ Address line \\ ... \\ Address line}
\newcommand*\samethanks[1][\value{footnote}]{\footnotemark[#1]}
\newcommand \footnoteONLYtext[1]
{
\let \mybackup \thefootnote
\let \thefootnote \relax
\footnotetext{#1}
\let \thefootnote \mybackup
\let \mybackup \imareallyundefinedcommand
}
\author{You Li\samethanks[1] \quad Jinhui Yin\samethanks[1] \quad Yuming Lin\samethanks[2] \\
Guangxi Key Laboratory of Trusted Software, Guilin University of Electronic Technology \\
\texttt{liyou@guet.edu.cn} \quad \texttt{yinjinhui55@gmail.com} \quad \texttt{ymlin@guet.edu.cn}\\
}
\begin{document}
\maketitle
\footnoteONLYtext{*Equal contribution.}
\footnoteONLYtext{\dag Corresponding Author.}
\begin{abstract}
Pretrained language models are expected to effectively map input text to a set of vectors while preserving the inherent relationships within the text. Consequently, designing a white-box model to compute metrics that reflect the presence of specific internal relations in these vectors has become a common approach for post-hoc interpretability analysis of pretrained language models. However, achieving interpretability in white-box models and ensuring the rigor of metric computation becomes challenging when the source model lacks inherent interpretability. Therefore, in this paper, we discuss striking a balance in this trade-off and propose a novel line to constructing metrics for understanding the mechanisms of pretrained language models. We have specifically designed a family of metrics along this line of investigation, and the model used to compute these metrics is referred to as the tree topological probe. We conducted measurements on BERT-large by using these metrics. Based on the experimental results, we propose a speculation regarding the working mechanism of BERT-like pretrained language models, as well as a strategy for enhancing fine-tuning performance by leveraging the topological probe to improve specific submodules.\footnote{Our code is available at \url{https://github.com/cclx/Effective_Metrics}}
\end{abstract}
\section{Introduction}
\label{sec:introduction}
Pretrained language models consisting of stacked transformer blocks \citep{DBLP:conf/nips/VaswaniSPUJGKP17} are commonly expected to map input text to a set of vectors, such that any relationship in the text corresponds to some algebraic operation on these vectors. However, it is generally unknown whether such operations exist. Therefore, designing a white-box model that computes a metric for a given set of vectors corresponding to a text, which reflects to some extent the existence of operations extracting specific information from the vectors, is a common approach for post-hoc interpretability analysis of such models \citep{DBLP:conf/acl/MaudslayVPWC20,DBLP:conf/acl/LimisiewiczM20,DBLP:conf/iclr/ChenFXXTCJ21,DBLP:conf/naacl/WhitePSC21, DBLP:conf/acl/ImmerHFC22}. However, even though we may desire strong interpretability from a white-box model and metrics computed by it that rigorously reflect the ability to extract specific information from a given set of vectors, it can be challenging to achieve both of these aspects simultaneously when the source model lacks inherent interpretability. Therefore, making implicit assumptions during metric computation is common \citep{DBLP:conf/icml/Kornblith0LH19, DBLP:conf/acl/WangK022}. A simple example is the cosine similarity of contextual embeddings. This metric is straightforward and has an intuitive geometric interpretation, making it easy to explain, but it tends to underestimate the similarity of high-frequency words \citep{DBLP:conf/acl/ZhouECJ22}.
On the other hand, due to the intuition that 'if a white-box model cannot distinguish embeddings that exhibit practical differences (such as context embeddings and static embeddings), it should be considered ineffective,' experimental validation of a white-box model's ability to effectively distinguish between embeddings with evident practical distinctions is a common practice in research. Furthermore, if the magnitude of metrics computed by a white-box model strongly correlates with the quality of different embeddings in practical settings, researchers usually trust its effectiveness. Therefore, in practice, traditional white-box models actually classify sets of vectors from different sources.
Taking the structural probe proposed by \citeauthor{DBLP:conf/naacl/HewittM19} as an example, they perform a linear transformation on the embedding of each complete word in the text and use the square of the L2 norm of the transformed vector as a prediction for the depth of the corresponding word in the dependency tree \citep{DBLP:conf/naacl/HewittM19}. In this way, the linear transformation matrix serves as a learning parameter, and the minimum risk loss between the predicted and true depths is used as a metric. Intuitively, the smaller the metric is, the more likely the embedding contains complete syntax relations. The experimental results indeed align with this intuition, showing that contextual embeddings (such as those generated by BERT \citep{DBLP:conf/naacl/DevlinCLT19}) outperform static embeddings. However, due to the unknown nature of the true deep distribution, it is challenging to deduce which geometric features within the representations influence the magnitude of structural probe measurements from the setup of structural probe. Overall, while the results of the structural probe provide an intuition that contextual embeddings, such as those generated by BERT, capture richer syntactic relations than those of the traditional embeddings, it is currently impossible to know what the geometric structure of a "good" embedding is for the metric defined by the structural probe.
In addition, to enhance the interpretability and flexibility of white-box models, it is common to include assumptions that are challenging to empirically validate. For example, \citeauthor{DBLP:conf/emnlp/Ethayarajh19} proposed to use anisotropy-adjusted self-similarity to measure the context-specificity of embeddings \citep{DBLP:conf/emnlp/Ethayarajh19}. Since the computation of this metric doesn't require the introduction of additional human labels, it is theoretically possible to conduct further analysis, such as examining how fundamental geometric features in the representation (e.g., rank) affect anisotropy-adjusted self-similarity, or simply consider this metric as defining a new geometric feature. Overall, this is a metric that can be discussed purely at the mathematical level. However, verifying whether the measured context-specificity in this metric aligns well with context-specificity in linguistics, without the use of, or with only limited additional human labels, may be challenging. Additionally, confirming whether the model leverages the properties of anisotropy-adjusted self-similarity during actual inference tasks might also be challenging.
There appears to be a trade-off here between two types of metrics:
1. Metrics that are constrained by supervised signals with ground truth labels, which provide more practical intuition.
2. Metrics that reflect the geometric properties of the vector set itself, which provide a more formal representation.
Therefore, we propose a new line that takes traditional supervised probes as the structure of the white-box model and then self-supervises it, trying to preserve both of the abovementioned properties as much as possible. The motivation behind this idea is that any feature that is beneficial for interpretability has internal constraints. If a certain feature has no internal constraints, it must be represented by a vector set without geometric constraints, which does not contain any interpretable factors. Therefore, what is important for interpretability is the correspondence between the internal constraints of the probed features and the vector set, which can describe the geometric structure of the vector set to some extent. \textbf{In the case where the internal constraints of the probed features are well defined, a probe that detects these features can naturally induce a probe that detects the internal constraints, which is self-supervised}.
In summary, the contributions of this work include:
\begin{enumerate}
\item We propose a novel self-supervised probe, referred to as the \textbf{tree topological probe}, to probe the hierarchical structure of sentence representations learned by pretrained language models like BERT.
\item We discuss the theoretical relationship between the tree topological probe and the structural probe, with the former bounding the latter.
\item We measure the metrics constructed based on the tree topological probe on BERT-large. According to the experimental results, we propose a speculation regarding the working mechanism of a BERT-like pretrained language model.
\item We utilize metrics constructed by the tree topological probe to enhance BERT's submodules during fine-tuning and observe that enhancing certain modules can improve the fine-tuning performance. We also propose a strategy for selecting submodules.
\end{enumerate}
\section{Related Work}
The probe is the most common approach for associating neural network representations with linguistic properties \citep{DBLP:conf/emnlp/VoitaT20}. This approach is widely used to explore part of speech knowledge \citep{DBLP:journals/tacl/BelinkovG19,DBLP:conf/emnlp/VoitaT20,DBLP:conf/acl/PimentelVMZWC20,DBLP:conf/emnlp/HewittELM21} and for sentence and dependency structures \citep{DBLP:conf/naacl/HewittM19,DBLP:conf/acl/MaudslayVPWC20,DBLP:conf/naacl/WhitePSC21,DBLP:conf/acl/LimisiewiczM20,DBLP:conf/iclr/ChenFXXTCJ21}. These studies demonstrate many important aspects of the linguistic information are encoded in pretrained representations. However, in some probe experiments, researchers have found that the probe precision obtained by both random representation and pretrained representation were quite close \citep{DBLP:conf/emnlp/ZhangB18,DBLP:conf/emnlp/HewittL19}. This demonstrates that it is not sufficient to use the probe precision to measure whether the representations contain specific language information. To improve the reliability of probes, some researchers have proposed the use of control tasks in probe experiments \citep{DBLP:conf/emnlp/HewittL19}. In recent research, \citeauthor{DBLP:conf/iclr/LoveringJLP21} realized that inductive bias can be used to describe the ease of extracting relevant features from representations. \citeauthor{DBLP:conf/acl/ImmerHFC22} further proposed a Bayesian framework for quantifying inductive bias with probes, and they used the Model Evidence Maximum instead of trivial precision.
\section{Methodology}
As the foundation of the white-box model proposed in this paper is built upon the traditional probe, we will begin by providing a general description of the probe based on the definition presented in \cite{DBLP:journals/corr/abs-2104-08197}. Additionally, we will introduce some relevant notation for better understanding.
\subsection{General Form of the Probe}
Given a character set, in a formal language, the generation rules uniquely determine the properties of the language. We assume that there also exists a set of generation rules $\mathcal{R}$ implicitly in natural language, and the language objects derived from these rules exhibit a series of features. Among these features, a subset $Y$ is selected as the probed feature for which the properties represent the logical constraints of the generation rule set. Assuming there is another model $\mathcal{M}$ that can assign a suitable representation vector to the generated language objects, the properties of $Y$ are then represented by the intrinsic geometric constraints of the vector set. By studying the geometric constraints that are implicit in the vector set and that correspond to $Y$, especially when $Y$ is expanded to all features of the language object, we can determine the correspondence between $\mathcal{M}$ and $\mathcal{R}$. The probe is a model that investigates the relationship between the geometric constraints of the vector set and $Y$. It is composed of a function set $F$ and a metric $E_{Y}$ defined on $Y$. The input of a function in $F$ is the representation vector of a language object, and the output is the predicted $Y$ feature of the input language object. The distance between the predicted feature and the true feature is calculated by using the metric $E_{Y}$, and a function $f$ in $F$ that minimizes the distance is determined. Here, $F$ limits the range of geometric constraints, and $E_{Y}$ limits the selection of a "good" geometry. Notably, this definition seems very similar to that of learning. Therefore, the larger the scope of $F$ is, the harder it is to discern the form of the geometric constraints, especially when $F$ is a neural network \citep{DBLP:conf/acl/PimentelVMZWC20, DBLP:conf/naacl/WhitePSC21}. However, the purpose of the probe is different from that of learning. The goal of learning is to construct a model $\mathcal{M}$ (usually a black box), which may have multiple construction methods, while the purpose of the probe is to analyze the relationship between $\mathcal{M}$ and $\mathcal{R}$.
\subsection{The Design Scheme for the Topological Probe}
One of the goals of topology is to find homeomorphic or homotopic invariants (including invariant quantities, algebraic structures, functors, etc.) and then to characterize the intrinsic structure of a topological space with these invariants. Analogously, we can view $R$ as a geometric object and $Y$ as its topology. Can we then define a concept similar to topological invariants with respect to $Y$?
We define a feature invariant for $Y$ as a set of conditions $C_{Y}$ such that any element in $Y$ satisfies $C_{Y}$. $C_{Y}$ reflects the internal constraints of the probed feature, as well as a part of the logical constraints of $R$. Furthermore, if $C_{Y}$ is well defined, it induces a set $X_{C_{Y}}$ consisting of all objects satisfying $C_{Y}$, which naturally extends the metric defined on $Y$ to $X_{C_{Y}}$.
Furthermore, just as the distance measure between two points can induce a distance measure between a point and a plane, the distance measure between the predicted feature $px$ and $X_{C_{Y}}$ can also be induced by $E_{Y}$ (denoted as $E_{C_{Y}}$):
\begin{equation}\label{eq1}
E_{C_{Y}}(px, X_{C_{Y}}) = \mathop{\min}\limits_{x\in X_{C_{Y}}}E_{Y}(px, x)
\end{equation}
It can be easily verified that if $E_{Y}$ is a well-defined distance metric on $Y$, then $E_{C_{Y}}$ should also be a well-defined distance metric on $px$. Once we have $E_{C_{Y}}$, the supervised probe $(F, E_{Y}, Y)$ can naturally induce a self-supervised probe $(F, E_{C_{Y}}, C_{Y})$. We refer to $(F, E_{C_{Y}}, C_{Y})$ as the self-supervised version of $(F, E_{Y}, Y)$, also known as the topological probe.
Notably, the prerequisite for obtaining $(F, E_{C_{Y}}, C_{Y})$ is that $C_{Y}$ must be well-defined, so $C_{Y}$ should not be a black box. Figure \ref{fig:relation.png} shows an intuitive illustration.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.5]{"relation.png"}
\caption{The relationship between the distance from the predicted feature A to $X_{C_{Y}}$ and the distance from A to $X_{Y}$.}
\label{fig:relation.png}
\end{figure}
Next, we present a specific topological probe that is based on the previously outlined design scheme and serves as a self-supervised variant of the structural probe.
\subsection{The Self-supervised Tree Topological Probe}
Given a sentence $W$, it is represented by a model $M$ as a set (or sequence) of vectors, denoted as $H=M(W)$. The number of vectors in $H$ is denoted as $L_{H}$, and we assign an index $(1,2,...L_{H})$ to each vector in $H$ so that the order of the indices matches the order of the corresponding tokens in the sentence. Additionally, we denote the dimension of the vectors as $n$. For each $W$, there exists a syntax tree $T_{W}$, where each complete word in $W$ corresponds to a node in $T_{W}$.
The probed feature $Y$ that the structural probe defines is the depth of the nodes corresponding to complete words. Following the work in \citep{DBLP:conf/naacl/HewittM19}, we set the parameter space of $F$ for the structural probe to be all real matrices of size $m*n$, where $m<n$. The specific form for predicting the depth is as follows:
\quad \qquad \; Given $p \in R,\, \forall 1\leq i\leq L_{H}$
\begin{equation}\label{eq2}
\bm{p}dep(h_{i}) = \|f*h_{i}\|^{p}
\end{equation}
where $\bm{p}dep(h_{i})$ is the prediction tree depth of $w_{i}$ in $T_{W}$ and $f$ is a real matrix of size $m*n$. Because $\forall p<2$, there is a tree that cannot be embedded as above \citep{DBLP:conf/nips/ReifYWVCPK19}, so $p$ is usually taken as $2$. $\bm{p}dep(h_{1})$, $\bm{p}dep(h_{2})$ $\cdots $, $\bm{p}dep(h_{L_{H}})$ form a sequence denoted as $\bm{p}dep_{H}$.
Moreover, we denote the true depth of $w_{i}$ as $dep(w_{i})$. Hence, $dep(w_{1})$, $dep(w_{2})$ $\cdots $, $dep(w_{L_{H}})$ also form a sequence denoted as $dep_{W}$. The metric $E$ in the structural probe is defined as follows:
$$
E(\bm{p}dep_{H}, dep_{W}) \qquad \qquad \qquad
$$
\begin{equation}\label{eq3}
=\frac{1}{L_{H}} \sum_{i=1}^{L_{H}}(\bm{p}dep(h_{i})-dep(w_{i}))^{2}
\end{equation}
Therefore, the structural probe is defined as $(|f*|^{2}$, $E$ , $dep)$.
Now we provide the constraints $C_{dep}$ for $dep$. An important limitation of $dep_{W}$ is that it is an integer sequence. Based on the characteristics of the tree structure, it is naturally determined that $dep_{W}$ must satisfy the following two conditions:
\textbf{(Boundary condition).} If $L_{H}\geq 1$, there is exactly one minimum element in $dep_{W}$, and it is equal to $1$; if $L_{H}\geq 2$, at least one element in $dep_{W}$ is equal to $2$.
\textbf{(Recursion condition).} If we sort $dep_{W}$ in ascending order to obtain the sequence $asdep_{W}$, then
$$
\forall 1\leq i \leq L_{H}-1
$$
$$
asdep(w_{i+1})=asdep(w_{i})
$$
or
$$
asdep(w_{i+1})=asdep(w_{i})+1
$$
We denote the set of all sequences that conform to $C_{dep}$ as $X_{C_{dep}}$. From equation $\ref{eq1}$, we can induce a metric $E_{C_{dep}}$:
\begin{equation}\label{eq4}
E_{C_{dep}}(\bm{p}dep_{H}, X_{C_{dep}}) = \mathop{\min}\limits_{x\in X_{C_{dep}}}E(\bm{p}dep_{H}, x)
\end{equation}
Assuming we can construct an explicit sequence $mins_{W}$ such that:
\begin{equation}\label{eq5}
mins_{W} = \mathop{\arg\min}\limits_{x\in X_{C_{dep}}}\sum_{i=1}^{L_{H}}(\bm{p}dep(h_{i})- x(w_{i}))^{2}
\end{equation}
We can obtain an analytical expression for $E_{C_{dep}}$ as follows:
\begin{equation}\label{eq6}
E_{C_{dep}}(\bm{p}dep_{H}, X_{C_{dep}}) = E(\bm{p}dep_{H}, mins_{W})
\end{equation}
Consider the following two examples:
\begin{enumerate}
\item When $\bm{p}dep_{H} = 0.8, 1.5, 1.8, 2.4, 4.5$, then $mins_{W} = 1, 2, 2, 3, 4$.
\item When $\bm{p}dep_{H} = 0.8, 1.5, 1.8, 2.4, 7.5$, then $mins_{W} = 1, 2, 3, 4, 5$.
\end{enumerate}
It can be observed that the predicted depths for nodes further down the hierarchy can also influence the corresponding values of $mins_{W}$ for nodes higher up in the hierarchy. In the examples provided, due to the change from 4.5 to 7.5, 1.8 changes from 2 to 3 at the corresponding $mins_{W}$. Therefore, using a straightforward local greedy approach may not yield an accurate calculation of $mins_{W}$, and if a simple enumeration method is employed, the computational complexity will become exponential.
However, while a local greedy approach may not always provide an exact computation of $mins_{W}$, it can still maintain a certain degree of accuracy for reasonable results of $\bm{p}dep_{H}$. This is because cases like the jump from 2.4 to 7.5 should be infrequent in a well-trained probe's computed sequence of predicted depths, unless the probed representation does not encode the tree structure well and exhibits a disruption in the middle.
Before delving into that, we first introduce some notations:
\begin{itemize}
\item $\bm{ap}dep_{H}$ denote the sequence obtained by sorting $\bm{p}dep_{H}$ in ascending order.
\item $\bm{ap}dep_{i}$ represents the $i$-th element of $\bm{ap}dep_{H}$.
\item $pre_{W}$ be a sequence in $X_{C_{dep}}$.
\end{itemize}
Here, we introduce a simple method for constructing $mins_{W}$ from a local greedy perspective.
\textbf{(Initialization).} If $L_{H}\geq 1$, let $pre(w_{1})=1$; if $L_{H}\geq 2$, let $pre(w_{2})=2$.
\textbf{(Recurrence).} If $L_{H}\geq 3$ and $3\leq i\leq L_{H}$, let
\begin{equation}\label{eq7}
pre(w_{i})=pre(w_{i-1})+bias_{i-1}
\end{equation}
where the values of $bias_{i-1}$ and $\bm{ap}dep_{H}$ are related if
$$
|pre(w_{i-1})+1-\bm{ap}dep_{i}| \leq |pre(w_{i-1})-\bm{ap}dep_{i}|
$$
$bias_{i-1}=1$; otherwise, $bias_{i-1}=0$.
\textbf{(Alignment).} Let $a_i$($1\leq i \leq L_{H}$) denote the index of $\bm{ap}dep_i$ in $\bm{p}dep_H$. Then, let
\begin{equation}\label{eq8}
pesu(w_{a_{i}})=pre(w_{i})
\end{equation}
It can be shown that $pesu_{W}$ constructed in the above manner satisfies the following theorem:
\newtheorem{thm}{Theorem}
\label{thm1}
\begin{thm}
If $\, \forall i=1$, $2$ $\cdots$, $L_{H}-1$, $\bm{ap}dep_{i+1}-\bm{ap}dep_{i}<=1$, then
$$
E(\bm{p}dep_{H}, pesu_{W}) = E(\bm{p}dep_{H}, mins_{W})
$$
\end{thm}
Therefore, $pesu_{W}$ can be considered an approximation to $mins_{W}$. Appendix \ref{sec:appendixA} contains the proof of this theorem. In the subsequent sections of this paper, we replace $E_{C_{dep}}(\bm{p}dep_{H}, X_{C_{dep}})$ with $E(\bm{p}dep_{H}, pesu_{W})$.
Additionally, an important consideration is determining the appropriate value of the minimum element for $dep_{W}$ in the boundary condition. In the preceding contents, we assumed a root depth of 1 for the syntactic tree. However, in traditional structural probe \citep{DBLP:conf/naacl/HewittM19, DBLP:conf/acl/MaudslayVPWC20,DBLP:conf/acl/LimisiewiczM20,DBLP:conf/iclr/ChenFXXTCJ21,DBLP:conf/naacl/WhitePSC21}, the root depth is typically assigned as 0 due to the annotation conventions of syntactic tree datasets. From a logical perspective, these two choices may appear indistinguishable.
However, in Appendix \ref{sec:appendixB}, we demonstrate that the choice of whether the root depth is 0 has a significant impact on the geometry defined by the tree topological probe. Furthermore, we can prove that as long as the assigned root depth is greater than 0, the optimal geometry defined by the tree topological probe remains the same to a certain extent. Therefore, in the subsequent sections of this paper, we adopt the setting where the value of the minimum element of $dep_{W}$ is $1$.
\subsection{Enhancements to the Tree Topological Probe}
Let the set of all language objects generated by rule $R$ be denoted as $\mathcal{X}_R$, and the cardinality of $\mathcal{X}_R$ be denoted as $|\mathcal{X}_R|$. The structural probe induces a metric that describes the relationship between model $M$ and $dep$:
\begin{equation}\label{eq9}
\mathcal{X}_{sp}(M) = \mathop{\min}\limits_{f\in F}\frac{1}{|\mathcal{X}_R|}\sum_{W\in \mathcal{X}_{R}}E(\bm{p}dep_{M(W)}, dep_{W})
\end{equation}
The tree topological probe can also induce a similar metric:
$\mathcal{X}_{ssp}(M) =$
\begin{equation}\label{eq10}
\mathop{\min}\limits_{f\in F}\frac{1}{|\mathcal{X}_R|}\sum_{W\in \mathcal{X}_{R}}E(\bm{p}dep_{M(W)}, mins_{W})
\end{equation}
On the other hand, we let
\begin{equation}\label{eq11}
maxs_{W} = \mathop{\arg\max}\limits_{x\in X_{C_{dep}}}\sum_{i=1}^{L_{H}}(\bm{p}dep(h_{i})- x(w_{i}))^{2}
\end{equation}
similar to $mins_{W}$, and $maxs_{W}$, inducing the following metrics:
$\mathcal{X}_{essp}(M)$
\begin{equation}\label{eq12}
= \mathop{\min}\limits_{f\in F}\frac{1}{|\mathcal{X}_R|}\sum_{W\in \mathcal{X}_{R}}E(\bm{p}dep_{M(W)}, maxs_{W})
\end{equation}
Since $dep_{W} \in X_{C_{dep}}$, when $f$ is given, we have:
$$
E(\bm{p}dep_{M(W)}, dep_{W}) \leq \mathop{\max}\limits_{x\in X_{C_{dep}}}E(\bm{p}dep_{M(W)}, x)
$$
Furthermore, as $\mathcal{X}{sp}(M)$ and $\mathcal{X}{essp}(M)$ share the same set of probing functions $F$, we have:
$$
\mathcal{X}_{sp}(M) \leq \mathcal{X}_{essp}(M)
$$
Therefore, $\mathcal{X}_{essp}(M)$ provides us with an upper bound for the structural probe metric. Similarly, for $\mathcal{X}{ssp}(M)$, we also have:
$$
\mathcal{X}_{ssp}(M) \leq \mathcal{X}_{sp}(M)
$$
Therefore, $\mathcal{X}_{ssp}(M)$ provides us with a lower bound for the structural probe metric.
In summary, we have the following:
$$
\mathcal{X}_{ssp}(M) \leq \mathcal{X}_{sp}(M) \leq \mathcal{X}_{essp}(M)
$$
If $\mathcal{X}{ssp}(M) = \mathcal{X}{essp}(M)$, then there is no difference between the tree topological probe and the structural probe. On the other hand, if it is believed that a smaller $\mathcal{X}{sp}(M)$ is desirable, then estimating $\mathcal{X}{sp}(M)$ within the range $[\mathcal{X}{ssp}(M), \mathcal{X}{essp}(M)]$ becomes an interesting problem. We consider the following:
$\theta_{W}=$
\begin{equation}\label{eq13}
\frac{E(\bm{p}dep_{M(W)}, dep_{W})-E(\bm{p}dep_{M(W)}, mins_{W})}{E(\bm{p}dep_{M(W)}, maxs_{W})-E(\bm{p}dep_{M(W)}, mins_{W})}
\end{equation}
This leads to an intriguing linguistic distribution, the distribution of $\theta_{W}\in [0, 1]$ when uniformly sampling $W$ from $\mathcal{X}_{R}$. We suppose the density function of this distribution is denoted as $P_{\theta}$, and the expectation with respect to $\theta$ is denoted as $E_{P_{\theta}}$. Then we can approximate $\mathcal{X}{sp}(M)$ as follows:
\begin{equation}\label{eq14}
\mathcal{X}{sp}(M) = E_{P_{\theta}}\mathcal{X}{essp}(M)+(1-E_{P_{\theta}})\mathcal{X}{ssp}(M)
\end{equation}
While the analysis of $P_{\theta}$ is not the primary focus of this paper, in the absence of any other constraints or biases on model $M$, we conjecture that the distribution curve of $\theta$ may resemble a uniform bell curve. Hence, we consider the following distribution approximation:
$$
P_{\theta}(x) = 6(x-x^2) \quad x\in [0, 1]
$$
At this point:
\begin{equation}\label{eq15}
\mathcal{X}{sp}(M) = \frac{1}{2}(\mathcal{X}{essp}(M)+\mathcal{X}{ssp}(M))
\end{equation}
Therefore, utilizing a self-supervised metric can approximate the unbiased optimal geometry defined by the structural probe:
\begin{equation}\label{eq16}
M_{G} = \mathop{\arg\min}\limits_{M} \frac{1}{2}(\mathcal{X}{essp}(M)+\mathcal{X}{ssp}(M))
\end{equation}
Moreover, $M_{G}$ is an analytically tractable object, implying that the metrics induced by the tree topological probe preserve to a certain extent the two metric properties discussed in the introduction. However, there is a crucial issue that remains unresolved. Can we explicitly construct $maxs_{W}$? Currently, we have not found a straightforward method similar to constructing $pesu_{W}$ for approximating $maxs_{W}$. However, based on the sorting inequality, we can construct a sequence that approximates $maxs_{W}$ based on $pre_{W}$. Let $d_{i}$($1\leq i \leq L_{H}$) denote $L_{H}-i+1$. Then, let
\begin{equation}\label{eq17}
xpesu(w_{a_{i}})=pre(w_{d_{i}})
\end{equation}
In our subsequent experiments, we approximate $E(\bm{p}dep_{H}, maxs_{W})$ with $E(\bm{p}dep_{H}, xpesu_{W})$.
\section{Experiments}
In this section, delve into a range of experiments conducted on the tree topological probe, along with the underlying motivations behind them. To accommodate space limitations, we include many specific details of the experimental settings in Appendices \ref{sec:appendixC} and \ref{sec:appendixD}. Moreover, we focus our experiments on BERT-large and its submodules. Moreover, conducting similar experiments on other models is also straightforward (refer to Appendix \ref{sec:appendixF} for supplementary results of experiments conducted using RoBERTa-large).
\subsection{Measuring $\mathcal{X}{ssp}$ and $\mathcal{X}{essp}$ on BERT}
We denote the model consisting of the input layer and the first $i$ transformer blocks of BERT-large as $M_{i}(0\leq i \leq 24)$. Since the input of $M_{i}$ consists of tokenized units, including special tokens [CLS], [SEP], [PAD], and [MASK], we can conduct at least four types of measurement experiments:
\begin{enumerate}[label=e\arabic*.]
\item Measurement of the vector set formed by token embedding and special token embedding.
\item Measurement of the vector set formed solely by token embedding.
\item Measurement of the vector set formed by estimated embedding of complete words using token embedding and special token embedding.
\item Measurement of the vector set formed solely by estimated embedding of complete words using token embedding.
\end{enumerate}
Similarly, due to space constraints, we focus on discussing e1 in this paper. The measurement results are shown in Tables \ref{tab:1} and \ref{tab:2}. The precise measurement values can be found in Appendix \ref{sec:appendixE}. Furthermore, as shown in Figure \ref{fig:log_metric.png}, we present the negative logarithm curves of three measurement values as a function of $M_{i}$ variation.
\begin{table}[htbp]
\centering
\begin{tabular}{cc}
\hline
\textbf{$\mathcal{X}{ssp}$} & \textbf{$M$} \\
\hline
0.01$\sim$0.05 & $M_{0}$$\sim$$M_{11}$ \\
0.05$\sim$0.1 & $M_{12}$$\sim$$M_{21}$ \\
0.1$\sim$0.15 & $M_{22}$$\sim$$M_{23}$ \\
\hline
\end{tabular}
\caption{\label{tab:1} Grouping $M_{i}$ based on $\mathcal{X}{ssp}$. $M_{l}$$\sim$$M_{r}$ denotes $M_{l}, M_{l+1}, M_{l+2}, \ldots, M_{r}$. For example, the first row of the table indicates that the exact values of $\mathcal{X}{ssp}$ for $M_{0}, M_{1}, M_{2}, \ldots, M_{11}$ fall within the range of 0.01 to 0.05.}
\end{table}
\begin{table}[htbp]
\centering
\begin{tabular}{cccc}
\hline
\textbf{$\mathcal{X}{essp}$} & \multicolumn{3}{c}{\textbf{$M$}} \\
\hline
0.3$\sim$0.4 & $M_{3}$$\sim$$M_{4}$ & $M_{7}$$\sim$$M_{12}$ & \phantom{} \\
0.4$\sim$0.5 & \phantom{} & $M_{13}$$\sim$$M_{14}$ & \phantom{} \\
0.5$\sim$1.0 & $M_{1}$$\sim$$M_{2}$ & $M_{5}$$\sim$$M_{6}$ & $M_{15}$$\sim$$M_{19}$ \\
1.0$\sim$2.0 & \phantom{} & \phantom{} & $M_{20}$$\sim$$M_{24}$ \\
$\geq 4.0$ & $M_{0}$ & \phantom{} & \phantom{}\\
\hline
\end{tabular}
\caption{\label{tab:2} Grouping $M_{i}$ based on $\mathcal{X}{essp}$. Similar to the explanation in the caption of Table \ref{tab:1}.}
\end{table}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.55]{"log_metric.png"}
\caption{Negative logarithm of $\mathcal{X}_{ssp}$, $\mathcal{X}_{essp}$, unbiased $\mathcal{X}_{sp}$ and true $\mathcal{X}_{sp}$ across $M_{i}$.}
\label{fig:log_metric.png}
\end{figure}
By examining the experimental results presented above, we can ascertain the following findings:
\begin{enumerate}[label=f\arabic*.]
\item $\mathcal{X}{ssp}$ and $\mathcal{X}{essp}$ indeed bound the actual $\mathcal{X}{sp}$, and for $M_{14}$ to $M_{18}$, their true $\mathcal{X}{sp}$ are very close to their $\mathcal{X}{ssp}$.
\item $M_{0}$ serves as a good baseline model. Furthermore, using $\mathcal{X}{essp}$ and unbiased $\mathcal{X}{sp}$ allows for effective differentiation between embeddings generated by models consisting solely of the regular input layer and those generated by models incorporating transformer blocks.
\item For $M_{1}$ to $M_{6}$, their true $\mathcal{X}{sp}$ are very close to their unbiased $\mathcal{X}{sp}$.
\item Both the curve of $-log(\mathcal{X}{essp})$ and the curve of the true $-log(\mathcal{X}{sp})$ follow an ascending-then-descending pattern. However, the models corresponding to their highest points are different, namely, $M_{8}$ and $M_{16}$, respectively.
\item For the curve of $-log(\mathcal{X}{ssp})$, its overall trend also shows an ascending-then-descending pattern but with some fluctuations in the range of $M_{3}$ to $M_{6}$. However, the model corresponding to its highest point is consistent with $-log(\mathcal{X}{essp})$, which is $M_{8}$.
\item The true $\mathcal{X}{sp}$ does not effectively distinguish between $M_{0}$ and $M_{1}$.
\end{enumerate}
Based on the above findings, we can confidently draw the following rigorous conclusions:
\begin{enumerate}[label=c\arabic*.]
\item Based on f1, we can almost infer that $dep_{W} \in \mathop{\arg\min}\limits_{x\in X_{C_{dep}}}\sum_{i=1}^{L_{H}}(\bm{p}dep(h_{i})- x(w_{i}))^{2}$ for $M_{14}$ to $M_{18}$. \textbf{This implies that they memorize the preferences of the real data and minimize as much as possible to approach the theoretical boundary}. Building upon f5, we can further conclude that the cost of memorizing $dep_{W}$ is an increase in $\mathcal{X}{ssp}$, which leads to a decrease in the accuracy of the embedding's linear encoding for tree structures.
\item Based on f1, we can conclude that there exists a model $M$ where the true $\mathcal{X}{sp}(M)$ aligns with the $\mathcal{X}{ssp}(M)$ determined by $C_{dep}$. \textbf{This indicates that $C_{dep}$ serves as a sufficiently tight condition}.
\item Based on f3, we can infer that $M_{1}$ to $M_{6}$ may not capture the distributional information of the actual syntactic trees, resulting in their generated embeddings considering only the most general case for linear encoding of tree structures. This implies that the distribution curve of their $\theta_{W}$ parameters is uniformly bell-shaped.
\item Based on f2 and f6, we can conclude that the tree topological probe provides a more fine-grained evaluation of the ability to linearly encode tree structures in embedding vectors compared to the structural probe.
\item Based on f3, f4 and f5, we can conclude that in BERT-large, embedding generated by $M_{8}$ and its neighboring models exhibit the strongest ability to linearly encode tree structures. Moreover, they gradually start to consider the distribution of real dependency trees, resulting in the true $\mathcal{X}{sp}(M)$ approaching $\mathcal{X}{ssp}(M)$ until reaching $M_{16}$.
\item Based on f4 and f5, we can conclude that starting from $M_{16}$, the embeddings generated by $M_{i}$ gradually lose their ability to linearly encode tree structures. The values of $\mathcal{X}{ssp}$ and $\mathcal{X}{essp}$ for these models are generally larger compared to models before $M_{16}$. However, they still retain some distributional information about the depth of dependency trees. This means that despite having a higher unbiased $\mathcal{X}{sp}$, their true $\mathcal{X}{sp}$ is still smaller than that of $M_{i}$ before $M_{8}$.
\end{enumerate}
From the above conclusions, we can further speculate about the workings of pretrained language models such as BERT, and we identify some related open problems.
Based on c5 and c6, we can speculate that the final layer of a pretrained language model needs to consider language information at various levels, but its memory capacity is limited. Therefore, it relies on preceding submodules to filter the information. The earlier submodules in the model encode the most generic (unbiased) structures present in the language features. As the model advances, the intermediate submodules start incorporating preferences for general structures based on actual data. Once a certain stage is reached, the later submodules in the model start to loosen their encoding of generic structures. However, due to the preference information passed from the intermediate submodules, the later submodules can still outperform the earlier submodules in encoding real structures, rather than generic ones.
Based on c3 and c6, it appears that true $\mathcal{X}{sp} \leq$ unbiased $\mathcal{X}{sp} <$ $\mathcal{X}{essp}$. This suggests that for BERT, unbiased $\mathcal{X}{sp}$ serves as a tighter upper bound for $\mathcal{X}{sp}$, and there exists a submodule that achieves this upper bound. Now, the question arises: Is this also the case for general pretrained models? If so, what are the underlying reasons?
\subsection{Using $\mathcal{X}{ssp}$ and $\mathcal{X}{essp}$ as Regularization Loss in Fine-tuning BERT}
Let us denote the downstream task loss as $T(M_{24})$. Taking $\mathcal{X}{ssp}$ as an example, using $\mathcal{X}{ssp}$ as a regularizing loss during fine-tuning refers to replacing the task loss with:
$$
T(M_{24})+\lambda * \mathcal{X}{ssp}(M_{i}) \quad (1 \leq i \leq 24)
$$
where $\lambda$ is a regularization parameter. The purpose of this approach is to explore the potential for enhancing the fine-tuning performance by improving the submodules of BERT in their ability to linearly encode tree structures. If there exists a submodule that achieves both enhancement in linear encoding capabilities and improved fine-tuning performance, it implies that the parameter space of this submodule, which has better linear encoding abilities, overlaps with the optimization space of fine-tuning. This intersection is smaller than the optimization space of direct fine-tuning, reducing susceptibility to local optima and leading to improved fine-tuning results.
Conversely, if enhancing certain submodules hinders fine-tuning or even leads to its failure, it suggests that the submodule's parameter space, which has better linear encoding abilities, does not overlap with the optimization space of fine-tuning. This indicates that the submodule has already attained the smallest $\mathcal{X}{ssp}$ value that greatly benefits the BERT's performance.
Based on f1, we can infer that $M_{14}$ to $M_{18}$ are not suitable as enhanced submodules. According to c5, the submodules most likely to improve fine-tuning performance after enhancement should be near $M_{8}$. We conducted experiments on a single-sentence task called the Corpus of Linguistic Acceptability (CoLA) \citep{DBLP:journals/tacl/WarstadtSB19}, which is part of The General Language Understanding Evaluation (GLUE) benchmark \citep{DBLP:conf/iclr/WangSMHLB19}.
The test results are shown in Table \ref{tab:3}. As predicted earlier, enhancing the submodules around $M_{14}$ to $M_{18}$ (now expanded to $M_{12}$ to $M_{19}$) proves to be detrimental to fine-tuning, resulting in failed performance. However, we did observe an improvement in fine-tuning performance for the sub-module $M_{10}$ near $M_{8}$ after enhancement. This gives us an intuition that if we have additional topological probes and similar metrics to $\mathcal{X}{ssp}$ and $\mathcal{X}{sp}$, we can explore enhancing submodules that are in the rising phase of true $\mathcal{X}{sp}$, away from the boundary of unbiased $\mathcal{X}{sp}$ and $\mathcal{X}{ssp}$, in an attempt to improve fine-tuning outcomes.
\begin{table}[htbp]
\centering
\begin{tabular}{cccc}
\hline
\textbf{Method} & \textbf{mean} & \textbf{std} & \textbf{max} \\
\hline
DF & 63.34 & 1.71 & 66.54 \\
EH $M_{3}$ & 63.90 & 2.66 & 68.73 \\
EH $M_{5}$ & 63.90 & 1.36 & 66.04 \\
EH $M_{10}$ & \textbf{64.87} & 2.07 & \textbf{68.47} \\
EH $M_{12}$$\sim$$M_{19}$ & 0.00 & 0.00 & 0.00 \\
EH $M_{20}$ & 5.48 & 16.46 & 54.87 \\
EH $M_{24}$ & 40.43 & 26.60 & 62.52 \\
\hline
\end{tabular}
\caption{\label{tab:3} Direct fine-tuning and sub-module enhancement test scores. Here, "DF" denotes direct fine-tuning, while "EH $M_{i}$" represents the fine-tuning with the enhancement of $M_{i}$ based on $\mathcal{X}{ssp}$. The evaluation metric used in CoLA is the Matthew coefficient, where a higher value indicates better performance.}
\end{table}
\section{Conclusion}
Consider a thought experiment where there is a planet in a parallel universe called "Vzjgs" with a language called "Vmtprhs". Like "English", "Vmtprhs" comprises 26 letters as basic units, and there is a one-to-one correspondence between the letters of "Vmtprhs" and "English". Moreover, these two languages are isomorphic under letter permutation operations. In other words, sentences in "English" can be rearranged so that they are equivalent to sentences in "Vmtprhs", while preserving the same meaning. If there were models like BERT or GPT in the "Vzjgs" planet, perhaps called "YVJIG" and "TLG," would the pretraining process of "YVJIG" on "Vmtprhs" be the same as BERT's pretraining on "English"?
In theory, there should be no means to differentiate between these two pretraining processes. For a blank model (without any training), extracting useful information from "Vmtprhs" and "English" would pose the same level of difficulty. However, it is true that "Vmtprhs" and "English" are distinct, with the letters of "Vmtprhs" possibly having different shapes or being the reverse order of the "English" alphabet. Therefore, we can say that they have different letter features, although this feature seems to be a mere coincidence. In natural language, there are many such features created by historical contingencies, such as slang or grammatical exceptions. Hence, when we aim to interpret the mechanisms of these black-box models by studying how language models represent language-specific features, we must consider which features are advantageous for interpretation and what we ultimately hope to gain from this research.
This paper presents a thorough exploration of a key issue, specifically examining the articulation of internal feature constraints. By enveloping the original feature within a feature space that adheres to such constraints, it is possible to effectively eliminate any unintended or accidental components. Within this explicitly defined feature space, metrics such as $\mathcal{X}{ssp}$ and $\mathcal{X}{essp}$ can be defined. We can subsequently examine the evolution of these metrics within the model to gain a deeper understanding of the encoding strategies employed by the model for the original feature, as described in the experimental section of this paper. Once we understand the encoding strategies employed by the model, we can investigate the reasons behind their formation and the benefits they bring to the model. By conducting studies on multiple similar features, we can gain a comprehensive understanding of the inner workings of the black box.
\section*{Limitations}
The main limitation of this research lies in the approximate construction of $mins_{W}$ and $maxs_{W}$, which leads to true $-log(\mathcal{X}{sp})$ surpassing $-log(\mathcal{X}{ssp})$ near $M_{16}$ to some extent. However, this may also be due to their proximity, resulting in fluctuations within the training error. On the other hand, the proposed construction scheme for the topological probe discussed in this paper lacks sufficient mathematical formalization. One possible approach is to restate it using the language of category theory.
\iffalse
\section*{Ethics Statement}
Due to the emergence of ChatGPT, large-scale language models have become one of the most discussed topics in the NLP community. These models are currently predominantly pre-trained through unsupervised or weakly supervised methods, and then fine-tuned using various techniques such as RLHF to specialize their capabilities. While fine-tuning plays a crucial role in the ultimate performance of the models, it is the pre-trained model itself that determines its upper and lower limits. Therefore, in the ongoing discussions about the potential impact of large models on human society in the near future, providing a specific explanation of the workings of the post-pre-training model becomes important to make these discussions more rational and scientific. Moreover, it may inspire further reflections on the advantages of human intelligence compared to the intelligence of these models and how to expand on those advantages. On the other hand, the large-scale data used during model pre-training may inevitably contain content that violates societal norms and decency. Having an accurate understanding of which portions of the data are memorized in which submodules of the model is crucial to ensure controllable output results of the final model after fine-tuning. This work provides some intriguing insights into how to address these issues.
\fi
\section*{Acknowledgements}
We thank the anonymous reviewers for their helpful comments and suggestions. This work was supported by National Natural Science Foundation of China (Nos. 62362015, 62062027 and U22A2099) and the project of Guangxi Key Laboratory of Trusted Software.
% Entries for the entire Anthology, followed by custom entries
\bibliography{custom}
\bibliographystyle{acl_natbib}
\appendix
\section{Proof of Theorem 1}
\label{sec:appendixA}
\begin{proof}[Proof]
For any sequence $x \in X_{C_{dep}}$ that is in the same order as $\bm{p}dep_{H}$, according to the inequality of rankings, for any permutation $\pi_{x}$ of $x$, we have:
$$
\sum_{i=1}^{L_{H}}\pi_{x}(w_{i})*\bm{p}dep(h_{i})\leq \sum_{i=1}^{L_{H}}x(w_{i})*\bm{p}dep(h_{i})
$$
Therefore,
$$
\sum_{i=1}^{L_{H}}(\pi_{x}(w_{i}) - \bm{p}dep(h_{i}))^{2}
$$
$$
\geq \sum_{i=1}^{L_{H}}(x(w_{i}) - \bm{p}dep(h_{i}))^{2}
$$
Since $pesu_{W}$ and $\bm{p}dep_{H}$ are in the same order, we just need to prove that any sequence $x \in X_{C_{dep}}$ and in the same order as $\bm{p}dep_{H}$ satisfies
$$
\sum_{i=1}^{L_{H}}(x(w_{i}) - \bm{p}dep(h_{i}))^{2}
$$
$$
\geq \sum_{i=1}^{L_{H}}(pesu(w_{i}) - \bm{p}dep(h_{i}))^{2}
$$
The theorem is automatically established. Because
$$
\sum_{i=1}^{L_{H}}(pesu(w_{i}) - \bm{p}dep(h_{i}))^{2}
$$
\begin{equation}\label{eq18}
=\sum_{i=1}^{L_{H}}(pre(w_{i}) - \bm{ap}dep_{i})^{2}
\end{equation},
without loss of generality, we can assume $pesu_{W}$ and $x$ to be ascending sequences and not equal and exist a $k$ such that when $1\leq i\leq k-1$
\begin{equation}\label{eq19}
pesu(w_{i}) = x(w_{i})
\end{equation}
and
\begin{equation}\label{eq20}
pesu(w_{k}) \neq x(w_{k})
\end{equation}
Based on the recursive condition, we can infer that
\begin{equation}\label{eq21}
|pesu(w_{k}) - x(w_{k})|=1
\end{equation}
Combined with the value condition of $bias_{k-1}$, we further find that
$$
|pesu(w_{k}) - \bm{ap}dep_{k}| \leq |x(w_{k}) - \bm{ap}dep_{k}|
$$
The inductive hypothesis when $i=m$ is
$$
|pesu(w_{m}) - \bm{ap}dep_{m}| \leq |x(w_{m}) - \bm{ap}dep_{m}|
$$
Due to the condition $\bm{ap}dep_{m+1}-\bm{ap}dep_{m} \leq 1$ and the value condition of $bias_{m}$, it still holds when $i=m+1$ that
$$
|pesu(w_{m+1}) - \bm{ap}dep_{m+1}|
$$
$$
\leq |x(w_{m+1}) - \bm{ap}dep_{m+1}|
$$
Thus, when $i\geq k$
$$
(x(w_{i}) - \bm{p}dep(h_{i}))^{2}\geq (pesu(w_{i}) - \bm{p}dep(h_{i}))^{2}.
$$
\end{proof}
\section{Analysis of Tree Depth Minimum}
\label{sec:appendixB}
The minimum of $pesu_{W}$ is denoted as $dep_{min}$. Fixing $pesu_{W}$, we let all sets (or sequences) of vectors satisfying the following conditions compose a set denoted by $\Omega_{pesu_{W}}$.
$$
\exists \, P\in R^{m*n}, \, \forall \, i(i=1, 2,\cdots L_{H})
$$
$$
pesu(w_{i})-\sqrt{\epsilon_{i}} < h_{i}^{T}P^{T}Ph_{i} < pesu(w_{i})+\sqrt{\epsilon_{i}}
$$
Here, $\epsilon_{i} \ll pesu(w_{i})^{2}$. Let $pesu(w_{1})$ be $dep_{min}$ and $pesu(w_{i})\leq pesu(w_{i+1})(i=1,2,\cdots L_{H}-1)$ without loss of generality, and the following theorem can be obtained.
\begin{thm}
\label{thm2}
For any two different sequences $pesu_{W}$ and $pesu_{W}^{'}$, if $pesu(w_{1}) > 0$ and $pesu^{'}(w_{1}) > 0$. there is a one-to-one mapping $\phi$ between $\Omega_{pesu_{W}}$ and $\Omega_{pesu_{W}^{'}}$.
\end{thm}
\begin{proof}[Proof]
We construct $\phi$ such that
$$
\forall H\in \Omega_{pesu_{W}}
$$
$$
\phi(H)=(h_{1}^{'}, h_{2}^{'},\cdots h_{L_{H}}^{'})=H^{'} \in \Omega_{pesu_{W}^{'}}
$$
Here, $h_{1}^{'}=h_{1}$, and when $i=2, 3 \cdots, L_{H}$
$$
h_{i}^{'}=\frac{\sqrt{pesu^{'}(w_{i})*pesu(w_{1})}}{\sqrt{pesu(w_{i})*pesu^{'}(w_{1})}}h_{i}
$$
Since $$\exists P\in R^{m*n}, \forall i(i=1, 2,\cdots L_{H})$$
$$
pesu(w_{i})-\sqrt{\epsilon_{i}} < h_{i}^{T}P^{T}Ph_{i} < pesu(w_{i})+\sqrt{\epsilon_{i}}
$$
$$
\epsilon_{i} \ll pesu(w_{i})^{2}
$$
Let $P^{'}=\frac{\sqrt{pesu^{'}(w_{1})}}{\sqrt{pesu(w_{1})}}P$ and when $i=1, 2, \cdots L_{H}$
$$
\epsilon_{i}^{'}=(\frac{pesu^{'}(w_{i})}{pesu(w_{i})})^{2}\epsilon_{i},
$$
then
$$
\epsilon_{i}^{'}\ll (\frac{pesu^{'}(w_{i})}{pesu(w_{i})})^{2}pesu(w_{i})^{2} =pesu^{'}(w_{i})^{2}
$$
After calculation,
$$
\forall i(i=1, 2,\cdots L_{H})
$$
$$
pesu^{'}(w_{i})-\sqrt{\epsilon_{i}^{'}}
$$
$$
<(h_{i}^{'})^{T}(P^{'})^{T}P^{'}h_{i}^{'}
$$
$$
<pesu^{'}(w_{i})+\sqrt{\epsilon_{i}^{'}}
$$
Therefore, $\phi$ is well defined, and $\forall H_{i}, H_{j}\in \Omega_{pesu_{W}}$ when $H_{i} \neq H_{j}$
$$
\phi(H_{i})\neq \phi(H_{j})
$$
Therefore, $\phi$ is also an injective function. It is easy to prove that the inverse map $\phi^{-1}$ of $\phi$ is also an injective function and satisfies the above conditions.
\end{proof}
The proof of the theorem above does not apply to the cases where $pesu(w_{1})=0$ or $pesu^{'}(w_{1})=0$. If $dep_{min}$ is greater than $0$, then the results of the tree topological probe do not necessarily depend on the selection of $dep_{min}$, and we may set it as $1$. However, we have not further explored whether Theorem \ref{thm2} is necessarily invalid. Nevertheless, we can examine the drawbacks that arise from setting $dep_{min}$ to $0$ from another perspective.
When $i\geq 2$, $h_{i}$ is projected by $P$ near the ($m$)-dimensional sphere with a radius of $\sqrt{pesu(w_{i})}$,
$$
\forall i=1, 2 \cdots, L_{H}
$$
$$
|h_{i}^{T}P^{T}Ph_{i}-pesu(w_{i})|<\epsilon_{i}
$$
If $dep_{min}=0$, then the topology of the geometric space composed of all vectors $Ph_{1}$ satisfying $|h_{1}^{T}P^{T}Ph_{1}-dep_{min}|<\epsilon_{1}$ is homeomorphic to an $m$-dimensional open ball. This may result in probes exhibiting different preferences for the root and other nodes. However, if $dep_{min}>0$, the topology of the geometric space is an $m$-dimensional annulus, which is the same for all nodes, thus avoiding the issue of preference.
\section{Data for Training and Evaluating Probes}
\label{sec:appendixC}
To ensure the reliability and diversity of data (appropriate sentences) sources, we separated the sentences participating in the probe experiment from the training, verification and test data sets of some tasks of The General Language Understanding Evaluation (GLUE) benchmark \citep{DBLP:conf/iclr/WangSMHLB19}.
We selected four small sample text classification tasks in GLUE with reference to \citep{DBLP:conf/naacl/HuaLDXL21}, namely, the Corpus of Linguistic Acceptability (CoLA) \citep{DBLP:journals/tacl/WarstadtSB19}, Microsoft Research Paraphrase Corpus (MRPC) \citep{DBLP:conf/acl-iwp/DolanB05}, Recognizing Textual Entailment (RTE) \citep{DBLP:conf/iclr/WangSMHLB19} and Semantic Textual Similarity Benchmark (STS-B) \citep{DBLP:journals/corr/abs-1708-00055}, which cover the three major task types of SINGLE-SENTENCE, SIMILARITY AND PARAPHRASE and INFERENCE in GLUE. MRPC, RTE and STS-B are all double sentence tasks, and the experiment needs only BERT to represent a single sentence; thus, we consider two sentences that belong to the same group of data independently, not spliced.
After the data sets of the four tasks are processed as above, the remaining statements are merged into a raw text data set $rtd_{mix}$, which contains 47136 sentences. This is close to the size of the Pennsylvania tree database \citep{DBLP:journals/coling/MarcusSM94} used by the structural probe \citep{DBLP:conf/naacl/HewittM19}; short and long sentences are evenly distributed.
\section{Experimental Setup for Training Probes and Fine-tuning}
\label{sec:appendixD}
We use the BERT implementation of \citeauthor{DBLP:journals/corr/abs-1910-03771} and set the rank of the probe matrix to be half the embedding dimension. The probe matrix is randomly initialized following a uniform distribution $U(-0.05, 0.05)$.
We employ the AdamW optimizer with the warmup technique, where the initial learning rate is set to 2e-5 and the epsilon value is set to 1e-8. The training stops after 10 epochs. The training setup for fine-tuning experiments is similar to that of training probes. One notable difference is the regularization coefficient $\lambda$, which is dynamically determined after one epoch of training, ensuring that $\frac{\lambda * \mathcal{X}_{ssp}(M_{i})}{T(M_{24})}\approx 0.1$, without any manual tuning.
We conduct experiments on each fine-tuning method by using 10 different random seeds, and we compute the mean, the standard deviation (std), and the maximum values.
\section{Supplementary Chart Materials}
\label{sec:appendixE}
Table \ref{tab:4} lists the exact measurements of $\mathcal{X}{ssp}$, $\mathcal{X}{essp}$, and true $\mathcal{X}_{sp}$ for BERT-Large.
\begin{table}[htbp]
\centering
\begin{tabular}{cccc}
\hline
\textbf{$M$} & \textbf{$\mathcal{X}_{ssp}$} & \textbf{$\mathcal{X}_{essp}$} & \textbf{$\mathcal{X}_{tsp}$}\\
\hline
$M_{0}$ & 0.039 & 5.382 & 0.3084 \\
$M_{1}$ & 0.017 & 0.536 & 0.2644 \\
$M_{2}$ & 0.017 & 0.526 & 0.244 \\
$M_{3}$ & 0.018 & 0.348 & 0.2016 \\
$M_{4}$ & 0.033 & 0.351 & 0.1701 \\
$M_{5}$ & 0.025 & 0.52 & 0.1622 \\
$M_{6}$ & 0.023 & 0.52 & 0.1559 \\
$M_{7}$ & 0.013 & 0.345 & 0.14 \\
$M_{8}$ & 0.01 & 0.347 & 0.1424 \\
$M_{9}$ & 0.011 & 0.352 & 0.1577 \\
$M_{10}$ & 0.013 & 0.359 & 0.1415 \\
$M_{11}$ & 0.021 & 0.375 & 0.1128 \\
$M_{12}$ & 0.054 & 0.391 & 0.0975 \\
$M_{13}$ & 0.076 & 0.42 & 0.0764 \\
$M_{14}$ & 0.084 & 0.467 & 0.0651 \\
$M_{15}$ & 0.088 & 0.525 & 0.0616 \\
$M_{16}$ & 0.09 & 0.663 & 0.0656 \\
$M_{17}$ & 0.086 & 0.785 & 0.0808 \\
$M_{18}$ & 0.09 & 0.883 & 0.1155 \\
$M_{19}$ & 0.09 & 0.999 & 0.1416 \\
$M_{20}$ & 0.092 & 1.045 & 0.1615 \\
$M_{21}$ & 0.094 & 1.447 & 0.2468 \\
$M_{22}$ & 0.102 & 1.715 & 0.28634 \\
$M_{23}$ & 0.107 & 1.709 & 0.3171 \\
$M_{24}$ & 0.113 & 1.837 & 0.328 \\
\hline
\end{tabular}
\caption{\label{tab:4} Exact values of $\mathcal{X}_{ssp}$, $\mathcal{X}_{essp}$, and true $\mathcal{X}_{sp}$ for $M_{i}$}
\end{table}
\section{Experimental data for RoBERTa-large}
\label{sec:appendixF}
Figure \ref{fig:log_metric2.png} shows the negative logarithm curves of three measurement values as a function of variation in Mi for RoBERTa-Large. Table \ref{tab:5} lists the exact measurements of $\mathcal{X}{ssp}$, $\mathcal{X}{essp}$, and true $\mathcal{X}_{sp}$ for RoBERTa-Large.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.55]{"log_metric2.png"}
\caption{Negative logarithm of $\mathcal{X}_{ssp}$, $\mathcal{X}_{essp}$, unbiased $\mathcal{X}_{sp}$ and true $\mathcal{X}_{sp}$ across $M_{i}$.}
\label{fig:log_metric2.png}
\end{figure}
From the experimental data, it is evident that the overall pattern of evolution in the graphs for RoBERTa-Large and BERT-Large is consistent. There's a slight initial increase followed by a decline, but the boundaries for $\mathcal{X}{sp}$ in the case of RoBERTa-Large are much tighter, especially in the earlier modules.
\begin{table}[htbp]
\centering
\begin{tabular}{cccc}
\hline
\textbf{$M$} & \textbf{$\mathcal{X}_{ssp}$} & \textbf{$\mathcal{X}_{essp}$} & \textbf{$\mathcal{X}_{tsp}$}\\
\hline
$M_{0}$ & 0.008 & 3.532 & 0.991 \\
$M_{1}$ & 0.145 & 0.515 & 0.243 \\
$M_{2}$ & 0.137 & 0.469 & 0.446 \\
$M_{3}$ & 0.139 & 0.470 & 0.331 \\
$M_{4}$ & 0.131 & 0.493 & 0.257 \\
$M_{5}$ & 0.132 & 0.500 & 0.199 \\
$M_{6}$ & 0.123 & 0.494 & 0.153 \\
$M_{7}$ & 0.117 & 0.491 & 0.110 \\
$M_{8}$ & 0.117 & 0.542 & 0.109 \\
$M_{9}$ & 0.114 & 0.491 & 0.087 \\
$M_{10}$ & 0.113 & 0.555 & 0.091 \\
$M_{11}$ & 0.113 & 0.567 & 0.091 \\
$M_{12}$ & 0.110 & 0.598 & 0.094 \\
$M_{13}$ & 0.114 & 0.667 & 0.101 \\
$M_{14}$ & 0.113 & 0.675 & 0.102 \\
$M_{15}$ & 0.124 & 0.738 & 0.120 \\
$M_{16}$ & 0.133 & 0.797 & 0.136 \\
$M_{17}$ & 0.136 & 0.789 & 0.131 \\
$M_{18}$ & 0.137 & 0.807 & 0.136 \\
$M_{19}$ & 0.136 & 0.831 & 0.135 \\
$M_{20}$ & 0.136 & 0.880 & 0.145 \\
$M_{21}$ & 0.150 & 1.086 & 0.169 \\
$M_{22}$ & 0.153 & 1.277 & 0.176 \\
$M_{23}$ & 0.190 & 2.006 & 0.197 \\
$M_{24}$ & 0.117 & 0.711 & 0.201 \\
\hline
\end{tabular}
\caption{\label{tab:5} Exact values of $\mathcal{X}_{ssp}$, $\mathcal{X}_{essp}$, and true $\mathcal{X}_{sp}$ for $M_{i}$}
\end{table}
\end{document}
\begin{thebibliography}{54}
\expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi
\bibitem[{Adi et~al.(2017)Adi, Kermany, Belinkov, Lavi, and
Goldberg}]{DBLP:conf/iclr/AdiKBLG17}
Yossi Adi, Einat Kermany, Yonatan Belinkov, Ofer Lavi, and Yoav Goldberg. 2017.
\newblock \href {https://openreview.net/forum?id=BJh6Ztuxl} {Fine-grained
analysis of sentence embeddings using auxiliary prediction tasks}.
\newblock In \emph{5th International Conference on Learning Representations,
{ICLR} 2017}.
\bibitem[{Aghajanyan et~al.(2021)Aghajanyan, Shrivastava, Gupta, Goyal,
Zettlemoyer, and Gupta}]{DBLP:conf/iclr/AghajanyanSGGZG21}
Armen Aghajanyan, Akshat Shrivastava, Anchit Gupta, Naman Goyal, Luke
Zettlemoyer, and Sonal Gupta. 2021.
\newblock \href {https://openreview.net/forum?id=OQ08SN70M1V} {Better
fine-tuning by reducing representational collapse}.
\newblock In \emph{9th International Conference on Learning Representations,
{ICLR} 2021}.
\bibitem[{Aho and Ullman(1972)}]{Aho:72}
Alfred~V. Aho and Jeffrey~D. Ullman. 1972.
\newblock \emph{The Theory of Parsing, Translation and Compiling}, volume~1.
\newblock Prentice-Hall, Englewood Cliffs, NJ.
\bibitem[{Alain and Bengio(2017)}]{DBLP:conf/iclr/AlainB17}
Guillaume Alain and Yoshua Bengio. 2017.
\newblock \href {https://openreview.net/forum?id=HJ4-rAVtl} {Understanding
intermediate layers using linear classifier probes}.
\newblock In \emph{5th International Conference on Learning Representations,
{ICLR} 2017, Workshop Track Proceedings}.
\bibitem[{{American Psychological Association}(1983)}]{APA:83}
{American Psychological Association}. 1983.
\newblock \emph{Publications Manual}.
\newblock American Psychological Association, Washington, DC.
\bibitem[{Ando and Zhang(2005)}]{Ando2005}
Rie~Kubota Ando and Tong Zhang. 2005.
\newblock A framework for learning predictive structures from multiple tasks
and unlabeled data.
\newblock \emph{Journal of Machine Learning Research}, 6:1817--1853.
\bibitem[{Andrew and Gao(2007)}]{andrew2007scalable}
Galen Andrew and Jianfeng Gao. 2007.
\newblock Scalable training of {L1}-regularized log-linear models.
\newblock In \emph{Proceedings of the 24th International Conference on Machine
Learning}, pages 33--40.
\bibitem[{Belinkov and Glass(2019)}]{DBLP:journals/tacl/BelinkovG19}
Yonatan Belinkov and James~R. Glass. 2019.
\newblock \href {https://doi.org/10.1162/tacl\_a\_00254} {Analysis methods in
neural language processing: {A} survey}.
\newblock \emph{Trans. Assoc. Comput. Linguistics}, 7:49--72.
\bibitem[{Cer et~al.(2017)Cer, Diab, Agirre, Lopez{-}Gazpio, and
Specia}]{DBLP:journals/corr/abs-1708-00055}
Daniel~M. Cer, Mona~T. Diab, Eneko Agirre, I{\~{n}}igo Lopez{-}Gazpio, and
Lucia Specia. 2017.
\newblock \href {http://arxiv.org/abs/1708.00055} {Semeval-2017 task 1:
Semantic textual similarity - multilingual and cross-lingual focused
evaluation}.
\newblock \emph{CoRR}, abs/1708.00055.
\bibitem[{Chandra et~al.(1981)Chandra, Kozen, and Stockmeyer}]{Chandra:81}
Ashok~K. Chandra, Dexter~C. Kozen, and Larry~J. Stockmeyer. 1981.
\newblock \href {https://doi.org/10.1145/322234.322243} {Alternation}.
\newblock \emph{Journal of the Association for Computing Machinery},
28(1):114--133.
\bibitem[{Chen et~al.(2021)Chen, Fu, Xu, Xie, Tan, Chen, and
Jing}]{DBLP:conf/iclr/ChenFXXTCJ21}
Boli Chen, Yao Fu, Guangwei Xu, Pengjun Xie, Chuanqi Tan, Mosha Chen, and
Liping Jing. 2021.
\newblock \href {https://openreview.net/forum?id=17VnwXYZyhH} {Probing {BERT}
in hyperbolic spaces}.
\newblock In \emph{9th International Conference on Learning Representations,
{ICLR} 2021}.
\bibitem[{Chicco et~al.(2021)Chicco, T{\"{o}}tsch, and
Jurman}]{DBLP:journals/biodatamining/ChiccoTJ21}
Davide Chicco, Niklas T{\"{o}}tsch, and Giuseppe Jurman. 2021.
\newblock \href {https://doi.org/10.1186/s13040-021-00244-z} {The matthews
correlation coefficient {(MCC)} is more reliable than balanced accuracy,
bookmaker informedness, and markedness in two-class confusion matrix
evaluation}.
\newblock \emph{BioData Min.}, 14(1):13.
\bibitem[{Conneau et~al.(2018)Conneau, Kruszewski, Lample, Barrault, and
Baroni}]{DBLP:conf/acl/BaroniBLKC18}
Alexis Conneau, Germ{\'{a}}n Kruszewski, Guillaume Lample, Lo{\"{\i}}c
Barrault, and Marco Baroni. 2018.
\newblock \href {https://doi.org/10.18653/v1/P18-1198} {What you can cram into
a single {\textbackslash}{\textdollar}{\&}!{\#}* vector: Probing sentence
embeddings for linguistic properties}.
\newblock In \emph{Proceedings of the 56th Annual Meeting of the Association
for Computational Linguistics, {ACL} 2018, Volume 1: Long Papers}, pages
2126--2136.
\bibitem[{Devlin et~al.(2019)Devlin, Chang, Lee, and
Toutanova}]{DBLP:conf/naacl/DevlinCLT19}
Jacob Devlin, Ming{-}Wei Chang, Kenton Lee, and Kristina Toutanova. 2019.
\newblock \href {https://doi.org/10.18653/v1/n19-1423} {{BERT:} pre-training of
deep bidirectional transformers for language understanding}.
\newblock In \emph{Proceedings of the 2019 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language
Technologies, {NAACL-HLT} 2019, Volume 1 (Long and Short Papers)}, pages
4171--4186.
\bibitem[{Dolan and Brockett(2005)}]{DBLP:conf/acl-iwp/DolanB05}
William~B. Dolan and Chris Brockett. 2005.
\newblock \href {https://aclanthology.org/I05-5002/} {Automatically
constructing a corpus of sentential paraphrases}.
\newblock In \emph{Proceedings of the Third International Workshop on
Paraphrasing, IWP@IJCNLP 2005}.
\bibitem[{Ethayarajh(2019)}]{DBLP:conf/emnlp/Ethayarajh19}
Kawin Ethayarajh. 2019.
\newblock \href {https://doi.org/10.18653/v1/D19-1006} {How contextual are
contextualized word representations? comparing the geometry of bert, elmo,
and {GPT-2} embeddings}.
\newblock In \emph{Proceedings of the 2019 Conference on Empirical Methods in
Natural Language Processing and the 9th International Joint Conference on
Natural Language Processing, {EMNLP-IJCNLP} 2019}, pages 55--65.
\bibitem[{Gabrielsson et~al.(2020)Gabrielsson, Nelson, Dwaraknath, and
Skraba}]{DBLP:conf/aistats/GabrielssonNDS20}
Rickard~Br{\"{u}}el Gabrielsson, Bradley~J. Nelson, Anjan Dwaraknath, and
Primoz Skraba. 2020.
\newblock \href {http://proceedings.mlr.press/v108/gabrielsson20a.html} {A
topology layer for machine learning}.
\newblock In \emph{The 23rd International Conference on Artificial Intelligence
and Statistics, {AISTATS} 2020}, volume 108 of \emph{Proceedings of Machine
Learning Research}, pages 1553--1563. {PMLR}.
\bibitem[{Gusfield(1997)}]{Gusfield:97}
Dan Gusfield. 1997.
\newblock \emph{Algorithms on Strings, Trees and Sequences}.
\newblock Cambridge University Press, Cambridge, UK.
\bibitem[{Hewitt et~al.(2021)Hewitt, Ethayarajh, Liang, and
Manning}]{DBLP:conf/emnlp/HewittELM21}
John Hewitt, Kawin Ethayarajh, Percy Liang, and Christopher~D. Manning. 2021.
\newblock \href {https://doi.org/10.18653/v1/2021.emnlp-main.122} {Conditional
probing: measuring usable information beyond a baseline}.
\newblock In \emph{Proceedings of the 2021 Conference on Empirical Methods in
Natural Language Processing, {EMNLP} 2021}, pages 1626--1639.
\bibitem[{Hewitt and Liang(2019)}]{DBLP:conf/emnlp/HewittL19}
John Hewitt and Percy Liang. 2019.
\newblock \href {https://doi.org/10.18653/v1/D19-1275} {Designing and
interpreting probes with control tasks}.
\newblock In \emph{Proceedings of the 2019 Conference on Empirical Methods in
Natural Language Processing and the 9th International Joint Conference on
Natural Language Processing, {EMNLP-IJCNLP} 2019}, pages 2733--2743.
\bibitem[{Hewitt and Manning(2019)}]{DBLP:conf/naacl/HewittM19}
John Hewitt and Christopher~D. Manning. 2019.
\newblock \href {https://doi.org/10.18653/v1/n19-1419} {A structural probe for
finding syntax in word representations}.
\newblock In \emph{Proceedings of the 2019 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language
Technologies, {NAACL-HLT} 2019, Volume 1 (Long and Short Papers)}, pages
4129--4138.
\bibitem[{Hua et~al.(2021)Hua, Li, Dou, Xu, and
Luo}]{DBLP:conf/naacl/HuaLDXL21}
Hang Hua, Xingjian Li, Dejing Dou, Cheng{-}Zhong Xu, and Jiebo Luo. 2021.
\newblock \href {https://doi.org/10.18653/v1/2021.naacl-main.258} {Noise
stability regularization for improving {BERT} fine-tuning}.
\newblock In \emph{Proceedings of the 2021 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language
Technologies, {NAACL-HLT} 2021}, pages 3229--3241.
\bibitem[{Hupkes and Zuidema(2018)}]{DBLP:conf/ijcai/HupkesZ18}
Dieuwke Hupkes and Willem~H. Zuidema. 2018.
\newblock \href {https://doi.org/10.24963/ijcai.2018/796} {Visualisation and
'diagnostic classifiers' reveal how recurrent and recursive neural networks
process hierarchical structure (extended abstract)}.
\newblock In \emph{Proceedings of the Twenty-Seventh International Joint
Conference on Artificial Intelligence, {IJCAI} 2018}, pages 5617--5621.
\bibitem[{Immer et~al.(2022)Immer, Hennigen, Fortuin, and
Cotterell}]{DBLP:conf/acl/ImmerHFC22}
Alexander Immer, Lucas~Torroba Hennigen, Vincent Fortuin, and Ryan Cotterell.
2022.
\newblock \href {https://doi.org/10.18653/v1/2022.acl-long.129} {Probing as
quantifying inductive bias}.
\newblock In \emph{Proceedings of the 60th Annual Meeting of the Association
for Computational Linguistics (Volume 1: Long Papers), {ACL} 2022}, pages
1839--1851.
\bibitem[{Ivanova et~al.(2021)Ivanova, Hewitt, and
Zaslavsky}]{DBLP:journals/corr/abs-2104-08197}
Anna~A. Ivanova, John Hewitt, and Noga Zaslavsky. 2021.
\newblock \href {http://arxiv.org/abs/2104.08197} {Probing artificial neural
networks: insights from neuroscience}.
\newblock \emph{CoRR}, abs/2104.08197.
\bibitem[{Jiang et~al.(2020)Jiang, He, Chen, Liu, Gao, and
Zhao}]{DBLP:conf/acl/JiangHCLGZ20}
Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Tuo
Zhao. 2020.
\newblock \href {https://doi.org/10.18653/v1/2020.acl-main.197} {{SMART:}
robust and efficient fine-tuning for pre-trained natural language models
through principled regularized optimization}.
\newblock In \emph{Proceedings of the 58th Annual Meeting of the Association
for Computational Linguistics, {ACL} 2020}, pages 2177--2190.
\bibitem[{Kornblith et~al.(2019)Kornblith, Norouzi, Lee, and
Hinton}]{DBLP:conf/icml/Kornblith0LH19}
Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey~E. Hinton. 2019.
\newblock \href {http://proceedings.mlr.press/v97/kornblith19a.html}
{Similarity of neural network representations revisited}.
\newblock In \emph{Proceedings of the 36th International Conference on Machine
Learning, {ICML} 2019}, volume~97 of \emph{Proceedings of Machine Learning
Research}, pages 3519--3529.
\bibitem[{Lee et~al.(2020)Lee, Cho, and Kang}]{DBLP:conf/iclr/LeeCK20}
Cheolhyoung Lee, Kyunghyun Cho, and Wanmo Kang. 2020.
\newblock \href {https://openreview.net/forum?id=HkgaETNtDB} {Mixout: Effective
regularization to finetune large-scale pretrained language models}.
\newblock In \emph{8th International Conference on Learning Representations,
{ICLR} 2020}.
\bibitem[{Lee et~al.(2022)Lee, Lee, Jang, and Yu}]{DBLP:conf/acl/LeeLJY22}
Seonghyeon Lee, Dongha Lee, Seongbo Jang, and Hwanjo Yu. 2022.
\newblock \href {https://doi.org/10.18653/v1/2022.acl-long.412} {Toward
interpretable semantic textual similarity via optimal transport-based
contrastive sentence learning}.
\newblock In \emph{Proceedings of the 60th Annual Meeting of the Association
for Computational Linguistics (Volume 1: Long Papers), {ACL} 2022}, pages
5969--5979.
\bibitem[{Limisiewicz and Marecek(2021)}]{DBLP:conf/acl/LimisiewiczM20}
Tomasz Limisiewicz and David Marecek. 2021.
\newblock \href {https://doi.org/10.18653/v1/2021.acl-long.36} {Introducing
orthogonal constraint in structural probes}.
\newblock In \emph{Proceedings of the 59th Annual Meeting of the Association
for Computational Linguistics and the 11th International Joint Conference on
Natural Language Processing, {ACL/IJCNLP} 2021, Volume 1: Long Papers}, pages
428--442.
\bibitem[{Liu et~al.(2019)Liu, Gardner, Belinkov, Peters, and
Smith}]{DBLP:conf/naacl/Liu0BPS19}
Nelson~F. Liu, Matt Gardner, Yonatan Belinkov, Matthew~E. Peters, and Noah~A.
Smith. 2019.
\newblock \href {https://doi.org/10.18653/v1/n19-1112} {Linguistic knowledge
and transferability of contextual representations}.
\newblock In \emph{Proceedings of the 2019 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language
Technologies, {NAACL-HLT} 2019, Volume 1 (Long and Short Papers)}, pages
1073--1094.
\bibitem[{Loshchilov and Hutter(2019)}]{DBLP:conf/iclr/LoshchilovH19}
Ilya Loshchilov and Frank Hutter. 2019.
\newblock \href {https://openreview.net/forum?id=Bkg6RiCqY7} {Decoupled weight
decay regularization}.
\newblock In \emph{7th International Conference on Learning Representations,
{ICLR} 2019}.
\bibitem[{Lovering et~al.(2021)Lovering, Jha, Linzen, and
Pavlick}]{DBLP:conf/iclr/LoveringJLP21}
Charles Lovering, Rohan Jha, Tal Linzen, and Ellie Pavlick. 2021.
\newblock \href {https://openreview.net/forum?id=mNtmhaDkAr} {Predicting
inductive biases of pre-trained models}.
\newblock In \emph{9th International Conference on Learning Representations,
{ICLR} 2021}.
\bibitem[{Marcus et~al.(1993)Marcus, Santorini, and
Marcinkiewicz}]{DBLP:journals/coling/MarcusSM94}
Mitchell~P. Marcus, Beatrice Santorini, and Mary~Ann Marcinkiewicz. 1993.
\newblock Building a large annotated corpus of english: The penn treebank.
\newblock \emph{Comput. Linguistics}, 19(2):313--330.
\bibitem[{Matthews(1975)}]{matthews1975comparison}
Brian~W Matthews. 1975.
\newblock Comparison of the predicted and observed secondary structure of t4
phage lysozyme.
\newblock \emph{Biochimica et Biophysica Acta (BBA)-Protein Structure},
405(2):442--451.
\bibitem[{Maudslay et~al.(2020)Maudslay, Valvoda, Pimentel, Williams, and
Cotterell}]{DBLP:conf/acl/MaudslayVPWC20}
Rowan~Hall Maudslay, Josef Valvoda, Tiago Pimentel, Adina Williams, and Ryan
Cotterell. 2020.
\newblock \href {https://doi.org/10.18653/v1/2020.acl-main.659} {A tale of a
probe and a parser}.
\newblock In \emph{Proceedings of the 58th Annual Meeting of the Association
for Computational Linguistics, {ACL} 2020}, pages 7389--7395.
\bibitem[{Pimentel and Cotterell(2021)}]{DBLP:conf/emnlp/PimentelC21}
Tiago Pimentel and Ryan Cotterell. 2021.
\newblock \href {https://doi.org/10.18653/v1/2021.emnlp-main.229} {A bayesian
framework for information-theoretic probing}.
\newblock In \emph{Proceedings of the 2021 Conference on Empirical Methods in
Natural Language Processing, {EMNLP} 2021}, pages 2869--2887.
\bibitem[{Pimentel et~al.(2020{\natexlab{a}})Pimentel, Saphra, Williams, and
Cotterell}]{DBLP:conf/emnlp/PimentelSWC20}
Tiago Pimentel, Naomi Saphra, Adina Williams, and Ryan Cotterell.
2020{\natexlab{a}}.
\newblock \href {https://doi.org/10.18653/v1/2020.emnlp-main.254} {Pareto
probing: Trading off accuracy for complexity}.
\newblock In \emph{Proceedings of the 2020 Conference on Empirical Methods in
Natural Language Processing, {EMNLP} 2020}, pages 3138--3153.
\bibitem[{Pimentel et~al.(2020{\natexlab{b}})Pimentel, Valvoda, Maudslay,
Zmigrod, Williams, and Cotterell}]{DBLP:conf/acl/PimentelVMZWC20}
Tiago Pimentel, Josef Valvoda, Rowan~Hall Maudslay, Ran Zmigrod, Adina
Williams, and Ryan Cotterell. 2020{\natexlab{b}}.
\newblock \href {https://doi.org/10.18653/v1/2020.acl-main.420}
{Information-theoretic probing for linguistic structure}.
\newblock In \emph{Proceedings of the 58th Annual Meeting of the Association
for Computational Linguistics, {ACL} 2020}, pages 4609--4622.
\bibitem[{Rasooli and Tetreault(2015)}]{rasooli-tetrault-2015}
Mohammad~Sadegh Rasooli and Joel~R. Tetreault. 2015.
\newblock \href {http://arxiv.org/abs/1503.06733} {Yara parser: {A} fast and
accurate dependency parser}.
\newblock \emph{Computing Research Repository}, arXiv:1503.06733.
\newblock Version 2.
\bibitem[{Reif et~al.(2019)Reif, Yuan, Wattenberg, Vi{\'{e}}gas, Coenen,
Pearce, and Kim}]{DBLP:conf/nips/ReifYWVCPK19}
Emily Reif, Ann Yuan, Martin Wattenberg, Fernanda~B. Vi{\'{e}}gas, Andy Coenen,
Adam Pearce, and Been Kim. 2019.
\newblock \href
{https://proceedings.neurips.cc/paper/2019/hash/159c1ffe5b61b41b3c4d8f4c2150f6c4-Abstract.html}
{Visualizing and measuring the geometry of {BERT}}.
\newblock In \emph{Advances in Neural Information Processing Systems 32: Annual
Conference on Neural Information Processing Systems 2019, NeurIPS 2019},
pages 8592--8600.
\bibitem[{Saphra and Lopez(2019)}]{DBLP:conf/naacl/SaphraL19}
Naomi Saphra and Adam Lopez. 2019.
\newblock \href {https://doi.org/10.18653/v1/n19-1329} {Understanding learning
dynamics of language models with {SVCCA}}.
\newblock In \emph{Proceedings of the 2019 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language
Technologies, {NAACL-HLT} 2019, Volume 1 (Long and Short Papers)}, pages
3257--3267.
\bibitem[{Vaswani et~al.(2017)Vaswani, Shazeer, Parmar, Uszkoreit, Jones,
Gomez, Kaiser, and Polosukhin}]{DBLP:conf/nips/VaswaniSPUJGKP17}
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
Aidan~N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017.
\newblock \href
{https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html}
{Attention is all you need}.
\newblock In \emph{Advances in Neural Information Processing Systems 30: Annual
Conference on Neural Information Processing Systems 2017}, pages 5998--6008.
\bibitem[{Voita et~al.(2019)Voita, Sennrich, and
Titov}]{DBLP:conf/emnlp/VoitaST19a}
Elena Voita, Rico Sennrich, and Ivan Titov. 2019.
\newblock \href {https://doi.org/10.18653/v1/D19-1448} {The bottom-up evolution
of representations in the transformer: {A} study with machine translation and
language modeling objectives}.
\newblock In \emph{Proceedings of the 2019 Conference on Empirical Methods in
Natural Language Processing and the 9th International Joint Conference on
Natural Language Processing, {EMNLP-IJCNLP} 2019}, pages 4395--4405.
\bibitem[{Voita and Titov(2020)}]{DBLP:conf/emnlp/VoitaT20}
Elena Voita and Ivan Titov. 2020.
\newblock \href {https://doi.org/10.18653/v1/2020.emnlp-main.14}
{Information-theoretic probing with minimum description length}.
\newblock In \emph{Proceedings of the 2020 Conference on Empirical Methods in
Natural Language Processing, {EMNLP} 2020}, pages 183--196.
\bibitem[{Wang et~al.(2019)Wang, Singh, Michael, Hill, Levy, and
Bowman}]{DBLP:conf/iclr/WangSMHLB19}
Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and
Samuel~R. Bowman. 2019.
\newblock \href {https://openreview.net/forum?id=rJ4km2R5t7} {{GLUE:} {A}
multi-task benchmark and analysis platform for natural language
understanding}.
\newblock In \emph{7th International Conference on Learning Representations,
{ICLR} 2019}.
\bibitem[{Wang et~al.(2022)Wang, Kuo, and Li}]{DBLP:conf/acl/WangK022}
Bin Wang, C.{-}C.~Jay Kuo, and Haizhou Li. 2022.
\newblock \href {https://doi.org/10.18653/v1/2022.acl-long.419} {Just rank:
Rethinking evaluation with word and sentence similarities}.
\newblock In \emph{Proceedings of the 60th Annual Meeting of the Association
for Computational Linguistics (Volume 1: Long Papers), {ACL} 2022}, pages
6060--6077.
\bibitem[{Warstadt et~al.(2019)Warstadt, Singh, and
Bowman}]{DBLP:journals/tacl/WarstadtSB19}
Alex Warstadt, Amanpreet Singh, and Samuel~R. Bowman. 2019.
\newblock \href {https://doi.org/10.1162/tacl\_a\_00290} {Neural network
acceptability judgments}.
\newblock \emph{Trans. Assoc. Comput. Linguistics}, 7:625--641.
\bibitem[{White et~al.(2021)White, Pimentel, Saphra, and
Cotterell}]{DBLP:conf/naacl/WhitePSC21}
Jennifer~C. White, Tiago Pimentel, Naomi Saphra, and Ryan Cotterell. 2021.
\newblock \href {https://doi.org/10.18653/v1/2021.naacl-main.12} {A non-linear
structural probe}.
\newblock In \emph{Proceedings of the 2021 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language
Technologies, {NAACL-HLT} 2021}, pages 132--138.
\bibitem[{Wolf et~al.(2019)Wolf, Debut, Sanh, Chaumond, Delangue, Moi, Cistac,
Rault, Louf, Funtowicz, and Brew}]{DBLP:journals/corr/abs-1910-03771}
Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue,
Anthony Moi, Pierric Cistac, Tim Rault, R{\'{e}}mi Louf, Morgan Funtowicz,
and Jamie Brew. 2019.
\newblock \href {http://arxiv.org/abs/1910.03771} {Huggingface's transformers:
State-of-the-art natural language processing}.
\newblock \emph{CoRR}, abs/1910.03771.
\bibitem[{Zeng et~al.(2022)Zeng, Jiang, Wu, Yin, and
Li}]{DBLP:conf/acl/ZengJWY022}
Jiali Zeng, Yufan Jiang, Shuangzhi Wu, Yongjing Yin, and Mu~Li. 2022.
\newblock \href {https://doi.org/10.18653/v1/2022.findings-acl.247}
{Task-guided disentangled tuning for pretrained language models}.
\newblock In \emph{Findings of the Association for Computational Linguistics:
{ACL} 2022}, pages 3126--3137.
\bibitem[{Zhang and Bowman(2018)}]{DBLP:conf/emnlp/ZhangB18}
Kelly~W. Zhang and Samuel~R. Bowman. 2018.
\newblock \href {https://doi.org/10.18653/v1/w18-5448} {Language modeling
teaches you more than translation does: Lessons learned through auxiliary
syntactic task analysis}.
\newblock In \emph{Proceedings of the Workshop: Analyzing and Interpreting
Neural Networks for NLP, BlackboxNLP@EMNLP 2018}, pages 359--361.
\bibitem[{Zhou et~al.(2022)Zhou, Ethayarajh, Card, and
Jurafsky}]{DBLP:conf/acl/ZhouECJ22}
Kaitlyn Zhou, Kawin Ethayarajh, Dallas Card, and Dan Jurafsky. 2022.
\newblock \href {https://doi.org/10.18653/v1/2022.acl-short.45} {Problems with
cosine as a measure of embedding similarity for high frequency words}.
\newblock In \emph{Proceedings of the 60th Annual Meeting of the Association
for Computational Linguistics (Volume 2: Short Papers), {ACL} 2022}, pages
401--423.
\bibitem[{Zhu et~al.(2020)Zhu, Cheng, Gan, Sun, Goldstein, and
Liu}]{DBLP:conf/iclr/ZhuCGSGL20}
Chen Zhu, Yu~Cheng, Zhe Gan, Siqi Sun, Tom Goldstein, and Jingjing Liu. 2020.
\newblock \href {https://openreview.net/forum?id=BygzbyHFvB} {Freelb: Enhanced
adversarial training for natural language understanding}.
\newblock In \emph{8th International Conference on Learning Representations,
{ICLR} 2020}.
\end{thebibliography}