The automatic verification of document authorships is important in various settings. Researchers are for example judged and compared by the amount and impact of their publications and public figures are confronted by their posts on social media platforms. Therefore, it is important that authorship information in frequently used web services and platforms is correct. The question whether a given document is written by a given author is commonly referred to as authorship verification (AV). While AV is a widely investigated problem in general, only few works consider settings where the documents are short and written in a rather uniform style. This makes most approaches unpractical for online databases and knowledge graphs in the scholarly domain. Here, authorships of scientific publications have to be verified, often with just abstracts and titles available. To this point, we present our novel approach LG4AV which combines language models and graph neural networks for authorship verification. By directly feeding the available texts in a pre-trained transformer architecture, our model does not need any hand-crafted stylometric features that are not meaningful in scenarios where the writing style is, at least to some extent, standardized. By the incorporation of a graph neural network structure, our model can benefit from relations between authors that are meaningful with respect to the verification process. For example, scientific authors are more likely to write about topics that are addressed by their co-authors and twitter users tend to post about the same subjects as people they follow. We experimentally evaluate our model and study to which extent the inclusion of co-authorships enhances verification decisions in bibliometric environments.READ FULL TEXT VIEW PDF
Authorship verification tries to answer the question if two documents wi...
We propose two models for a special case of authorship verification prob...
Twitter accounts have already been used in many scientometrics studies, ...
We propose an unsupervised solution to the Authorship Verification task ...
Author name disambiguation in bibliographic databases is the problem of
Authorship verification is the task of analyzing the linguistic patterns...
A central problem that has been researched for many years in the field o...
Evaluation of research strongly depends on bibliometric databases. Today, they are used for the assessment of productivity and impact of researchers, conferences and affiliations. This implies an increasing importance of search engines and web services that store, present and collect bibliometric data. Because of their rising relevance for the evaluation of the scientific output of individual authors, it is crucial that the information which is stored and collected by scholarly search engines, databases and knowledge graphs is complete and accurate. However, with the rapid growth of publication output (bornmann15), automatic inspections and corrections of information in bibliometric databases is needed. One of the major challenges in this area is authorship verification (AV), which aims to verify if a document is written by a specific author. In general, AV is widely investigated (koppel04; halteren04; halvani19). Available approaches range from the use of hand-crafted features which are based on lexical or syntactic patterns (hurlimann15; pothas20; jankowska14) to methods that make use of neural networks and language models (LMs) (bagnall15; barlas20).
A majority of existing work handles author verification by capturing writing styles (castro15; halvani19), assuming that they are unique among different authors. This assumption does not hold in environments where the available texts are short and contain uniform language patterns. An example of this is given by verification tasks for scientific documents. In such settings, the availability of full texts is rare because bibliometric databases often contain only abstracts and titles. In such scenarios the variety of writing styles and linguistic usage is rather limited. Hence, methods that are solely based on stylometric features are less promising.
Additionally, the focus in AV research is on documents with one author, while verification of multi-author documents is seldom done. In such scenarios, the information about known multi-authorships can enhance the verification process because it provides a graph structure between the authors which is meaningful with respect to the verification process. There are many scenarios where verification decisions can benefit from such graph structures. For example, scientific authors are more likely to write papers that would also fit to their co-authors and twitter users are expected to post about the same topics as the persons they follow. The incorporation of such graph structures is rarely investigated.
Most of the current approaches are based on an a setting which breaks the AV problem down to either “Are the documents and by the same author?” or “Is the unknown document from the same author as the set of known documents ?”. Here, is assumed to be of small cardinality. In the PAN@CLEF111https://pan.webis.de/ tasks on AV, which most approaches are focused on, the set of known documents for each unknown document was never larger then elements. Hence, the developed methods are in general not fitted to scenarios where larger sets of known documents are possible. For example, established researches can have hundred publications or more which can reflect a variety of different topics that they have worked on over time. Thus, a large amount of publications of these research can be relevant for the verification of potential papers. This makes approaches unfeasible which need to compare or combine the unknown document explicitly with the known documents.
Here we step in with LG4AV. Our novel architecture combines language models and graph neural networks (GNNs) to verify whether a document belongs to a potential author. This is done without the explicit recap of the known documents of this author at decision time which can be a bottleneck with respect to the computation time. This is especially true for authors with a large amount of known documents. Additionally, LG4AV does not rely on any hand-crafted stylometric features.
By incorporating a graph neural network structure into our architecture, we use known relations between potential authors to enhance the verification process. In this way, we are able to account for the fact that authors are more likely to turn to topics that are present in their social neighborhood. We experimentally evaluate the ability of our model to make verification decisions in bibliometric environments and we review the influence of the individual components on the quality of the verification decisions. Applications of LG4AV to other data sources where texts are connected with graph information, such as for example social networks, are possible. Our contribution is as follows.
We study authorship verification in a setting where information of known documents is only used at training time. An explicit recap of the known documents to verify authorships of specific authors with new documents is not needed.
We present LG4AV, a novel network architecture that incorporates language models and graph neural networks. LG4AV does not depend on any stylometric hand-crafted features and allows to incorporate meaningful relations between authors into the verification process.
We evaluate LG4AV in bibliometric environments and study how the different forms of information and the different parts of the module enhance the verification process.
We additionally investigate to which extent LG4AV is capable of verifying potential publications of authors that were not seen at training time.
LG4AV is available at https://github.com/mstubbemann/LG4AV.
Authorship verification is a commonly studied problem. PAN@CLEF provides regular competitions in this realm. However, their past author verification challenges were based in a setting where either small samples of up to ten known documents for each unknown document were provided (2013-2015) or pairs of documents were given where the task was to decide whether they were written by the same person (2020). Both scenarios are not applicable to our situation, where the amount of known documents of an author strongly varies and can reach up to hundreds.
Many well-established methods for author verification develop specific hand-crafted features that capture stylometric and syntactic patterns of documents. For example, (hurlimann15)
uses features such as sentence-lengths, punctuation marks and frequencies of n-grams to make verification decisions. While the use of n-grams was already studied in earlier works such as(kevselj03), there are still recent methods that build upon them, such as (pothas20). Here, the authors use them to determine similarities between known, unknown and external documents to compute verification scores. Another well established approach is given by (koppel07) where the authors successively remove features and observe how this reduces the distinction between two works. This approach is still known to be the gold standard (pothas20; bevendorff19). Despite its advantages, it is known to perform worse on short texts. Therefore, (bevendorff19) proposes a modification that is also applicable to shorter texts. However, in their work the authors experiment with documents of 4,000 words per document, which is still much longer then abstracts of scientific publications.
Recently, methods based on neural network architectures emerged. For example, (bagnall15)barlas20) proposes to replace the recurrent network by a pre-trained language model. Note, that both of these approaches need to train head layers for each individual author. This makes them unpractical for verification in bibliometric databases where information of thousands of authors has to be stored.
Author verification has been applied to many scenarios. While earlier studies often focused on books (koppel07; kevselj03), recent methods consider also online sources such as mails, social media data and forum postings (brocardo15; boenninghoff19-2; halvani16). However, author verification for bibliometric data is rarely done. One of the few works that experiments with bibliometric data is (halvani19), which uses full text of a small subset of authors and ignores co-authorship relations. Most other works that deal with bibliometric data tackle the closely related problem of authorship attribution (AA), i.e., with questions of the kind “who is the author of ” instead of “is author of ”. For example, (hill03; bradley08) do author verification of research papers only on citation information while (caragea19) also incorporates full texts.
Most of the work that combines graph information and text features in the realm of bibliometric data is restricted to the paper level (cohan20; hamilton17; kipf17). Even if there are few works that use text information and co-author connections, they tackle other problems then AV, such as link prediction or author clustering (ganesh16; wang21). One of the few works in the realm of AV that explicitly takes into account that research papers are multi-author documents is (sarwar20). In their work, the authors derive a similarity based graph structure of text fragments for authorship attribution for multi-author documents. In contrast, our aim is to incorporate past co-authorship relations to verify potential authorships.
In the realm of bibliometric data, many works also have focused on author disambiguation which has been tackled with the use of graph information (kang09; xu18) or text information (song07). However, in contrast to AV, author disambiguation does not tackle the problem of identifying authors of documents but the related problem of distinguishing authors with (nearly) equal names.
In the following, we formulate the problem we consider, introduce our architecture and discuss the individual components of LG4AV. For the rest of this work, we assume vectorsto be row-vectors. For a matrix , we denote by the -th row of and with the -th entry of the -th row of .
Let be a fixed time point and let be a graph with being a set of authors and a set of undirected and unweighted edges that represent connections until . Additionally, let be a set of documents. Let, for all authors , be the set of their known documents until . Let be a set of documents created after with unknown authorships. The goal is to verify for a set of potential author-document pairs, whether for each is an author of the unknown document . More formally, the aim is to find for each tuple a verification score .
Hence, we aim to infer from the data present at training time information about authors to verify potential documents of them that occur at testing time. Our formulation differs from the usual setting where the problem is either broken down to sequences of pairs where the task is to determine for each if the unknown document is from the same author as the set of known documents . These settings are closely connected in the sense that each author can be interpreted as the set of his known documents at training time. However, approaches adopted to this setting often assume to have already pairs of known sets and unknown document (hurlimann15) at training time or they explicitly use the set of known documents for verification (pothas20; jankowska14; seidman13). In contrast, we train a neural network on incorporating information of known documents of an author to make it possible to make verification decisions at testing time without explicitly recapping these documents for each unknown document. This is especially useful in settings where the amount of unknown documents can be large for some authors. Here, a need to explicitly use all these documents for every single verification of an unknown document can be a potential computational bottleneck.
We develop an end-to-end model to tackle the author verification problem. For this, we additionally assume to have for each author a vector representation . At training time, our model gets as input pairs and is trained on predicting whether is author of , i.e., . For inference at testing time, the model gets as input pairs and decides whether is author of by computing a verification score . For both training and testing, author-document pairs are forwarded through the network in the following manner. We add a special token to represent the current author to the beginning of . The resulting document is then guided through a language model. Additionally, we incorporate a graph neural network structure by 1.) computing vector representations of that depend on its graph neighbors, 2.) combining these vector representations with the output of a language model 3.) and forwarding through a fully connected layer to get a verification score. In the following, we discuss the individual components and give a detailed explanation of the network inference. A scratch of LG4AV is given by Figure 1.
We choose a neural-network based language model build upon the transformer architecture (vaswani17). More specifically, we choose the standard BERT (devlin19) model which we assume the reader to be familiar with. While this model is originally intended for sentences, it is possible to feed arbitrary sequences of specific maximum length (512 tokens for regular BERT models) through the network. Hence, we feed the full text of the document at once through BERT and extract the output of the first token (a specific classification token, denoted by [CLS]). This practice is established and has already been applied to abstracts of scientific documents (cohan20) as well as social media data (zhu19). To sum up, our language model can be interpreted as a map
To combine the output of the language model with the neighborhood aggregated author vector representation, we need to ensure that the output of the language model has the same size as the author features, i.e., . Hence,we extend by a linear layer on top of the BERT model, if needed.
To give the language model information about the current author , we replace the regular [CLS] token by an author-dependent classification token [CLS-]. Hence, the information of the current author is encoded into the input of the language model. Roughly speaking, if a pair is fed through the whole network at training time, the author information is not only incorporated into the
layers via backpropagation. Instead,also sees in the forward step. To nourish from the optimization of the [CLS] token that was done in the pre-training procedure of BERT, we initialize for each author the token-embedding of [CLS-] with the token-embedding of [CLS].
We use this paragraph to explain how we incorporate the co-author information and to discuss the benefits and underlying considerations of our approach.
The Graph Convolutional Network (GCN) (kipf17) is a base for many modern graph neural network architectures. It is a 2-layer neural network with an additional neighborhood aggregating step at the input and hidden layer. This method leads to problems in batch-processing. Since the feature vector of each node is merged with feature vectors of adjacent nodes at the input and hidden layers the corresponding node vectors have to be in the same batch. This problem is circumvented by (kipf17) since they only consider problems where full-batch training is possible, i.e., training on the whole feature matrix at once. Various approaches have been proposed to solve this problem. In (hamilton17) the authors enable mini-batch training by sampling the neighborhood of each node. Another approach is given by (chiang19) where the authors propose to cluster the graph into smaller subgraphs which build the batches for training. While these approaches allow to train neural networks with smaller batches, they are still not sufficient for our needs since each batch has to continue a reasonable amount of neighbors and non-neighbors. This makes it challenging to incorporate BERT which only allows for small batch sizes. Additionally, in our architecture the question arises how to combine author-vectors with output of the language model if the author is only part of the batch for neighborhood aggregation and not as part of a verification example.
To solve this problem we use a GNN architecture that only aggregates neighborhood information before weight matrices are multiplied with feature vectors. Then, the neighborhood-aggregation can be done once in a preprocessing step over the full feature-matrix. Such an approach is for example introduced by (wu19). Here, the authors propose a linear model of the form , with being a normalized adjacency matrix and a trainable weight matrix. In contrast, SIGN (rossi20) proposes to have multiple input layers with different neighborhood aggregations of the form
. The outputs of these layers are then concatenated, fed through an activation function and another fully-connected layer. Here,can for example be a power of a normalized adjacency matrix or powers of matrices that are based on triangles in the graph. While the model in (wu19) is more simple, it has the disadvantage that it only uses and not incorporates X itself. For LG4AV, we use a new model that is strongly inspired by both of the introduced approaches. Our complete architecture looks as follows.
The exact network inference is done in the following manner. Let be the feature matrix of all authors, i.e. . Let . Let be the adjacency matrix of with added self loops. For each , let be the degree of in G and be with and if . Let then be the normalized adjacency matrix. For a given pair of an author and a document the network inference is done in the following manner. For all , we concatenate the vectors to . Here, denotes the element-wise product. To derive a verification score from this concatenated vector, we feed into a fully connected layer with weight matrix , a bias and sigmoid activation. Hence, the full network inference of LG4AV is given by the equation
For training we use all pairs with as positive examples. For each we sample documents
to generate negative examples. We use binary cross entropy as loss function.
We experimentally evaluate to which extent LG4AV is able to verify potential authorships in bibliometric environments.
We use a data set which consists of publication information of AI research communities (koopmann21). It is based on data from DBLP222https://dblp.org/ and Semantic Scholar (ammar18). We use this data source because it contains titles and abstracts from Semantic Scholar which are needed for LG4AV but are not included in DBLP. However, the relations between authors and papers are based on DBLP which is, in our experience, comparably tidy and accurate. This is crucial for us since we want to prevent wrong authorship information in the data itself.
The data set consists of two subsets, 1.) the international AI researchers and all their publications, 2.) the German AI researchers and all their publications. We use the data set of the German AI researchers as our first data set. As a second data set, we extract all authors with publications at the KDD conference and all their publications from the data set of the international AI researchers. We refer to the first as the GAI and to the second as the KDD data.
To evaluate to which extent LG4AV is capable of handling the problem proposed in Section 3.1 we need to generate pairs of authors and potential co-authors with binary labels for training and testing. We generate this data for the GAI and the KDD data set in the following manner.
We discard all publications without an English abstract.
For each publication, we use the title and the abstract as input for the language model and concatenate them via a new line char to generate the text that represents this publication.
We build the co-author graph of all authors until 2015. From this co-author graph, we discard all author nodes that do not belong to the biggest connected component. Let be the set of authors that are nodes in this graph. We use this graph for the neighborhood aggregation. We denote the set of publications until 2015 of these authors with .
We generate for all authors and each of their publications a positive training example. For all authors, we then sample papers from which they are not an author of. We sample in such a way that we have for each author an equal amount of positive and negative examples.
We use data from the year for validation. More specifically, we use for all authors all publications that they have (co-) authored in 2016 as positive validation examples. Let be the set of these publications. We sample for all authors papers from that they are not author of as negative validation examples. Again, we sample in such a way that we have for each author an equal amount of positive and negative validation examples.
We use all publications from 2017 and newer to generate test examples. We create the test data analogously to the training examples and validation examples.
|# Authors||# Edges||# Train||# Validation||# Test|
Basic statistics of the resulting data can be found in Table 1. Note, that at validation and testing time, we only use the edges available at training time for neighborhood aggregation. Results for new co-authorships (and authors!) not seen at training time will be evaluated in Section 4.7.
We want to point out that we study AV in a balanced settings, i.e., with an equal amount of positive and negative examples. This is common practice in the realm of author verification (pothas20; halvani19; halvani16).
Since we study author verification in a novel manner where the examples are author-paper pairs instead of pairs of known documents and an unknown document, most of the existing AV approaches are not directly applicable.
There are existing methods that use language models and neural networks. However, they often deal with the related problem of authorship attribution (barlas20; gupta19), or they consider the problem whether pairs of documents are written by the same author (boeninghoff19).
For our experiments, we focus on baselines that can be easily adapted to author document pairs. For each baseline, we explain the needed transformation steps to fit it to our problem. Additionally, we briefly outline the function principles and the parameter choices.
This baseline is strongly inspired by the baseline script of the AV challenge of PAN@CLEF 2020. For all authors we generate a “superdocument” by concatenating all their documents that are available for training.
This baseline measures the cosine similarity between the known and unknown documents. More specifically, this baselines works as follows: For all pairs of authorsand documents at validation and testing time, we measure the similarity between the document and the superdocument of to decide whether is author of . If the similarity is above a given threshold , we classify the pair as a positive example. To measure similarities, we build a character-based -gram TF-IDF vectorizer upon all papers available at training time. Here, we only use the 3000 features that correspond to the most frequent -grams across the documents that the vectorizer is build on. We use this vectorizer to compute for each pair at validation or testing time a vector representation of the superdocument of and a vector representation of and then compute cosine-similarity.
We use the validation data to tune the parameter and the threshold . We tune via grid-search on and choose the value that corresponds to the highest AUC on the validation set. We tune the threshold on the set . Note, that this means to sample 1000 evenly spaced samples in . We also test the median of the distances of the validation examples as threshold. Since the AUC is independent of this threshold, we tune on the validation F1 score after the best is chosen.
This method is intended for pairs of the form where is a set of documents and is a single document. The question is whether is of the same author as . Note, that GLAD needs such pairs already for training. Since our data consists of pairs with an author and a document, we build training examples for GLAD in the following manner. Let be the set of all author-document pairs available at training time and let, for all , be the label of that pair. For each author we collect the set of positive and the set of negative training examples. For each , we feed GLAD with the training examples with a positive label and with a negative label. For validation and testing, we replace pairs by .
GLAD works as follows. For each pair a vector representation is computed that consists of features that are solely build on or , such as for example average sentence length and joint features which are build on and , such as entropy of concatenations of documents of and
. This vector representations are then fed into a support-vector machine. While the authors of(hurlimann15) uses a linear support-vector machine with default parameter setting of scikit-learn (pedregosa11), we enhance GLAD by tuning the -parameter of the support-vector machine and additionally experiment with radial kernels where we tune the -parameter. For both parameters, we grid search over . Again, we report the test results for the model with the best AUC on the validation data.
The ranking-based impostors method decides for a pair if they are from the same author with the help of a set of external documents. To use exactly the information available for training and thus have a fair comparison, we use as the known documents and as the external documents for each pair .
By studying for each how many documents of are closer to , the impostors method computes a verification score where pairs with higher scores are more likely to be positive examples. To compute vector representations for documents, we stick to the procedure in (pothas20). We choose the following parameters for RBI. We grid search , choose cosine similarity as the similarity function and select the aggregation function between mean, minimum and maximum function. For the meaning of the parameters, we refer to (pothas20). We use the AUC score on the validation data to choose the best parameters. To derive binary predictions from the verification scores, we use the median of all verification scores from the validation data. In contrast to GLAD, RBI does not need any pairs for a training procedure. Hence, verification scores can be directly computed for validation and testing data.
We implement LG4AV in PyTorch with Lightning and use the Hugging Face Transformers library for the language model. For our model, we stick to standard parameters in the context of GNNs and BERT. We train LG4AV for 3 epochs. After the element-wise multiplication of the BERT output and the text features, we dropout with probability of. We use the ADAM optimizer with weight decay of and a learning rate of with linear decay. We do not use warm-up steps as it hurts performance. We use as batch size and do gradient accumulation of for an effective batch size of . We set which is a common choice in the realm of GNNs. As BERT fine-tuning is known to be unstable (mosbach21; zhang21), we do runs for LG4AV and report mean scores.
To generate text features for each author, we feed all their documents through the not fine-tuned BERT model and build the mean point vector of the vector representations of the [CLS] tokens. At earlier stages, we also experimented with LSA (deerwester90) and doc2vec (le14) features and witnessed only negligible differences in performance.
In bibliometric settings, SciBert (beltagy19) would be the natural choice for the BERT model of LG4AV as it is trained on scientific text. However, since SciBert is trained on the Semantic Scholar Corpus which is also incorporated in the data sets used for our experiment, it can not be ruled out that SciBert is trained on texts included in our test data. Hence, to have a fair comparison to the baselines, we use the “regular” BERT base uncased model.
The results can be found in Table 2. For both data sets, LG4AV outperforms all baselines. Considering the baselines, N-Gram leads to the best results. It stands out that the scores for the KDD data sets are generally lower.
Our results indicate that LG4AV is comparably well suited for the task of verifying author-document pairs. Note, that for all authors their papers until 2015 were used to both build the positive training examples and for computing their feature vector. Hence, considering the positive examples, the network is trained on verifying author-document pairs where the document is additionally used to build the features of this author. Thus, it is reasonable to expect that the availability to generalize to unseen documents is potentially limited. However, the results on the test set, which contains only documents which where not used to build the features vectors of the authors, show that this is not the case.
While conducting our experiments, we noticed that validation scores are consistently higher then the test scores. We attribute this to our splitting procedure. The validation data is, in terms of time, “closer” to the training data then the test data. As the topics researchers work on change over time an increasing timely distance to the training data hinders the verification of new documents.
While the performance of LG4AV surpasses all baselines, we find that it is comparably unstable which is a well known problem for transformer models (mosbach21; zhang21). For some random seeds, LG4AV does not converge to a reasonable solution at all and results in AUC scores are around . Hence, we recommend to try different random seeds for training and not rely on one training run.
Lastly, the large gap between the performances on the two data sets is remarkable. One possible explanation is the fact that the GAI data is based on German researches from all domains of AI while the KDD data is limited to authors whose research interests overlaps with the topics of the KDD conference. Hence, it stands to reason that the documents of the KDD data set are more topically related. In consequence, the worse results on the KDD data support our hypothesis that author verification is mainly not about identifying writing style of unique authors but about identifying the relevant topics of authors by capturing important words and formulations.
|LG4AV-2||.9247333 Significantly outperforms LG4AV- with .||.8520||.8501||.85223||.76833||.76813|
The novelty of LG4AV lays in the connection of two components, namely a GNN and a fine-tunable language model, to perform author verification tasks. Hence, it is crucial to evaluate the individual components with respect to their influence on the verification decisions. In order to do so, we run the experiments in Section 4 with two additional LG4AV models.
This model coincides with the model used in Section 4 with the difference, that we freeze all BERT parameters and just train the weight matrix . This allows us to understand if the process of fine-tuning the BERT parameters is indeed necessary for successful author verification.
Here, the parameter is set to . Hence, this model does not use any graph information. It just uses the individual author features and does not include any neighborhood aggregation. We use this model to evaluate to which extent the graph information enhances the verification process.
The LG4AV model with which was also used in Table 2. This is our “regular” LG4AV which uses both BERT fine-tuning and neighborhood aggregation.
We also run rounds of LG4AV- and LG4AV-F and report mean scores. For this runs, we use the same 10 different random seeds for weight initialization and for shuffling the training data which were used for LG4AV-. We decided for this approach as early experiments indicated that the performances of LG4AV- and LG4AV- were better for the same random seeds (and therefore same shuffles of training data). This means, that the seeds which lead to the higher/lower values for LG4AV- generally also lead to better results for LG4AV-.
and the observation that both models have the tendency of having better/worse results for the same random seeds, we decide to use a paired t-test over the 10 runs for significance testing. On the KDD data set, LG4AV-outperforms LG4AV- with a significance level of on all metrics. On the GAI data set, LG4AV- outperforms LG4AV- with a significance level of with respect to the AUC score. Considering accuracy, LG4AV- outperform LG4AV- on the significance level . However, for the common value , it does not. There is no significant difference for the F1-score with respect to a reasonable p-value.
To sum up, the results indicate that the incorporation of co-author information can lead to additional enhancements. Still, the comparable results for show that our idea of combining text features with a language model even works without the addition of co-author information. On the other hand, if the BERT layers are frozen, the performance declines considerably. Hence, the fine-tuning of the language model is an integral point for the successful author verification with LG4AV.
Comparing the performance of LG4AV-F with the results in Table 2, the frozen models perform on line with GLAD and RBI, which shows that LG4AV-F is (to some extent) capable of successful author verification. We especially point that out because the freezing of the BERT layers significantly decreases the runtime since only the parameters of the top layer have to be trained. On a NVIDIA RTX 2060 SUPER, each epoch of LG4AV-F and LG4AV-2 on the GAI data last for about 40 minutes and for about 2 hour and 15 minutes, respectively. For the KDD data, one epoch needs about 1 hour and about 4 hours and 15 minutes, respectively. With limited resources available it would be a reasonable compromise to, for example, train the whole model for just one epoch and then freeze the layers of the language model.
Until now, we focused on verification cases for authors that were already seen at training time. However, as LG4AV takes as input examples an author represented by a feature-vector and a document text to decide whether this document fits to the feature-vector, it is possible to make verification decisions for (feature-vectors) of authors which where not seen at training time. We use this section to evaluate LG4AV on such not seen authors and with the use of known edges not available at training time.
For this, we assume to have an additional set of new authors and with additional known documents . Again let for be the set of known documents of this new author and let be a new set of documents of unknown authorship. Additionally we assume to have a new unweighted, undirected graph . Hence, we assume to have also edges between the new authors and the old authors. The goal is again, to make verification decisions for a set .
Note, that for each author seen at training time, we train a specific cls token [CLS-] which we place at the beginning of each document for each input pair . Hence, if we want to infer verification scores for pairs of new authors and potential documents, the question arises on how to choose the cls token [CLS-] as it is not trained. For this, we use the information of co-author relations to authors that were present at training time. More specific, we choose the cls token in the following manner.
Let be the set of neighbors of in that are in . Let for all be the vector representation of [CLS-]. For we set
To sum up, we set the set the embedding of the cls-token of a new authors to the mean point vector of the cls-tokens of its neighbors that were already present at training time. If there are no such old neighbors, we use the mean point vector of the embeddings of the cls-tokens of all old authors instead.
|# Authors||# Edges Back||Edges||# Test|
For both the GAI and the KDD data set, we choose the set of new authors in the following manner. For the GAI data set, we choose all authors, that have their first publication in our former validation time windows, i.e., in the year 2016. For the KDD data set, we choose all authors, that have a publication at the KDD conference and which have their first publication in 2016. For , we choose all co-author edges between authors of until 2016. Note, that this also includes new co-authorships between old authors that first arise in 2016.
For we choose the authored publications of 2016 of this author as the known documents . Let be the set of the documents of all new authors in 2016. For the RBI baseline, we sample for each a set of external documents has not authored from such that . The positive test examples are given by pairs where and is a document authored by after 2016. Let be the set of all documents of all authors after 2016, i.e., . For each author we sample documents that has not authored from . For each positive example for we sample one negative example. Statistics of the resulting data can be found in Table 4.
For all baselines we use the parameters and models for which we reported scores in Table 2 to infer the verification scores for the new test examples. For LG4AV, we report mean results over the 10 models used in Table 2 and Table 3.
The results of this experiment can be found in Table 5. On the GAI data set, LG4AV outperforms all baselines. On the KDD data set, GLAD leads to the best performance with respect to the AUC and accuracy score, followed by LG4AV. However, LG4AV still outperforms all baselines with respect to the F1-score. It stands out, that all methods except GLAD perform remarkably worse for unseen authors then in the experiment with the same author set for training and testing. This effect is remarkable strong for LG4AV on the KDD data set.
Having a look at Table 1, it stands out, that in the KDD data set a smaller ratio of new authors have connections to old authors () then for the GAI data (). Hence, it may be promising to study with more data sets if a general connection between the ratio of new authors with connection to old authors and the performance exists. If this is indeed the case, modifications of the generation of the author-dependent cls-token for new authors would be a potential starting point for suiting LG4AV especially to verification problems of authors not seen at training time.
In this work, we presented LG4AV, a novel architecture for author verification. By combining a language model with a graph neural network, our model does not depend on any handcrafted features and is able to incorporate relations between authors into the verification process. LG4AV surpasses methods that use handcrafted stylometric and n-gram text features when it comes to verification of short and, to some extent, standardized texts, as for example titles and abstracts of research papers. Hence, LG4AV is especially helpful to correct authorship information collected and stored by search engines and online data bases in the scholarly domain.
Future work could include applications of LG4AV to different settings, as for example authorship verification of tweets.
While LG4AV is generally designed for the verification of potential authorships of authors seen at training time, it also lead to comparable good results for verification cases on authors that were not seen at training time. However, as noted and discussed in Section 4.7, the performance drops in such scenarios and modifications of LG4AV to such tasks would be a useful direction for future research. Here, modifications of the incorporation of author information via the author-dependent cls-token could be promising.
From our observation that LG4AV and the baselines generally perform better on the validation then the testing data, we concluded that bigger time spans between training and inference data lead to worse results since the topics author consider continuously change over time. Building up on this, it would be interesting to study the influence of the discarding of old training examples. Does the performance increase if only recently published papers are considered to verify new potential authorships?