This is an internal, incomplete preview of a proposed change to the ACL Anthology.
For efficiency reasons, we generate only three BibTeX files per volume, and the preview may be incomplete in other ways, or contain mistakes.
Do not treat this content as an official publication.
Identifying important neurons for final predictions is essential for understanding the mechanisms of large language models. Due to computational constraints, current attribution techniques struggle to operate at neuron level. In this paper, we propose a static method for pinpointing significant neurons. Compared to seven other methods, our approach demonstrates superior performance across three metrics. Additionally, since most static methods typically only identify “value neurons” directly contributing to the final prediction, we propose a method for identifying “query neurons” which activate these “value neurons”. Finally, we apply our methods to analyze six types of knowledge across both attention and feed-forward network (FFN) layers. Our method and analysis are helpful for understanding the mechanisms of knowledge storage and set the stage for future research in knowledge editing. The code is available on https://github.com/zepingyu0512/neuron-attribution.
We investigate the mechanism of in-context learning (ICL) on sentence classification tasks with semantically-unrelated labels (“foo”/“bar”). We find intervening in only 1% heads (named “in-context heads”) significantly affects ICL accuracy from 87.6% to 24.4%. To understand this phenomenon, we analyze the value-output vectors in these heads and discover that the vectors at each label position contain substantial information about the corresponding labels. Furthermore, we observe that the prediction shift from “foo” to “bar” is due to the respective reduction and increase in these heads’ attention scores at “foo” and “bar” positions. Therefore, we propose a hypothesis for ICL: in in-context heads, the value-output matrices extract label features, while the query-key matrices compute the similarity between the features at the last position and those at each label position. The query and key matrices can be considered as two towers that learn the similarity metric between the last position’s features and each demonstration at label positions. Using this hypothesis, we explain the majority label bias and recency bias in ICL and propose two methods to reduce these biases by 22% and 17%, respectively.
We find arithmetic ability resides within a limited number of attention heads, with each head specializing in distinct operations. To delve into the reason, we introduce the Comparative Neuron Analysis (CNA) method, which identifies an internal logic chain consisting of four distinct stages from input to prediction: feature enhancing with shallow FFN neurons, feature transferring by shallow attention layers, feature predicting by arithmetic heads, and prediction enhancing among deep FFN neurons. Moreover, we identify the human-interpretable FFN neurons within both feature-enhancing and feature-predicting stages. These findings lead us to investigate the mechanism of LoRA, revealing that it enhances prediction probabilities by amplifying the coefficient scores of FFN neurons related to predictions. Finally, we apply our method in model pruning for arithmetic tasks and model editing for reducing gender bias. Code is on https://github.com/zepingyu0512/arithmetic-mechanism.
Maintaining factual consistency is a critical issue in abstractive text summarisation, however, it cannot be assessed by traditional automatic metrics used for evaluating text summarisation, such as ROUGE scoring. Recent efforts have been devoted to developing improved metrics for measuring factual consistency using pre-trained language models, but these metrics have restrictive token limits, and are therefore not suitable for evaluating long document text summarisation. Moreover, there is limited research and resources available for evaluating whether existing automatic evaluation metrics are fit for purpose when applied in long document settings. In this work, we evaluate the efficacy of automatic metrics for assessing the factual consistency of long document text summarisation. We create a human-annotated data set for evaluating automatic factuality metrics, LongSciVerify, which contains fine-grained factual consistency annotations for long document summaries from the scientific domain. We also propose a new evaluation framework, LongDocFACTScore, which is suitable for evaluating long document summarisation. This framework allows metrics to be efficiently extended to any length document and outperforms existing state-of-the-art metrics in its ability to correlate with human measures of factuality when used to evaluate long document summarisation data sets. We make our code and LongSciVerify data set publicly available: https://github.com/jbshp/LongDocFACTScore.
Automatically identifying genetic mutations in the cancer literature using text mining technology has been an important way to study the vast amount of cancer medical literature. However, novel knowledge regarding the genetic variants proliferates rapidly, though current supervised learning models struggle with discovering these unknown entity types. Few-shot learning allows a model to perform effectively with great generalization on new entity types, which has not been explored in recognizing cancer mutation detection. This paper addresses cancer mutation detection tasks with few-shot learning paradigms. We propose GDPN framework, which models the label dependency from the training examples in the support set and approximates the transition scores via Gaussian distribution. The experiments on three benchmark cancer mutation datasets show the effectiveness of our proposed model.
The goal of temporal relation extraction is to infer the temporal relation between two events in the document. Supervised models are dominant in this task. In this work, we investigate ChatGPT’s ability on zero-shot temporal relation extraction. We designed three different prompt techniques to break down the task and evaluate ChatGPT. Our experiments show that ChatGPT’s performance has a large gap with that of supervised methods and can heavily rely on the design of prompts. We further demonstrate that ChatGPT can infer more small relation classes correctly than supervised methods. The current shortcomings of ChatGPT on temporal relation extraction are also discussed in this paper. We found that ChatGPT cannot keep consistency during temporal inference and it fails in actively long-dependency temporal inference.
Early identification of depression is beneficial to public health surveillance and disease treatment. There are many models that mainly treat the detection as a binary classification task, such as detecting whether a user is depressed. However, identifying users’ depression severity levels from posts on social media is more clinically useful for future prevention and treatment. Existing severity detection methods mainly model the semantic information of posts while ignoring the relevant sentiment information, which can reflect the user’s state of mind and could be helpful for severity detection. In addition, they treat all severity levels equally, making the model difficult to distinguish between closely-labeled categories. We propose a sentiment-guided Transformer model, which efficiently fuses social media posts’ semantic information with sentiment information. Furthermore, we also utilize a supervised severity-aware contrastive learning framework to enable the model to better distinguish between different severity levels. The experimental results show that our model achieves superior performance on two public datasets, while further analysis proves the effectiveness of all proposed modules.
We propose a distantly supervised pipeline NER which executes entity span detection and entity classification in sequence named DISTANT (DIstantly Supervised enTity spAN deTection and classification).The former entity span detector extracts possible entity mention spans by the distant supervision. Then the later entity classifier assigns each entity span to one of the positive entity types or none by employing a positive and unlabeled (PU) learning framework. Two models were built based on the pre-trained SciBERT model and fine-tuned with the silver corpus generated by the distant supervision. Experimental results on BC5CDR and NCBI-Disease datasets show that our method outperforms the end-to-end NER baselines without PU learning by a large margin. In particular, it increases the recall score effectively.
This paper presents the results of the shared task on Lay Summarisation of Biomedical Research Articles (BioLaySumm), hosted at the BioNLP Workshop at ACL 2023. The goal of this shared task is to develop abstractive summarisation models capable of generating “lay summaries” (i.e., summaries that are comprehensible to non-technical audiences) in both a controllable and non-controllable setting. There are two subtasks: 1) Lay Summarisation, where the goal is for participants to build models for lay summary generation only, given the full article text and the corresponding abstract as input; and2) Readability-controlled Summarisation, where the goal is for participants to train models to generate both the technical abstract and the lay summary, given an article’s main text as input. In addition to overall results, we report on the setup and insights from the BioLaySumm shared task, which attracted a total of 20 participating teams across both subtasks.
Argument mining (AM) is a natural language processing task that aims to generate an argumentative graph given an unstructured argumentative text. An argumentative graph that consists of argumentative components and argumentative relations contains completed information of an argument and exhibits the logic of an argument. As the argument structure of an argumentative text can be regarded as an answer to a “why” question, the whole argument structure is therefore similar to the “chain of thought” concept, i.e., the sequence of ideas that lead to a specific conclusion for a given argument (Wei et al., 2022). For argumentative texts in the same specific genre, the “chain of thought” of such texts is usually similar, i.e., in a student essay, there is usually a major claim supported by several claims, and then a number of premises which are related to the claims are included (Eger et al., 2017). In this paper, we propose a new perspective which transfers the argument mining task into a multi-hop reading comprehension task, allowing the model to learn the argument structure as a “chain of thought”. We perform a comprehensive evaluation of our approach on two AM benchmarks and find that we surpass SOTA results. A detailed analysis shows that specifically the “chain of thought” information is helpful for the argument mining task.
The latest large language models (LLMs) such as ChatGPT, exhibit strong capabilities in automated mental health analysis. However, existing relevant studies bear several limitations, including inadequate evaluations, lack of prompting strategies, and ignorance of exploring LLMs for explainability. To bridge these gaps, we comprehensively evaluate the mental health analysis and emotional reasoning ability of LLMs on 11 datasets across 5 tasks. We explore the effects of different prompting strategies with unsupervised and distantly supervised emotional information. Based on these prompts, we explore LLMs for interpretable mental health analysis by instructing them to generate explanations for each of their decisions. We convey strict human evaluations to assess the quality of the generated explanations, leading to a novel dataset with 163 human-assessed explanations. We benchmark existing automatic evaluation metrics on this dataset to guide future related works. According to the results, ChatGPT shows strong in-context learning ability but still has a significant gap with advanced task-specific methods. Careful prompt engineering with emotional cues and expert-written few-shot examples can also effectively improve performance on mental health analysis. In addition, ChatGPT generates explanations that approach human performance, showing its great potential in explainable mental health analysis.
The information bottleneck (IB) principle has been proven effective in various NLP applications. The existing work, however, only used either generative or information compression models to improve the performance of the target task. In this paper, we propose to combine the two types of IB models into one system to enhance Named Entity Recognition (NER).For one type of IB model, we incorporate two unsupervised generative components, span reconstruction and synonym generation, into a span-based NER system. The span reconstruction ensures that the contextualised span representation keeps the span information, while the synonym generation makes synonyms have similar representations even in different contexts. For the other type of IB model, we add a supervised IB layer that performs information compression into the system to preserve useful features for NER in the resulting span representations. Experiments on five different corpora indicate that jointly training both generative and information compression models can enhance the performance of the baseline span-based NER system. Our source code is publicly available at https://github.com/nguyennth/joint-ib-models.
Rumour detection on social media is an important topic due to the challenges of misinformation propagation and slow verification of misleading information. Most previous work focus on the response posts on social media, ignoring the useful characteristics of involved users and their relations. In this paper, we propose a novel framework, Post-User Fusion Network (PESTO), which models the patterns of rumours from both post diffusion and user social networks. Specifically, we propose a novel Chronologically-masked Transformer architecture to model both temporal sequence and diffusion structure of rumours, and apply a Relational Graph Convolutional Network to model the social relations of involved users, with a fusion network based on self-attention mechanism to incorporate the two aspects. Additionally, two data augmentation techniques are leveraged to improve the robustness and accuracy of our models. Empirical results on four datasets of English tweets show the superiority of the proposed method.
We present a coherence-aware evaluation of document-level Text Simplification (TS), an approach that has not been considered in TS so far. We improve current TS sentence-based models to support a multi-sentence setting and the implementation of a state-of-the-art neural coherence model for simplification quality assessment. We enhanced English sentence simplification neural models for document-level simplification using 136,113 paragraph-level samples from both the general and medical domains to generate multiple sentences. Additionally, we use document-level simplification, readability and coherence metrics for evaluation. Our contributions include the introduction of coherence assessment into simplification evaluation with the automatic evaluation of 34,052 simplifications, a fine-tuned state-of-the-art model for document-level simplification, a coherence-based analysis of our results and a human evaluation of 300 samples that demonstrates the challenges encountered when moving towards document-level simplification.
Biomedical argument mining (BAM) aims at automatically identifying the argumentative structure in biomedical texts. However, identifying and classifying argumentative relations (AR) between argumentative components (AC) is challenging since it not only needs to understand the semantics of ACs but also need to capture the interactions between them. We argue that entities can serve as bridges that connect different ACs since entities and their mentions convey significant semantic information in biomedical argumentation. For example, it is common that related AC pairs share a common entity. Capturing such entity information can be beneficial for the Relation Identification (RI) task. In order to incorporate this entity information into BAM, we propose an Entity Coreference and Co-occurrence aware Argument Mining (ECCAM) framework based on an edge-oriented graph model for BAM. We evaluate our model on a benchmark dataset and from the experimental results we find that our method improves upon state-of-the-art methods.
We present PromptLS, a method for fine-tuning large pre-trained Language Models (LM) to perform the task of Lexical Simplification. We use a predefined template to attain appropriate replacements for a term, and fine-tune a LM using this template on language specific datasets. We filter candidate lists in post-processing to improve accuracy. We demonstrate that our model can work in a) a zero shot setting (where we only require a pre-trained LM), b) a fine-tuned setting (where language-specific data is required), and c) a multilingual setting (where the model is pre-trained across multiple languages and fine-tuned in an specific language). Experimental results show that, although the zero-shot setting is competitive, its performance is still far from the fine-tuned setting. Also, the multilingual is unsurprisingly worse than the fine-tuned model. Among all TSAR-2022 Shared Task participants, our team was ranked second in Spanish and third in English.
Cancer immunology research involves several important cell and protein factors. Extracting the information of such cells and proteins and the interactions between them from text are crucial in text mining for cancer immunology research. However, there are few available datasets for these entities, and the amount of annotated documents is not sufficient compared with other major named entity types. In this work, we introduce our automatically annotated dataset of key named entities, i.e., T-cells, cytokines, and transcription factors, which engages the recent cancer immunotherapy. The entities are annotated based on the UniProtKB knowledge base using dictionary matching. We build a neural named entity recognition (NER) model to be trained on this dataset and evaluate it on a manually-annotated data. Experimental results show that we can achieve a promising NER performance even though our data is automatically annotated. Our dataset also enhances the NER performance when combined with existing data, especially gaining improvement in yet investigated named entities such as cytokines and transcription factors.
Text summarization (TS) is an important NLP task. Pre-trained Language Models (PLMs) have been used to improve the performance of TS. However, PLMs are limited by their need of labelled training data and by their attention mechanism, which often makes them unsuitable for use on long documents. To this end, we propose a hybrid, unsupervised, abstractive-extractive approach, in which we walk through a document, generating salient textual fragments representing its key points. We then select the most important sentences of the document by choosing the most similar sentences to the generated texts, calculated using BERTScore. We evaluate the efficacy of generating and using salient textual fragments to guide extractive summarization on documents from the biomedical and general scientific domains. We compare the performance between long and short documents using different generative text models, which are finetuned to generate relevant queries or document titles. We show that our hybrid approach out-performs existing unsupervised methods, as well as state-of-the-art supervised methods, despite not needing a vast amount of labelled training data.
Negation and uncertainty modeling are long-standing tasks in natural language processing. Linguistic theory postulates that expressions of negation and uncertainty are semantically independent from each other and the content they modify. However, previous works on representation learning do not explicitly model this independence. We therefore attempt to disentangle the representations of negation, uncertainty, and content using a Variational Autoencoder. We find that simply supervising the latent representations results in good disentanglement, but auxiliary objectives based on adversarial learning and mutual information minimization can provide additional disentanglement gains.
Different from general documents, it is recognised that the ease with which people can understand a biomedical text is eminently varied, owing to the highly technical nature of biomedical documents and the variance of readers’ domain knowledge. However, existing biomedical document summarization systems have paid little attention to readability control, leaving users with summaries that are incompatible with their levels of expertise.In recognition of this urgent demand, we introduce a new task of readability controllable summarization for biomedical documents, which aims to recognise users’ readability demands and generate summaries that better suit their needs: technical summaries for experts and plain language summaries (PLS) for laymen.To establish this task, we construct a corpus consisting of biomedical papers with technical summaries and PLSs written by the authors, and benchmark multiple advanced controllable abstractive and extractive summarization models based on pre-trained language models (PLMs) with prevalent controlling and generation techniques.Moreover, we propose a novel masked language model (MLM) based metric and its variant to effectively evaluate the readability discrepancy between lay and technical summaries.Experimental results from automated and human evaluations show that though current control techniques allow for a certain degree of readability adjustment during generation, the performance of existing controllable summarization methods is far from desirable in this task.
We present a case study on the application of text classification and legal judgment prediction for flight compensation. We combine transformer-based classification models to classify responses from airlines and incorporate text data with other data types to predict a legal claim being successful. Our experimental evaluations show that our models achieve consistent and significant improvements over baselines and even outperformed human prediction when predicting a claim being successful. These models were integrated into an existing claim management system, providing substantial productivity gains for handling the case lifecycle, currently supporting several thousands of monthly processes.
The goal of text zoning is to segment a text into zones (i.e., Background, Conclusion) that serve distinct functions. Argumentative zoning, a specific text zoning scheme for the scientific domain, is considered as the antecedent for argument mining by many researchers. Surprisingly, however, little work is concerned with exploiting zoning information to improve the performance of argument mining models, despite the relatedness of the two tasks. In this paper, we propose two transformer-based models to incorporate zoning information into argumentative component identification and classification tasks. One model is for the sentence-level argument mining task and the other is for the token-level task. In particular, we add the zoning labels predicted by an off-the-shelf model to the beginning of each sentence, inspired by the convention commonly used biomedical abstracts. Moreover, we employ multi-head attention to transfer the sentence-level zoning information to each token in a sentence. Based on experiment results, we find a significant improvement in F1-scores for both sentence- and token-level tasks. It is worth mentioning that these zoning labels can be obtained with high accuracy by utilising readily available automated methods. Thus, existing argument mining models can be improved by incorporating zoning information without any additional annotation cost.
Several existing resources are available for sentiment analysis (SA) tasks that are used for learning sentiment specific embedding (SSE) representations. These resources are either large, common-sense knowledge graphs (KG) that cover a limited amount of polarities/emotions or they are smaller in size (e.g.: lexicons), which require costly human annotation and cover fine-grained emotions. Therefore using knowledge resources to learn SSE representations is either limited by the low coverage of polarities/emotions or the overall size of a resource. In this paper, we first introduce a new directed KG called ‘RELATE’, which is built to overcome both the issue of low coverage of emotions and the issue of scalability. RELATE is the first KG of its size to cover Ekman’s six basic emotions that are directed towards entities. It is based on linguistic rules to incorporate the benefit of semantics without relying on costly human annotation. The performance of ‘RELATE’ is evaluated by learning SSE representations using a Graph Convolutional Neural Network (GCN).
Recently, neural topic models (NTMs) have been incorporated into pre-trained language models (PLMs), to capture the global semantic information for text summarization. However, in these methods, there remain limitations in the way they capture and integrate the global semantic information. In this paper, we propose a novel model, the graph contrastive topic enhanced language model (GRETEL), that incorporates the graph contrastive topic model with the pre-trained language model, to fully leverage both the global and local contextual semantics for long document extractive summarization. To better capture and incorporate the global semantic information into PLMs, the graph contrastive topic model integrates the hierarchical transformer encoder and the graph contrastive learning to fuse the semantic information from the global document context and the gold summary. To this end, GRETEL encourages the model to efficiently extract salient sentences that are topically related to the gold summary, rather than redundant sentences that cover sub-optimal topics. Experimental results on both general domain and biomedical datasets demonstrate that our proposed method outperforms SOTA methods.
We propose a multi-task, probabilistic approach to facilitate distantly supervised relation extraction by bringing closer the representations of sentences that contain the same Knowledge Base pairs. To achieve this, we bias the latent space of sentences via a Variational Autoencoder (VAE) that is trained jointly with a relation classifier. The latent code guides the pair representations and influences sentence reconstruction. Experimental results on two datasets created via distant supervision indicate that multi-task learning results in performance benefits. Additional exploration of employing Knowledge Base priors into theVAE reveals that the sentence space can be shifted towards that of the Knowledge Base, offering interpretability and further improving results.
Emotion recognition (ER) is an important task in Natural Language Processing (NLP), due to its high impact in real-world applications from health and well-being to author profiling, consumer analysis and security. Current approaches to ER, mainly classify emotions independently without considering that emotions can co-exist. Such approaches overlook potential ambiguities, in which multiple emotions overlap. We propose a new model “SpanEmo” casting multi-label emotion classification as span-prediction, which can aid ER models to learn associations between labels and words in a sentence. Furthermore, we introduce a loss function focused on modelling multiple co-existing emotions in the input sentence. Experiments performed on the SemEval2018 multi-label emotion data over three language sets (i.e., English, Arabic and Spanish) demonstrate our method’s effectiveness. Finally, we present different analyses that illustrate the benefits of our method in terms of improving the model performance and learning meaningful associations between emotion classes and words in the sentence.
In this paper, we present Paladin, an open-source web-based annotation tool for creating high-quality multi-label document-level datasets. By integrating active learning and proactive learning to the annotation task, Paladin makes the task less time-consuming and requiring less human effort. Although Paladin is designed for multi-label settings, the system is flexible and can be adapted to other tasks in single-label settings.
Previous neural Seq2Seq models have shown the effectiveness for jointly extracting relation triplets. However, most of these models suffer from incompletion and disorder problems when they extract multi-token entities from input sentences. To tackle these problems, we propose a generative, multi-task learning framework, named GenerativeRE. We firstly propose a special entity labelling method on both input and output sequences. During the training stage, GenerativeRE fine-tunes the pre-trained generative model and learns the special entity labels simultaneously. During the inference stage, we propose a novel copy mechanism equipped with three mask strategies, to generate the most probable tokens by diminishing the scope of the model decoder. Experimental results show that our model achieves 4.6% and 0.9% F1 score improvements over the current state-of-the-art methods in the NYT24 and NYT29 benchmark datasets respectively.
Coreference resolution is the task of identifying all mentions in a text that refer to the same real-world entity. Collecting sufficient labelled data from expert annotators to train a high-performance coreference resolution system is time-consuming and expensive. Crowdsourcing makes it possible to obtain the required amounts of data rapidly and cost-effectively. However, crowd-sourced labels can be noisy. To ensure high-quality data, it is crucial to infer the correct labels by aggregating the noisy labels. In this paper, we split the aggregation into two subtasks, i.e, mention classification and coreference chain inference. Firstly, we predict the general class of each mention using an autoencoder, which incorporates contextual information about each mention, while at the same time taking into account the mention’s annotation complexity and annotators’ reliability at different levels. Secondly, to determine the coreference chain of each mention, we use weighted voting which takes into account the learned reliability in the first subtask. Experimental results demonstrate the effectiveness of our method in predicting the correct labels. We also illustrate our model’s interpretability through a comprehensive analysis of experimental results.
Risk management is a vital activity to ensure employee safety in construction projects. Various documents provide important supporting evidence, including details of previous incidents, consequences and mitigation strategies. Potential hazards may depend on a complex set of project-specific attributes, including activities undertaken, location, equipment used, etc. However, finding evidence about previous projects with similar attributes can be problematic, since information about risks and mitigations is usually hidden within and may be dispersed across a range of different free text documents. Automatic named entity recognition (NER), which identifies mentions of concepts in free text documents, is the first stage in structuring knowledge contained within them. While developing NER methods generally relies on annotated corpora, we are not aware of any such corpus targeted at concepts relevant to construction safety. In response, we have designed a novel named entity annotation scheme and associated guidelines for this domain, which covers hazards, consequences, mitigation strategies and project attributes. Four health and safety experts used the guidelines to annotate a total of 600 sentences from accident reports; an average inter-annotator agreement rate of 0.79 F-Score shows that our work constitutes an important first step towards developing tools for detailed semantic analysis of construction safety documents.
Unsupervised relation extraction (URE) extracts relations between named entities from raw text without manually-labelled data and existing knowledge bases (KBs). URE methods can be categorised into generative and discriminative approaches, which rely either on hand-crafted features or surface form. However, we demonstrate that by using only named entities to induce relation types, we can outperform existing methods on two popular datasets. We conduct a comparison and evaluation of our findings with other URE techniques, to ascertain the important features in URE. We conclude that entity types provide a strong inductive bias for URE.
Inter-sentence relation extraction deals with a number of complex semantic relationships in documents, which require local, non-local, syntactic and semantic dependencies. Existing methods do not fully exploit such dependencies. We present a novel inter-sentence relation extraction model that builds a labelled edge graph convolutional neural network model on a document-level graph. The graph is constructed using various inter- and intra-sentence dependencies to capture local and non-local dependency information. In order to predict the relation of an entity pair, we utilise multi-instance learning with bi-affine pairwise scoring. Experimental results show that our model achieves comparable performance to the state-of-the-art neural models on two biochemistry datasets. Our analysis shows that all the types in the graph are effective for inter-sentence relation extraction.
When constructing models that learn from noisy labels produced by multiple annotators, it is important to accurately estimate the reliability of annotators. Annotators may provide labels of inconsistent quality due to their varying expertise and reliability in a domain. Previous studies have mostly focused on estimating each annotator’s overall reliability on the entire annotation task. However, in practice, the reliability of an annotator may depend on each specific instance. Only a limited number of studies have investigated modelling per-instance reliability and these only considered binary labels. In this paper, we propose an unsupervised model which can handle both binary and multi-class labels. It can automatically estimate the per-instance reliability of each annotator and the correct label for each instance. We specify our model as a probabilistic model which incorporates neural networks to model the dependency between latent variables and instances. For evaluation, the proposed method is applied to both synthetic and real data, including two labelling tasks: text classification and textual entailment. Experimental results demonstrate our novel method can not only accurately estimate the reliability of annotators across different instances, but also achieve superior performance in predicting the correct labels and detecting the least reliable annotators compared to state-of-the-art baselines.
We tackle the nested and overlapping event detection task and propose a novel search-based neural network (SBNN) structured prediction model that treats the task as a search problem on a relation graph of trigger-argument structures. Unlike existing structured prediction tasks such as dependency parsing, the task targets to detect DAG structures, which constitute events, from the relation graph. We define actions to construct events and use all the beams in a beam search to detect all event structures that may be overlapping and nested. The search process constructs events in a bottom-up manner while modelling the global properties for nested and overlapping structures simultaneously using neural networks. We show that the model achieves performance comparable to the state-of-the-art model Turku Event Extraction System (TEES) on the BioNLP Cancer Genetics (CG) Shared Task 2013 without the use of any syntactic and hand-engineered features. Further analyses on the development set show that our model is more computationally efficient while yielding higher F1-score performance.
Document-level relation extraction is a complex human process that requires logical inference to extract relationships between named entities in text. Existing approaches use graph-based neural models with words as nodes and edges as relations between them, to encode relations across sentences. These models are node-based, i.e., they form pair representations based solely on the two target node representations. However, entity relations can be better expressed through unique edge representations formed as paths between nodes. We thus propose an edge-oriented graph neural model for document-level relation extraction. The model utilises different types of nodes and edges to create a document-level graph. An inference mechanism on the graph edges enables to learn intra- and inter-sentence relations using multi-instance learning internally. Experiments on two document-level biomedical datasets for chemical-disease and gene-disease associations show the usefulness of the proposed edge-oriented approach.
This paper describes our system developed for the coreference resolution task of the CRAFT Shared Tasks 2019. The CRAFT corpus is more challenging than other existing corpora because it contains full text articles. We have employed an existing span-based state-of-theart neural coreference resolution system as a baseline system. We enhance the system with two different techniques to capture longdistance coreferent pairs. Firstly, we filter noisy mentions based on parse trees with increasing the number of antecedent candidates. Secondly, instead of relying on the LSTMs, we integrate the highly expressive language model–BERT into our model. Experimental results show that our proposed systems significantly outperform the baseline. The best performing system obtained F-scores of 44%, 48%, 39%, 49%, 40%, and 57% on the test set with B3, BLANC, CEAFE, CEAFM, LEA, and MUC metrics, respectively. Additionally, the proposed model is able to detect coreferent pairs in long distances, even with a distance of more than 200 sentences.
The availability of large-scale and real-time data on social media has motivated research into adverse drug reactions (ADRs). ADR classification helps to identify negative effects of drugs, which can guide health professionals and pharmaceutical companies in making medications safer and advocating patients’ safety. Based on the observation that in social media, negative sentiment is frequently expressed towards ADRs, this study presents a neural model that combines sentiment analysis with transfer learning techniques to improve ADR detection in social media postings. Our system is firstly trained to classify sentiment in tweets concerning current affairs, using the SemEval17-task4A corpus. We then apply transfer learning to adapt the model to the task of detecting ADRs in social media postings. We show that, in combination with rich representations of words and their contexts, transfer learning is beneficial, especially given the large degree of vocabulary overlap between the current affairs posts in the SemEval17-task4A corpus and posts about ADRs. We compare our results with previous approaches, and show that our model can outperform them by up to 3% F-score.
Entity mentions embedded in longer entity mentions are referred to as nested entities. Most named entity recognition (NER) systems deal only with the flat entities and ignore the inner nested ones, which fails to capture finer-grained semantic information in underlying texts. To address this issue, we propose a novel neural model to identify nested entities by dynamically stacking flat NER layers. Each flat NER layer is based on the state-of-the-art flat NER model that captures sequential context representation with bidirectional Long Short-Term Memory (LSTM) layer and feeds it to the cascaded CRF layer. Our model merges the output of the LSTM layer in the current flat NER layer to build new representation for detected entities and subsequently feeds them into the next flat NER layer. This allows our model to extract outer entities by taking full advantage of information encoded in their corresponding inner entities, in an inside-to-outside way. Our model dynamically stacks the flat NER layers until no outer entities are extracted. Extensive evaluation shows that our dynamic model outperforms state-of-the-art feature-based systems on nested NER, achieving 74.7% and 72.2% on GENIA and ACE2005 datasets, respectively, in terms of F-score.
We present a novel graph-based neural network model for relation extraction. Our model treats multiple pairs in a sentence simultaneously and considers interactions among them. All the entities in a sentence are placed as nodes in a fully-connected graph structure. The edges are represented with position-aware contexts around the entity pairs. In order to consider different relation paths between two entities, we construct up to l-length walks between each pair. The resulting walks are merged and iteratively used to update the edge representations into longer walks representations. We show that the model achieves performance comparable to the state-of-the-art systems on the ACE 2005 dataset without using any external tools.
Currently, news articles are produced, shared and consumed at an extremely rapid rate. Although their quantity is increasing, at the same time, their quality and trustworthiness is becoming fuzzier. Hence, it is important not only to automate information extraction but also to quantify the certainty of this information. Automated identification of certainty has been studied both in the scientific and newswire domains, but performance is considerably higher in tasks focusing on scientific text. We compare the differences in the definition and expression of uncertainty between a scientific domain, i.e., biomedicine, and newswire. We delve into the different aspects that affect the certainty of an extracted event in a news article and examine whether they can be easily identified by techniques already validated in the biomedical domain. Finally, we present a comparison of the syntactic and lexical differences between the the expression of certainty in the biomedical and newswire domains, using two annotated corpora.
Existing biomedical coreference resolution systems depend on features and/or rules based on syntactic parsers. In this paper, we investigate the utility of the state-of-the-art general domain neural coreference resolution system on biomedical texts. The system is an end-to-end system without depending on any syntactic parsers. We also investigate the domain specific features to enhance the system for biomedical texts. Experimental results on the BioNLP Protein Coreference dataset and the CRAFT corpus show that, with no parser information, the adapted system compared favorably with the systems that depend on parser information on these datasets, achieving 51.23% on the BioNLP dataset and 36.33% on the CRAFT corpus in F1 score. In-domain embeddings and domain-specific features helped improve the performance on the BioNLP dataset, but they did not on the CRAFT corpus.
In this paper, we present APLenty, an annotation tool for creating high-quality sequence labeling datasets using active and proactive learning. A major innovation of our tool is the integration of automatic annotation with active learning and proactive learning. This makes the task of creating labeled datasets easier, less time-consuming and requiring less human effort. APLenty is highly flexible and can be adapted to various other tasks.
Descriptive document clustering aims to automatically discover groups of semantically related documents and to assign a meaningful label to characterise the content of each cluster. In this paper, we present a descriptive clustering approach that employs a distributed representation model, namely the paragraph vector model, to capture semantic similarities between documents and phrases. The proposed method uses a joint representation of phrases and documents (i.e., a co-embedding) to automatically select a descriptive phrase that best represents each document cluster. We evaluate our method by comparing its performance to an existing state-of-the-art descriptive clustering method that also uses co-embedding but relies on a bag-of-words representation. Results obtained on benchmark datasets demonstrate that the paragraph vector-based method obtains superior performance over the existing approach in both identifying clusters and assigning appropriate descriptive labels to them.
The goal of active learning is to minimise the cost of producing an annotated dataset, in which annotators are assumed to be perfect, i.e., they always choose the correct labels. However, in practice, annotators are not infallible, and they are likely to assign incorrect labels to some instances. Proactive learning is a generalisation of active learning that can model different kinds of annotators. Although proactive learning has been applied to certain labelling tasks, such as text classification, there is little work on its application to named entity (NE) tagging. In this paper, we propose a proactive learning method for producing NE annotated corpora, using two annotators with different levels of expertise, and who charge different amounts based on their levels of experience. To optimise both cost and annotation quality, we also propose a mechanism to present multiple sentences to annotators at each iteration. Experimental results for several corpora show that our method facilitates the construction of high-quality NE labelled datasets at minimal cost.
Named entity recognition (NER) in social media (e.g., Twitter) is a challenging task due to the noisy nature of text. As part of our participation in the W-NUT 2016 Named Entity Recognition Shared Task, we proposed an unsupervised learning approach using deep neural networks and leverage a knowledge base (i.e., DBpedia) to bootstrap sparse entity types with weakly labelled data. To further boost the performance, we employed a more sophisticated tagging scheme and applied dropout as a regularisation technique in order to reduce overfitting. Even without hand-crafting linguistic features nor leveraging any of the W-NUT-provided gazetteers, we obtained robust performance with our approach, which ranked third amongst all shared task participants according to the official evaluation on a gold standard named entity-annotated corpus of 3,856 tweets.
Classifying research grants into useful categories is a vital task for a funding body to give structure to the portfolio for analysis, informing strategic planning and decision-making. Automating this classification process would save time and effort, providing the accuracy of the classifications is maintained. We employ five classification models to classify a set of BBSRC-funded research grants in 21 research topics based on unigrams, technical terms and Latent Dirichlet Allocation models. To boost precision, we investigate methods for combining their predictions into five aggregate classifiers. Evaluation confirmed that ensemble classification models lead to higher precision. It was observed that there is not a single best-performing aggregate method for all research topics. Instead, the best-performing method for a research topic depends on the number of positive training instances available for this topic. Subject matter experts considered the predictions of aggregate models to correct erroneous or incomplete manual assignments.
Assessing the suitability of an Open Source Software project for adoption requires not only an analysis of aspects related to the code, such as code quality, frequency of updates and new version releases, but also an evaluation of the quality of support offered in related online forums and issue trackers. Understanding the content types of forum messages and issue trackers can provide information about the extent to which requests are being addressed and issues are being resolved, the percentage of issues that are not being fixed, the cases where the user acknowledged that the issue was successfully resolved, etc. These indicators can provide potential adopters of the OSS with estimates about the level of available support. We present a detailed hierarchy of content types of online forum messages and issue tracker comments and a corpus of messages annotated accordingly. We discuss our experiments to classify forum messages and issue tracker comments into content-related classes, i.e.~to assign them to nodes of the hierarchy. The results are very encouraging.
The process of annotating text corpora involves establishing annotation schemata which define the scope and depth of an annotation task at hand. We demonstrate this activity in Argo, a Web-based workbench for the analysis of textual resources, which facilitates both automatic and manual annotation. Annotation tasks in the workbench are defined by building workflows consisting of a selection of available elementary analytics developed in compliance with the Unstructured Information Management Architecture specification. The architecture accommodates complex annotation types that may define primitive as well as referential attributes. Argo aids the development of custom annotation schemata and supports their interoperability by featuring a schema editor and specialised analytics for schemata alignment. The schema editor is a self-contained graphical user interface for defining annotation types. Multiple heterogeneous schemata can be aligned by including one of two type mapping analytics currently offered in Argo. One is based on a simple mapping syntax and, although limited in functionality, covers most common use cases. The other utilises a well established graph query language, SPARQL, and is superior to other state-of-the-art solutions in terms of expressiveness. We argue that the customisation of annotation schemata does not need to compromise their interoperability.
Causality lies at the heart of biomedical knowledge, being involved in diagnosis, pathology or systems biology. Thus, automatic causality recognition can greatly reduce the human workload by suggesting possible causal connections and aiding in the curation of pathway models. For this, we rely on corpora that are annotated with classified, structured representations of important facts and findings contained within text. However, it is impossible to correctly interpret these annotations without additional information, e.g., classification of an event as fact, hypothesis, experimental result or analysis of results, confidence of authors about the validity of their analyses etc. In this study, we analyse and automatically detect this type of information, collectively termed meta-knowledge (MK), in the context of existing discourse causality annotations. Our effort proves the feasibility of identifying such pieces of information, without which the understanding of causal relations is limited.
This article provides an overview of the dissemination work carried out in META-NET from 2010 until early 2014; we describe its impact on the regional, national and international level, mainly with regard to politics and the situation of funding for LT topics. This paper documents the initiatives work throughout Europe in order to boost progress and innovation in our field.
As a first step towards assessing the quality of support offered online for Open Source Software (OSS), we address the task of locating requests, i.e., messages that raise an issue to be addressed by the OSS community, as opposed to any other message. We present a corpus of online communication messages randomly sampled from newsgroups and bug trackers, manually annotated as requests or non-requests. We identify several linguistically shallow, content-based heuristics that correlate with the classification and investigate the extent to which they can serve as independent classification criteria. Then, we train machine-learning classifiers on these heuristics. We experiment with a wide range of settings, such as different learners, excluding some heuristics and adding unigram features of various parts-of-speech and frequency. We conclude that some heuristics can perform well, while their accuracy can be improved further using machine learning, at the cost of obtaining manual annotations.
Cross-lingual information retrieval (CLIR) involving the Chinese language has been thoroughly studied in the general language domain, but rarely in the biomedical domain, due to the lack of suitable linguistic resources and parsing tools. In this paper, we describe a Chinese-English CLIR system for biomedical literature, which exploits a bilingual ontology, the ``eCMeSH Tree"""". This is an extension of the Chinese Medical Subject Headings (CMeSH) Tree, based on Medical Subject Headings (MeSH). Using the 2006 and 2007 TREC Genomics track data, we have evaluated the performance of the eCMeSH Tree in expanding queries. We have compared our results to those obtained using two other approaches, i.e. pseudo-relevance feedback (PRF) and document translation (DT). Subsequently, we evaluate the performance of different combinations of these three retrieval methods. Our results show that our method of expanding queries using the eCMeSH Tree can outperform the PRF method. Furthermore, combining this method with PRF and DT helps to smooth the differences in query expansion, and consequently results in the best performance amongst all experiments reported. All experiments compare the use of two different retrieval models, i.e. Okapi BM25 and a query likelihood language model. In general, the former performs slightly better.
Due to the rapid growth in the volume of biomedical literature, there is an increasing requirement for high-performance semantic search systems, which allow biologists to perform precise searches for events of interest. Such systems are usually trained on corpora of documents that contain manually annotated events. Until recently, these corpora, and hence the event extraction systems trained on them, focussed almost exclusively on the identification and classification of event arguments, without taking into account how the textual context of the events could affect their interpretation. Previously, we designed an annotation scheme to enrich events with several aspects (or dimensions) of interpretation, which we term meta-knowledge, and applied this scheme to the entire GENIA corpus. In this paper, we report on our experiments to automate the assignment of one of these meta-knowledge dimensions, i.e. Manner, to recognised events. Manner is concerned with the rate, strength intensity or level of the event. We distinguish three different values of manner, i.e., High, Low and Neutral. To our knowledge, our work represents the first attempt to classify the manner of events. Using a combination of lexical, syntactic and semantic features, our system achieves an overall accuracy of 99.4%.
Challenges in creating comprehensive text-processing worklows include a lack of the interoperability of individual components coming from different providers and/or a requirement imposed on the end users to know programming techniques to compose such workflows. In this paper we demonstrate Argo, a web-based system that addresses these issues in several ways. It supports the widely adopted Unstructured Information Management Architecture (UIMA), which handles the problem of interoperability; it provides a web browser-based interface for developing workflows by drawing diagrams composed of a selection of available processing components; and it provides novel user-interactive analytics such as the annotation editor which constitutes a bridge between automatic processing and manual correction. These features extend the target audience of Argo to users with a limited or no technical background. Here, we focus specifically on the construction of advanced workflows, involving multiple branching and merging points, to facilitate various comparative evalutions. Together with the use of user-collaboration capabilities supported in Argo, we demonstrate several use cases including visual inspections, comparisions of multiple processing segments or complete solutions against a reference standard, inter-annotator agreement, and shared task mass evaluations. Ultimetely, Argo emerges as a one-stop workbench for defining, processing, editing and evaluating text processing tasks.
The analysis of a corpus of micro-blogs on the topic of the 2011 UK referendum about the Alternative Vote has been undertaken as a joint activity by text miners and social scientists. To facilitate the collaboration, the corpus and its analysis is managed in a Web-accessible framework that allows users to upload their own textual data for analysis and to manage their own text annotation resources used for analysis. The framework also allows annotations to be searched, and the analysis to be re-run after amending the analysis resources. The corpus is also doubly human-annotated stating both whether each tweet is overall positive or negative in sentiment and whether it is for or against the proposition of the referendum.
In this paper, we present the main features of a text mining based search engine for the UK Educational Evidence Portal available at the UK National Centre for Text Mining (NaCTeM), together with a user-centred framework for the evaluation of the search engine. The framework is adapted from an existing proposal by the ISLE (EAGLES) Evaluation Working group. We introduce the metrics employed for the evaluation, and explain how these relate to the text mining based search engine. Following this, we describe how we applied the framework to the evaluation of a number of key text mining features of the search engine, namely the automatic clustering of search results, classification of search results according to a taxonomy, and identification of topics and other documents that are related to a chosen document. Finally, we present the results of the evaluation in terms of the strengths, weaknesses and improvements identified for each of these features.
Biomedical corpora annotated with event-level information provide an important resource for the training of domain-specific information extraction (IE) systems. These corpora concentrate primarily on creating classified, structured representations of important facts and findings contained within the text. However, bio-event annotations often do not take into account additional information (meta-knowledge) that is expressed within the textual context of the bio-event, e.g., the pragmatic/rhetorical intent and the level of certainty ascribed to a particular bio-event by the authors. Such additional information is indispensible for correct interpretation of bio-events. Therefore, an IE system that simply presents a list of bare bio-events, without information concerning their interpretation, is of little practical use. We have addressed this sparseness of meta-knowledge available in existing bio-event corpora by developing a multi-dimensional annotation scheme tailored to bio-events. The scheme is intended to be general enough to allow integration with different types of bio-event annotation, whilst being detailed enough to capture important subtleties in the nature of the meta-knowledge expressed about different bio-events. To our knowledge, our scheme is unique within the field with regards to the diversity of meta-knowledge aspects annotated for each event.
Language resources, including corpus and tools, are normally required to be combined in order to achieve a users specific task. However, resources tend to be developed independently in different, incompatible formats. In this paper we describe about U-Compare, which consists of the U-Compare component repository and the U-Compare platform. We have been building a highly interoperable resource library, providing the world largest ready-to-use UIMA component repository including wide variety of corpus readers and state-of-the-art language tools. These resources can be deployed as local services or web services, even possible to be hosted in clustered machines to increase the performance, while users do not need to be aware of such differences. In addition to the resource library, an integrated language processing platform is provided, allowing workflow creation, comparison, evaluation and visualization, using the resources in the library or any UIMA component, without any programming via graphical user interfaces, while a command line launcher is also available without GUIs. The evaluation itself is processed in a UIMA component, users can create and plug their own evaluation metrics in addition to the predefined metrics. U-Compare has been successfully used in many projects including BioCreative, Conll and the BioNLP shared task.
L’analyse qualitative des données demande au sociologue un important travail de sélection et d’interprétation des documents. Afin de faciliter ce travail, cette communauté c’est dotée d’outils informatique mais leur fonctionnalités sont encore limitées. Le projet ASSIST est une étude exploratoire pour préciser les modules de traitement automatique des langues (TAL) permettant d’assister le sociologue dans son travail d’analyse. Nous présentons le moteur de recherche réalisé et nous justifions le choix des composants de TAL intégrés au prototype.
It is a challenging task to match similar or related terms/expressions in NLP and Text Mining applications. Two typical areas in need for such work are terminology and ontology constructions, where terms and concepts are extracted and organized into certain structures with various semantic relations. In the EU BOOTSTrep Project we test various techniques for matching terms that can assist human domain experts in building and enriching ontologies. This paper reports on a work in which we evaluated a text comparing and clustering tool for this task. Particularly, we explore the feasibility of matching related terms with their definitions. Ontology terms, such as Gene Ontology terms, are often assigned with detailed definitions, which provide a fundamental information source for detecting relations between terms. Here we focus on the exploitation of term definitions for the term matching task. Our experiment shows that the tool is capable of grouping many related terms using their definitions.
This paper reports on the design and construction of a bio-event annotated corpus which was developed with a specific view to the acquisition of semantic frames from biomedical corpora. We describe the adopted annotation scheme and the annotation process, which is supported by a dedicated annotation tool. The annotated corpus contains 677 abstracts of biomedical research articles.
Many systems have been developed in the past few years to assist researchers in the discovery of knowledge published as English text, for example in the PubMed database. At the same time, higher level collective knowledge is often published using a graphical notation representing all the entities in a pathway and their interactions. We believe that these pathway visualizations could serve as an effective user interface for knowledge discovery if they can be linked to the text in publications. Since the graphical elements in a Pathway are of a very different nature than their corresponding descriptions in English text, we developed a prototype system called PathText. The goal of PathText is to serve as a bridge between these two different representations. In this paper, we first describe the overall architecture and the interfaces of the PathText system, and then provide some details about the core Text Mining components.
Given the increasing number of neologisms in biomedicine (names of genes, diseases, molecules, etc.), the rate of acronyms used in literature also increases. Existing acronym dictionaries cannot keep up with the rate of new creations. Thus, discovering and disambiguating acronyms and their expanded forms are essential aspects of text mining and terminology management. We present a method for clustering long forms identified by an acronym recognition method. Applying the acronym recognition method to MEDLINE abstracts, we obtained a list of short/long forms. The recognized short/long forms were classified by abiologist to construct an evaluation set for clustering sets of similar long forms. We observed five types of term variation in the evaluation set and defined four similarity measures to gathers the similar longforms (i.e., orthographic, morphological, syntactic, lexico semantic variants, nested abbreviations). The complete-link clustering with the four similarity measures achieved 87.5% precision and 84.9% recall on the evaluation set.
One of the main challenges in biomedical text mining is the identification of terminology, which is a key factor for accessing and integrating the information stored in literature. Manual creation of biomedical terminologies cannot keep pace with the data that becomes available. Still, many of them have been used in attempts to recognise terms in literature, but their suitability for text mining has been questioned as substantial re-engineering is needed to tailor the resources for automatic processing. Several approaches have been suggested to automatically integrate and map between resources, but the problems of extensive variability of lexical representations and ambiguity have been revealed. In this paper we present a methodology to automatically maintain a biomedical terminological database, which contains automatically extracted terms, their mutual relationships, features and possible annotations that can be useful in text processing. In addition to TermDB, a database used for terminology management and storage, we present the following modules that are used to populate the database: TerMine (recognition, extraction and normalisation of terms from literature), AcroTerMine (extraction and clustering of acronyms and their long forms), AnnoTerm (annotation and classification of terms), and ClusTerm (extraction of term associations and clustering of terms).