Paper

Content

Event

• Zero-shot Label-Aware Event Trigger and Argument Classification (2020) [Hongming Zhang, Haoyu Wang, Dan Roth]

  • Goal:
    Proposed a zero-shot event extraction approach, which first identifies events with existing tools (e.g., SRL) and then maps them to a given taxonomy of event types in a zero-shot manner.

  • Core Method:

    1. Event Identification：find event triggers by the SRL models and classify whether triggers belong to the pre-defined event ontology or not (classification task)
    2. Represent the labels with a cluster of contextualized embeddings (rather than just words). More specially, select several sentences that contain the word “attack” from an external corpus (e.g., New York Times (NYT)). Use a pre-trained language model (e.g., BERT (Devlin et al., 2019)) to get the word embedding as a data point in the “attack" cluster.
    3. At the inference time, acquire the contextualized representation of identified triggers and arguments, and map them to their corresponding types based on their
       similarities to those clusters in the embedding space.
    4. Beyond the labels, event definitions also provide constraints between types and roles. For example, the role “Attacker” can only appear as an argument of “Conflict:Attack” rather than “Life:Marry,” and only a person or a nation can take the role of an “Attacker". These constraints to regularize the zero-shot model by formulating as an integer linear programming (ILP) (Roth and tau Yih, 2004) and only produce decisions that satisfy all the constraints.

  • Achievement:

    1. Map 83% of the triggers and 54% of the arguments to correct types, almost doubling the performance of the previous best zero-shot event classification approach
    2. When pipelined with our improved zero-shot identification step, we exhibit an zero-shot event extraction pipeline that rivals a SOTA supervised system (Lin et al., 2020) trained on over 6,000 sentences.

  • Data Source:

    1. ACE-2005, which has 33 event trigger types and 22 argument role types.

NLI

• A large annotated corpus for learning natural language inference (EMNLP2015) Samuel R. Bowman, etc.
  • Abstract:

Event Schema and Script

• Connecting the Dots: Event Graph Schema Induction with Path Language Modeling (EMNLP2020) Manling Li, etc.

  • Abstract:

• The Future is not One-dimensional: Complex Event Schema Induction by Graph Modeling for Event Prediction (EMNLP2021) Manling Li, etc.

  • Abstract:

Event Coreference

• BERT for Coreference Resolution: Baselines and Analysis (ACL 2019) Mandar etc.
  • Abstract:

Event Temporal

• Temporal Reasoning on Implicit Events from Distant Supervision (NAACI 2021) Ben Zhou, Kyle Richardson, Qiang Ning, Tushar Khot, Ashish Sabharwal1, Dan Roth Git repository

  • Abstract:
    Propose a neuro-symbolic temporal reasoning model, SYMTIME, which exploits distant supervision signals from large-scale text and uses temporal rules to combine start times and durations to infer end times. Introduce a new dataset TRACIE (TempoRAL Closure InfErence) foucses on temporal relations on implicit events in short stories.
  • Core Methods:
    1. Use T5-Large fine-tuned on TRACIE to achieve binary prediction accuracy of 67.9% as baseline.
    2. Propose a novel distant supervision technique that improves generalization by extracting temporal patterns in large-scale free text as part of an additional pre- training step. Use large windows of text such as paragraphs in contrast to sentence level. Result a mode PINTIME(Pattern-Time) achieving 76.6% on TRACIE.
    3. Couple PTNTIME with a duration model from Zhou et al. (2020) to create a neural-symbolic reasoning model called SYMTIME, which achieves 78.9% on TRACIE.

• Conditional Generation of Temporally-ordered Event Sequences Shih-Ting Lin et al. Git [repository]

  • Abstract:
    Proposes a single model of events to suupport inferenes in two tasks: (1) temporal event ordering; (2)event infilling or inferring unseen or unmentioned event infilling.

  • Core Methods:

    1. Take temporally-ordered event sequences, shuffle them, delete some events, and then attempt to recover the original event sequence.
    2. BART-based model.

  • Evaluation Methods and Data Source:

  • 1. EventsNarrative Dataset. (Training). Assuming discourse order reflect temporal order.
  • 2. CaTeRS. (Evaluation). Used for both temporal relations and event infilling evaluation.
  • 3. MCTaco. (Evaluation). For temporal relations evaluation.
  • 4. BERT-based pairwise model and BERT-based pointerr network are used as comparison for temporal ordering task.
  • 5. GPT-2 baselines is used as comparison for infilled events generation task.

• Context-dependent Semantic Parsing for Time Expressions (ACL 2014) Kenton Lee, etc.

  • Abstract: The goal is to find all time expressions in an input document. Problem is devided into two parts: detection and resolution(normalization). Make use of a hand-engineered Combinatory Categorial Grammar (CCG) to construct a set of meaning representations that identify the time being described.
  • Core Methods:
  • Achievement:
  • Data Source:

• Multi-TimeLine Summarization (MTLS): Improving Timeline Summarization by Generating Multiple Summaries (ACL/IJCNLP 2021) Yi Yu, etc.

  • Contribution:

    • propose a novel task (MTLS), which auto-matically generates multiple, informative, and diverse timelines from an input time-stamped document collection.
    • introduce a superior MTLS model that outperforms all TLS-adapted MTLS baselines.
    • design an evaluation measure for MTLS systems by extending the original TLS evaluation framework (adapt the tilse framework to MTLS task).

    (Use cos score to detect the distance of two sentences in BERT semantic space, and add a function to calcuate time gap between the two verbs in sentences by applying a reverse exponential equation to idenitfy if they describe the same event.)

Other Interesting topic

• Probing Across Time: What Does RoBERTa Know and When? (2021) Leo Z. Liu etc.

  • Abstract: Following this effort, we systematically answer a question: for various types of knowledge a language model learns, when during (pre)training are they acquired? Using RoBERTa as a case study, we find: linguistic knowledge is acquired fast, stably, and robustly across domains. Facts and commonsense are slower and more domain- sensitive. Reasoning abilities are, in general, not stably acquired. As new datasets, pretraining protocols, and probes emerge, we believe that probing-across-time analyses can help researchers understand the complex, intermingled learning that these models undergo and guide us toward more efficient approaches that accomplish necessary learning faster.

• How Can We Know When Language Models Know? On the Calibration of Language Models for Question Answering (2021) Zhengbao Jiang etc.

  • Abstract:
    In this paper, we ask the question "how can we know when language models know, with confidence, the answer to a particular query?" We examine this question from the point of view of calibration, the property of a probabilistic model's predicted probabilities actually being well correlated with the probabilities of correctness. We examine three strong generative models -- T5, BART, and GPT-2 -- and study whether their probabilities on QA tasks are well calibrated, finding the answer is a relatively emphatic no. We then examine methods to calibrate such models to make their confidence scores correlate better with the likelihood of correctness through fine-tuning, post-hoc probability modification, or adjustment of the predicted outputs or inputs.

• Title (Conference year) Author etc.

  • Abstract:
  • Achievement:
  • Data Source: