PaperlayerView exact revision on arXiv

Source-linked AI summary

A Question Answering Approach to Emotion Cause Extraction

Lin Gui, Jiannan Hu, Yulan He, Ruifeng Xu, Qin Lu, Jiachen Du

arXiv:1708.05482v2cs.CL

TL;DR

Emotion cause extraction seeks to identify why an expressed emotion occurs, a harder task than emotion classification. The paper treats it as QA and introduces a convolutional multi-slot deep memory network that models contextual information. It reports state-of-the-art performance, surpassing the previous best method by 3.01% in F-measure.

  • Problem

    Emotion cause extraction identifies the reasons behind expressed emotions and is more difficult than emotion classification because it requires understanding the text conveying an emotion.

  • Method

    The paper formulates emotion cause extraction as QA and uses a deep memory network with convolutional multiple memory slots to capture sequential context.

  • Results

    3.01% in F-measure: ConvMS-Memnet outperforms the previous best-performing multi-kernel method on the emotion cause dataset.

  • Takeaways & Limitations

    The proposed architecture captures context information for emotion cause identification and achieves state-of-the-art performance against competitive baselines.

  • Takeaways & Limitations

    The model may struggle with long-distance dependencies, negations, emotion transitions, and discourse relations among clauses.

Abstract

from arXiv · show

Emotion cause extraction aims to identify the reasons behind a certain emotion expressed in text. It is a much more difficult task compared to emotion classification. Inspired by recent advances in using deep memory networks for question answering (QA), we propose a new approach which considers emotion cause identification as a reading comprehension task in QA. Inspired by convolutional neural networks, we propose a new mechanism to store relevant context in different memory slots to model context information. Our proposed approach can extract both word level sequence features and lexical features. Performance evaluation shows that our method achieves the state-of-the-art performance on a recently released emotion cause dataset, outperforming a number of competitive baselines by at least 3.01% in F-measure.

1 Introduction

Emotion cause extraction identifies why an emotion occurs, a harder problem than classifying emotions. The paper recasts it as a QA task and proposes a convolutional multi-slot memory network to model contextual relations.

  • Emotion cause extraction identifies the reason behind an emotion expression and requires deeper text understanding than emotion classification.
  • Existing approaches rely mainly on linguistic rules or lexicon features while overlooking semantic information and emotion–cause relations.
  • The proposed QA formulation uses an emotion word as a query and predicts whether a described event causes that emotion.
  • The model extends memory networks with convolutional operations and multiple memory slots to capture sequential context around words.
  • The paper evaluates the proposed deep memory network against existing approaches for emotion cause extraction.

2 Related Work

Prior emotion-analysis research largely focuses on emotion classification and information extraction, while emotion cause extraction has used rules, knowledge bases, and machine-learning methods. This paper positions memory networks as a QA-based way to model relations between emotion words and their causes.

  • Emotion analysis research primarily addresses emotion classification and emotion-related information extraction.
  • Emotion cause extraction was introduced as a task requiring identification of the clause containing an emotion cause.
  • Prior methods include rule-based systems, common-sense knowledge bases, conditional random fields, and multi-kernel models.
  • Existing work often fails to model the relation between an emotion word and its cause, motivating a QA formulation based on memory networks.
  • The study presents memory networks as a first application of that architecture to emotion cause extraction.

3 Our Approach

The paper formulates emotion cause extraction as a question-answering task and develops memory-network architectures that model word relevance, multiple hops, and local context. Its ConvMS-Memnet uses convolutional multiple memory slots to capture sequence-level information.

  • 3.1 Task Definition: Emotion cause extraction is defined as identifying the clause containing the cause of an emotion word in a document.The approach selects the clause with the highest predicted probability as the emotion cause.
  • 3.2 Memory Network: The basic memory network matches clause words with an emotion-word query, converts correlations into attention weights, and produces a yes/no prediction.Word and emotion representations are embedded in the same space, with attention weights normalized across clause positions.
  • 3.2 Memory Network: Deep memory networks stack multiple hops so successive prediction vectors refine the relation between the emotion query and the clause.The model uses the emotion embedding at the first hop and propagates prediction vectors through later hops.
  • 3.2 Memory Network: The basic memory network is more principled than rule-based or linguistic-feature methods but does not capture sequential context because it models words separately.Context is important for word meaning, negation, and emotion transitions.
  • 3.3 Convolutional Multiple-Slot Deep Memory Network: ConvMS-Memnet models clause context with convolutional multiple memory slots, using neighboring words and three output vectors for each position.The convolutional kernel considers the previous and following words, while concatenated slot outputs provide a sequence-level feature.
  • 3.3 Convolutional Multiple-Slot Deep Memory Network: The convolutional architecture is extended to multiple computational layers, and training uses stochastic gradient descent, backpropagation, skip-gram embeddings, and dropout.The reported implementation uses 20-dimensional word embeddings and dropout set to 0.4.

4 Experiments and Evaluation

The experiments section introduces its evaluation settings before reporting results.

  • The section first presents the experimental settings and then reports the results.

4.1 Experimental Setup and Dataset

The experiments use a simplified Chinese emotion-cause corpus and evaluate clause-level cause identification with precision, recall, and F-measure across repeated train/test splits.

  • 4.1 Experimental Setup and Dataset: The corpus contains 2,105 SINA city-news documents with one emotion word, one or more causes, and manually segmented clauses.
  • 4.1 Experimental Setup and Dataset: The main task is to identify which clause contains the emotion cause, using an annotated clause as correct when it covers the answer.
  • 4.1 Experimental Setup and Dataset: Evaluation follows precision, recall, and F-measure metrics used in prior work for comparable emotion-cause extraction results.
  • 4.1 Experimental Setup and Dataset: The study randomly assigns 90% of documents to training and 10% to testing, repeating evaluation 25 times with different splits.

4.2 Evaluation and Comparison

ConvMS-Memnet is compared with rule-based, common-sense, machine-learning, neural, memory-network, and multi-kernel baselines. It improves over Memnet and the prior best method on the main benchmark.

  • 4.2 Evaluation and Comparison: The baselines include rule-based, common-sense, combined machine-learning, SVM, Word2vec, multi-kernel, CNN, and Memnet methods.
  • 4.2 Evaluation and Comparison: The ConvMS-Memnet configuration uses pre-trained skip-gram word embeddings and three hops, with performance also examined across embedding choices and hop counts.
  • 4.2 Evaluation and Comparison: RB has fairly high precision but low recall, CB has the highest recall but worst precision, and RB+CB improves F-measure over RB by only 1.27%.
  • 4.2 Evaluation and Comparison: RB+CB+ML reaches an F-measure of 0.5597, while the multi-kernel method is the strongest baseline because it models context structurally and uses an emotion lexicon.
  • 4.2 Evaluation and Comparison: ConvMS-Memnet improves over Memnet by 11.54% in precision, 4.84% in recall, and 8.24% in F-measure, and exceeds multi-kernel by 3.01% in F-measure.The improvement over Memnet is reported as statistically significant with p-value less than 0.01.

4.3 More Insights into the ConvMS-Memnet

Additional experiments examine embeddings, hop depth, attention, and training duration. Pre-trained embeddings and deeper memory processing improve performance, while attention shifts toward the annotated cause.

  • 4.3.1 Pre-trained Word Embeddings: Pre-trained word embeddings produce 2.59% higher F-measure than random initialization, partly because the training data are limited.
  • 4.3.3 Word-Level Attention Weights: With at least three hops, attention shifts from “more” toward “insistence,” the emotion-cause keyword, and better results are obtained.
  • 4.3.3 Word-Level Attention Weights: The keyword-level metric treats the highest-attention word within an identified clause as correct when it falls within the annotated cause boundary.
  • 4.3.3 Word-Level Attention Weights: ConvMS-Memnet outperforms Memnet by 5.6% in keyword-level F-measure, indicating better word-level emotion-cause identification.
  • 4.3.4 Training Epochs: In a training-epoch case study, the model identifies the correct emotion-cause clause after five epochs, with its probability increasing as training continues.

4.4 Limitations

The model has difficulty with complex passages involving long-distance dependencies, negations, emotion transitions, and discourse relations. Its binary answers and limited annotated data also prevent directly generating the emotion cause.

  • 4.4 Limitations: Long-distance dependencies, negations, and emotion transitions can make the model misidentify the clause containing an emotion cause.The authors identify discourse relations among clauses as a challenging modeling issue.
  • 4.4 Limitations: The authors propose exploring discourse-aware architectures, transfer learning from larger annotated datasets, automatic annotation, and improved evaluation mechanisms.These directions target both discourse modeling and the data and evaluation constraints of the QA formulation.
  • 4.4 Limitations: Small annotated corpora constrain the model to generating only “yes” or “no” rather than a natural-language cause.The authors attribute the limitation to the expense of manually annotating data.

5 Conclusions

The paper formulates emotion cause extraction as question answering and proposes a deep memory network that models context sequentially. The model achieves state-of-the-art performance on an emotion cause detection dataset, while future work targets discourse modeling and direct cause generation.

  • 5 Conclusions: The proposed QA-based deep memory network achieves state-of-the-art performance against competitive baselines on an emotion cause detection dataset.The conclusion reports the overall benchmark outcome without providing numeric values.
  • 5 Conclusions: The architecture stores context in multiple memory slots and captures its sequence through convolutional operations.This context modeling is presented as the approach’s key property.
  • 5 Conclusions: Future work will model discourse relations among clauses and develop QA answers that directly output emotion causes.The proposed directions address limitations of the current formulation.
Loading 1708.05482v2…