Source-linked AI summary

Sequential Short-Text Classification with Recurrent and Convolutional Neural Networks

Ji Young Lee, Franck Dernoncourt

arXiv:1603.03827v1cs.CLcs.AIcs.LGcs.NEstat.ML

TL;DR

Most ANN systems classify short texts without using preceding texts, despite short texts commonly occurring in sequences. The paper introduces an RNN- and CNN-based sequential classifier for dialog act prediction and reports state-of-the-art results on three datasets.

  • Problem

    Most ANN systems classify short texts in isolation, although short texts commonly occur in sequences where preceding texts may improve accuracy.

  • Method

    The paper combines recurrent neural networks and convolutional neural networks to incorporate preceding short texts for sequential short-text classification.

  • Results

    The model achieves state-of-the-art results on three different datasets for dialog act prediction.

  • Takeaways & Limitations

    Adding sequential information improves prediction quality, and performance depends on which sequential information the model uses.

  • Takeaways & Limitations

    Rigorous comparisons are difficult because preprocessing, data splits, and the number of training runs vary across studies.

Abstract

from arXiv · show

Recent approaches based on artificial neural networks (ANNs) have shown promising results for short-text classification. However, many short texts occur in sequences (e.g., sentences in a document or utterances in a dialog), and most existing ANN-based systems do not leverage the preceding short texts when classifying a subsequent one. In this work, we present a model based on recurrent neural networks and convolutional neural networks that incorporates the preceding short texts. Our model achieves state-of-the-art results on three different datasets for dialog act prediction.

1 Introduction

Short-text classification is important across several NLP applications, but most ANN systems classify each text without preceding context. The paper addresses this gap with an RNN- and CNN-based model for sequential short-text classification.

  • Short-text classification supports applications including sentiment analysis, question answering, and dialog management.
  • Most ANN systems classify short texts in isolation without considering preceding short texts.
  • Short texts usually occur in sequences, motivating the use of preceding texts to improve classification accuracy.
  • The paper introduces an RNN- and CNN-based model for sequential short-text classification and evaluates it on dialog act classification.The model achieves state-of-the-art results on three different datasets.

2 Model

The model represents each short text with an RNN or CNN, then classifies the current text using its representation together with representations of preceding texts. A two-layer feedforward ANN produces the probability distribution over classes.

  • Short-text representation: The model first generates a vector representation for each short text using either an RNN or CNN.The input is a sequence of word vectors, and the resulting representation is an n-dimensional vector s.
  • RNN-based representation: The RNN uses an LSTM to transform word-level inputs into hidden-state vectors, which are combined by last, mean, or max pooling.Last pooling selects h_ℓ, mean pooling averages the hidden states, and max pooling takes their element-wise maximum.
  • CNN-based representation: The CNN applies n filters to windows of h consecutive words and uses element-wise max pooling over the resulting features to form s.Each feature dimension comes from a distinct filter, and the convolution is followed by a ReLU operation.
  • Sequential classification: The sequential classifier feeds the representation sequence s_i−d1−d2:i into a two-layer feedforward ANN to predict the class for short text i.The history sizes d1 and d2 determine how many preceding representations enter the first and second layers.
  • Sequential classification: The final output z_i is a probability distribution over k classes, with each element giving the probability of one class for the ith short text.The first layer produces class representations, and the second layer transforms the relevant sequence into z_i.

3 Datasets and Experimental Setup

The model is evaluated for dialog act classification across three datasets, using sequence-oriented dataset splits and standard neural-network training with regularization.

  • The evaluation covers DSTC 4, MRDA, and SwDA dialog act classification datasets.
  • MRDA provides train, validation, and test splits, whereas DSTC 4 and SwDA provide only train and test splits.
  • The dataset overview reports class counts, vocabulary sizes, and dialog and utterance counts for the train, validation, and test sets.
  • Training minimizes negative log-likelihood with stochastic gradient descent and Adadelta, updating weights, biases, and word vectors.
  • Dropout follows the pooling layer, and early stopping uses the validation set with a patience of 10 epochs.

4 Results and Discussion

Experiments examine hyperparameters, representations, history sizes, and comparisons with prior methods. Sequential information improves performance, especially when incorporated at the short-text representation level, while comparisons are constrained by differing experimental details.

  • Hyperparameters were selected by varying one hyperparameter at a time while keeping the others fixed.
  • Sequential classification outperforms non-sequential classification, and CNN outperforms LSTM across all datasets, usually by a small margin except for SwDA.
  • Increasing short-text history size d1 with d2 = 0 improves performance by 1.3–4.2 percentage points in both LSTM and CNN models.
  • Sequential information at the short-text representation level is more effective than at the class representation level.
  • Using sequential information at both representation levels does not help in most cases and may lower performance.
  • The authors hypothesize that class representations add less information and are more likely to propagate errors from previous misclassifications.
  • The model shows competitive results without human-engineered features, but rigorous comparisons are difficult because preprocessing, data splits, and training runs vary.

5 Conclusion

The paper presents an ANN-based approach to sequential short-text classification and reports improved prediction quality from sequential information. It achieves state-of-the-art dialog act prediction results on three datasets.

  • The paper presents an ANN-based approach to sequential short-text classification.
  • Adding sequential information improves prediction quality, with performance depending on which sequential information the model uses.
  • The model achieves state-of-the-art results on three datasets for dialog act prediction.
Loading 1603.03827v1…