TL;DR
This paper introduces "Di-LSTM Contrast," a deep neural network designed to detect metaphorical language by focusing on the contrast between a word's general meaning and its contextual usage. By leveraging bidirectional LSTMs and a novel feature selection mechanism, the model achieves competitive results on the Metaphor Shared Task 2018, outperforming baselines in certain categories.
What Problem Does This Paper Address?
Understanding and generating human language is a significant challenge for artificial intelligence, and a key aspect of this is comprehending figurative language, particularly metaphors. Metaphors, while enriching language, pose difficulties for machines because they involve mapping concepts from one domain to another, making literal interpretation insufficient. This paper tackles the problem of word-level metaphor detection, aiming to identify whether a specific word in a sentence is used metaphorically or literally. This is crucial for downstream NLP tasks like sentiment analysis, machine translation, and question answering, where misinterpreting metaphors can lead to significant errors.
Methods
The core of the proposed system is the Di-LSTM Contrast Network, a deep neural architecture trained end-to-end. It comprises three main modules:
- Context Encoder: This module uses Bidirectional Long Short-Term Memory (BiLSTM) networks to encode the context of a target word within a sentence. It splits the sentence into two parts: one preceding the target word and one following it, feeding these into separate forward and backward LSTMs. The resulting hidden states are then combined (concatenated or averaged) and passed through a fully connected layer to produce a context encoding vector.
- Feature Selection: This module generates contrast-based features. It takes the word vector of the target word and combines it with the difference between the target word's vector and the context encoding vector. The intuition here is that the difference between a word's general meaning (represented by its vector) and its contextual meaning (represented by the context encoding) is a strong indicator of metaphoricity.
- Classifier: The generated features are fed into a classifier, which consists of a hidden layer followed by an output layer. This classifier outputs the softmax probabilities for the target word being metaphoric.
The authors also explore cost-sensitive learning by re-weighting examples to address class imbalance and incorporate additional baseline features such as WordNet features, POS tags, and concreteness ratings.
Datasets
The primary dataset used is the VU Amsterdam Metaphor Corpus (VUAMC). This corpus is derived from the British National Corpus (BNC Baby) and annotated using the MIPVU Procedure. It includes examples from four text genres: Academic, News, Fiction, and Conversation. The paper also refers to the Metaphor Shared Task 2018 benchmark, which splits the data into two categories: "All POS" (covering nouns, adjectives, adverbs, and verbs) and "Verbs" (specifically for verbs).
Key Insights and Ideas
- Contextual Contrast is Key: The central idea that the difference between a word's general meaning and its specific contextual usage is a powerful signal for metaphor detection is a key insight. This "contrast" is explicitly modeled by subtracting the context encoding from the word embedding.
- BiLSTM for Contextual Encoding: Using bidirectional LSTMs effectively captures the surrounding context of a word, which is essential for disambiguating literal from metaphorical uses.
- Feature Engineering Still Matters: While deep learning models are powerful, incorporating well-chosen baseline features (like concreteness and WordNet) significantly boosts performance, suggesting a hybrid approach can be highly effective.
- Cost-Sensitive Learning Benefits: Re-weighting training examples proves to be an effective strategy for improving performance, particularly in imbalanced datasets common in NLP tasks.
- Concatenation Outperforms Averaging: For combining the forward and backward LSTM outputs, concatenation of hidden states followed by a fully connected layer yielded better results than simple averaging, indicating that more complex interactions between contextual representations are beneficial.
Key Results
The Di-LSTM Contrast model achieved competitive results on the Metaphor Shared Task 2018. The best-performing system achieved an overall F1 score of 0.570 on the All POS category and 0.605 on the Verbs category. The paper also shows that their model, particularly the "DC +RL" variant (DiLSTM Contrast with concatenation and re-weighting), surpassed the task baseline in the Verbs category and achieved comparable performance on the All POS category. Analysis by text genre showed better performance on academic and news texts compared to conversation and fiction texts.
Limitations and Future Work
The authors acknowledge that their model achieves a "lesser but comparable performance" to the baseline on the All POS category.
For future work, they suggest exploring:
- CNNs alongside LSTMs: To potentially capture context representations more effectively.
- Attention Mechanisms: To allow the model to focus on more relevant parts of the context.
- Experimenting with other NLP tasks: To further validate the utility of their metaphor detection approach.
Summary updated