[Paper Review] Lexical Sememe Prediction using Dictionary Definitions by Capturing Local Semantic Correspondence

Jiaju Du, Fanchao Qi+2 more|arXiv (Cornell University)|Jan 16, 2020

Topic Modeling31 references4 citations

TL;DR

This paper proposes Sememe Correspondence Pooling (SCorP), a novel model that predicts lexical sememes by capturing local semantic correspondence between sememes and words in dictionary definitions. By modeling correspondence scores and applying max-pooling, SCorP achieves state-of-the-art performance on HowNet, demonstrating robustness across word frequency ranges and effective alignment of sememes with definition words.

ABSTRACT

Sememes, defined as the minimum semantic units of human languages in linguistics, have been proven useful in many NLP tasks. Since manual construction and update of sememe knowledge bases (KBs) are costly, the task of automatic sememe prediction has been proposed to assist sememe annotation. In this paper, we explore the approach of applying dictionary definitions to predicting sememes for unannotated words. We find that sememes of each word are usually semantically matched to different words in its dictionary definition, and we name this matching relationship local semantic correspondence. Accordingly, we propose a Sememe Correspondence Pooling (SCorP) model, which is able to capture this kind of matching to predict sememes. We evaluate our model and baseline methods on a famous sememe KB HowNet and find that our model achieves state-of-the-art performance. Moreover, further quantitative analysis shows that our model can properly learn the local semantic correspondence between sememes and words in dictionary definitions, which explains the effectiveness of our model. The source codes of this paper can be obtained from https://github.com/thunlp/scorp.

Motivation & Objective

• To address the high cost of manual sememe knowledge base (KB) construction and updating by enabling automatic sememe prediction.
• To leverage dictionary definitions as a rich, accessible resource for sememe prediction, especially for low-frequency words.
• To discover and exploit local semantic correspondence—where each sememe matches specific words in a word's definition—thereby improving prediction accuracy.
• To develop a model that avoids imposing artificial order on sememes and mitigates issues from character-level ambiguity in prior approaches.

Proposed method

• SCorP models the semantic correspondence between each sememe and every word in a word's dictionary definition using a learnable scoring mechanism.
• It computes correspondence scores via a neural attention-like mechanism and applies max-pooling over definition words for each sememe to aggregate the strongest matches.
• The model uses word-level embeddings as input, avoiding character-level ambiguity and enabling better representation learning.
• Two retrofitting operations are introduced: incorporating sememes of definition words and integrating target word embeddings to enhance performance.
• The framework is a multi-label classification model that treats sememe prediction as a sequence-to-set task, avoiding sequence ordering constraints.
• The model is trained end-to-end on HowNet using cross-entropy loss, with fine-tuning enabled via the retrofitting techniques.

Experimental results

Research questions

• RQ1 Can dictionary definitions be effectively leveraged to improve lexical sememe prediction, especially for low-frequency words?
• RQ2 Is there a local semantic correspondence between sememes and individual words in dictionary definitions, and can it be exploited to improve prediction?
• RQ3 How does modeling correspondence between sememes and definition words compare to sequence-to-sequence or character-level models in performance and robustness?
• RQ4 To what extent do retrofitting operations—incorporating sememes of definition words and target word embeddings—enhance model performance?

Key findings

• SCorP achieves state-of-the-art performance on the HowNet sememe prediction benchmark, outperforming existing methods including sequence-to-sequence and character-level models.
• The model shows strong robustness across word frequency ranges, maintaining high performance even for low-frequency words (frequency < 50), unlike CSP which suffers a sharp drop.
• Quantitative analysis confirms that SCorP effectively learns local semantic correspondence, with correct sememes consistently achieving the highest correspondence scores for relevant definition words.
• The case study in Table 5 demonstrates that the model correctly identifies the top sememes for 'observatory' by aligning them with semantically matching definition words such as 'institution' for 'location'.
• The two retrofitting operations significantly improve performance, with the final model achieving higher F1-scores than baseline models.
• The model's ability to capture correspondence between sememes and definition words is validated through attention-style attention maps and correspondence score analysis.