TL;DR
This paper introduces a novel graph-based approach to building reverse dictionaries, which map phrases to single words. By leveraging word definitions and a "reverse map" concept, their method achieves performance comparable to existing commercial solutions like Onelook, particularly on smaller lexicons. This work offers a computationally efficient and interpretable alternative for tackling the "tip-of-the-tongue" problem.
What Problem Does This Paper Address?
Ever struggled to recall a specific word, even when you can clearly describe its meaning? This is the "tip-of-the-tongue" phenomenon, and it's a common linguistic challenge. Traditional dictionaries (forward dictionaries) map words to their definitions. Reverse dictionaries, on the other hand, aim to solve this by taking a descriptive phrase as input and outputting a list of words that semantically match that phrase. This is also crucial for individuals with word-finding difficulties, such as those with anomic aphasia. The challenge lies in dynamically computing the meaning of a phrase and finding a single word that best approximates it, especially given the potentially infinite number of phrases that can be constructed from a finite vocabulary. Existing approaches often rely on complex methods like compositional distributional semantics or recurrent neural networks, with varying degrees of success.
Methods
The core of this paper's contribution is a novel node-graph architecture for reverse dictionaries. Instead of directly mapping phrases to words, they first extract "content words" from the input phrase. Then, they construct a "reverse map" from a forward dictionary. In a forward map, words lead to the words in their definitions. In a reverse map, words lead to the words whose definitions contain them. This reverse map is then transformed into a connectivity matrix (Back-Linked Matrix - BLM).
Each word is represented as a node in a graph, and the BLM defines the connections between these nodes. To find a word for a given phrase, they create a graph for each input word, activate the corresponding node, and then "evolve" the graph through a "n-layered search" (or search depth). A similarity measure is then computed between the input phrase and each word in the lexicon, based on the inverse of the distances in this graph. This measure is weighted by the inverse frequency of the input words, giving more importance to "semantically important" words. To ensure all words can connect, they also construct a mixed back-linked matrix (mBLM) that incorporates forward links for words with insufficient backward connections.
Datasets
The primary dataset used is the Oxford 3000 wordlist, which served as the basis for their forward dictionary. They also utilized:
- Oxford Learner's Dictionary (OLD)
- Merriam-Webster Dictionary (MW)
- WordNet (WN)
These were used as forward dictionaries to construct the BLMs and mBLMs for the Oxford 3000 wordlist. For evaluation, they also used the WordNet lexicon (WL), which is significantly larger (80k words).
For performance testing, they used:
- Macmillan Dictionary Definitions: 179 definitions were used for evaluation.
- User Concept Descriptions: 179 phrases generated by users were used for evaluation.
Key Insights and Ideas
- The "Reverse Map" is Key: Shifting from a forward dictionary's "word -> definition" to a "definition -> word" perspective (the reverse map) is crucial for efficiently finding words that define a concept.
- Graph-Based Similarity is Powerful: Representing word relationships as a graph and using graph traversal (n-layered search) provides a structured way to measure semantic similarity between phrases and words.
- Semantic Importance Matters: Weighting the similarity measure by the inverse frequency of input words helps to capture the "semantic importance" of each word in the phrase.
- Shallow Search is Often Sufficient: The research found that a relatively shallow search depth (around 3 hops) is often enough to achieve stable and good performance, making the approach computationally efficient.
- Lexicon Size is a Bottleneck: While the approach performs well on smaller lexicons (like the 3k wordlist), performance significantly deteriorates on larger ones (80k WordNet lexicon) due to a high branching factor and similar distances between many words.
Key Results
- On a 3k lexicon, their approach (specifically the Fusion mBLM) achieved an accuracy of .25/.55/.84 (at ranks 1, 10, and 100 respectively) for user-generated phrases, which is comparable to or better than Onelook (.19/.41/.65) and significantly outperforms the word2vec baseline (.01/.06/.20).
- The Fusion mBLM performed better than individual dictionary BLMs, suggesting the benefit of combining multiple dictionaries.
- When tested on the larger 80k WordNet lexicon, performance dropped significantly, with accuracies of .03/.15/.36 for the WL BLM.
- The paper concludes that their approach is "much better than the distributional semantics method, and as good as Onelook, on a 3k lexicon."
Limitations and Future Work
- Lexicon Size Limitation: The most significant limitation acknowledged is the performance degradation on larger lexicons. The high branching factor in large dictionaries leads to similar distances between many words, obscuring semantic similarity. The authors note there's "no trivial fix" for this.
- Multi-Word Expressions: Handling multi-word expressions (like "kick the bucket") is not directly addressed. The current approach treats each word individually. Future work could involve recursive parsing or treating such expressions as single nodes.
- Improving Similarity Measure: The authors suggest exploring semantic rules, lexical relations, or user feedback to enrich the similarity measure beyond just dictionary definitions.
- Alternative Graph Implementations: Investigating alternative graph implementations, such as using a single graph with frequency tagging or exploring Spiking Neural Networks, could be beneficial.
- Generalization: The approach is presented as generalizable to any language given an appropriate forward dictionary, lemmatizer, and functional word list.
Summary updated