Leviplus
commented
5 years ago Weekly report at June 14, 2019
1 Paper reading
Paper (1): Coarse-to-fine decoding for neural semantic parsing. Direction: Sentence-level Semantics, Word-level Semantics in Nature Language Processing. Link: https://arxiv.org/pdf/1805.04793.pdf (ACL 2018)
Keynote: Semantic parsing aims at mapping natural language utterances into structured meaning representations. This paper presented a coarse-to-fine decoding framework for neural semantic parsing. In this work, a structure-aware neural architecture which decomposes the semantic parsing process into two stages is proposed. Given an input utterance, a rough sketch of its meaning is firstly generated, where low-level information (such as variable names and arguments) is glossed over. Missing details are then filled in the sketch by taking into account the natural language input and the sketch itself. The proposed framework can be easily adapted to different domains and meaning representations. Experimental results show that this method consistently improves performance on four datasets, achieving competitive results despite the use of relatively simple decoders.
Main contributions (and my comments):
- The decomposition disentangles high-level from low-level semantic information, which enables the decoders to model meaning at different levels of granularity. (Other methods tend to only roughly take the semantic information of sentences as a whole, and cannot achieve semantics separation at different levels.)
- The model can explicitly share knowledge of coarse structures for the examples that have the same sketch (i.e., basic meaning), even though their actual meaning representations are different (e.g., due to different details). (Different sentences often have the same syntactic structure, the only difference is the specific content and detailed information. The semantic analysis method proposed in this paper conforms to the structural characteristics of natural language.)
- After generating the sketch, the decoder knows what the basic meaning of the utterance looks like, and the model can use it as global context to improve the prediction of the final details. (The two steps of semantic parsing in this paper are not independent of each other, but closely related. The sketch obtained by rough parsing can supervise the expression details and parameters generated by fine parsing, thus improving the accuracy of parsing.)
Paper (2): Improving event coreference resolution by modeling correlations between event coreference chains and document topic structures. Direction: Document Analysis in Nature Language Processing. Link: https://www.aclweb.org/anthology/P18-1045 (ACL 2018)
Keynote: This paper models correlations between event coreference chains and document topical structures through an Integer Linear Programming formulation. The correlations between the main event chains of a document are firstly constructed through topic transition sentences, inter-coreference chain correlations, event mention distributional characteristics and sub-event structure, and then use them with scores obtained from a local coreference relation classifier for jointly resolving multiple event chains in a document. Experiments across KBP 2016 and 2017 datasets suggest that each of the structures contribute to improving event coreference resolution performance.
Main contributions (and my comments):
- Model correlations between event coreference chains and document topic structures by designing constraints in ILP, modifying the objective function, and encouraging associating more coreferent event mentions to a chain that has a large stretch. (Topic transition sentences often overlap in content (for reminding purposes) so that coreference links between event mentions that appear in topic transition sentences should be encouraged.)
- Model correlations across semantically associated event chains. (Semantically associated events often co-occur in the same sentence, thus creating coreference links between event mentions in sentences that contain other already known coreferent event mentions should be encouraged to boost performance.)
- Model document level distributional patterns of coreferent event mentions. (In most cases, a majority of event coreference chains tend to be initiated in the early sections of the document, and event mentions in the later paragraphs may exist as coreferent mentions of an established coreference chain or as singleton event mentions which, however, are less likely to initiate a new coreference chain. The characteristic of this distribution reminds the model that it should be encouraged to generate more event coreference links in early sections of a document.)
- Restraining subevents from being included in coreference chains by modifying the global optimization goal to add a new objective function. (Subevents are known to be a major source of false coreference links due to their high surface similarity with their parent events. Since that subevents referring to specific actions were seldomly referred back in a document and are often singleton events, so such specific action events should be identified and coreference links between a specific action event and other event mentions should be discouraged.)
2 Something to say I will study the direction of combining Graph Neural Network with Natural Language Processing for the next two months. Through a period of entanglement, I gradually realized that it is difficult to introduce Graph Neural Network from Natural Language Processing. However, starting with Graph Neural Network, it is a good shortcut to introduce natural language processing. The specific approach is to select a basic Graph Neural Network, preferably an automatic encoder-decoder based Graph Convolution Network or a Graph Attention Network, and improve it to be able to be used for specific NLP tasks. Then at the NLP end, a graph structure with stronger expressive ability is constructed from raw language, which can be better embedded in the Graph Neural Network. Graph based structure mainly includes two specific parts: structure information matrix (Adjacency matrix or Degree matrix) and node feature matrix. Structural information matrix is relatively fixed, and the possible improvement direction is to introduce attention mechanism or construct additional substructure graph matrix, so that the weight of the corresponding edges of the important nodes can well reflect its importance. The design of feature matrix is relatively flexible, and the node expression with rich semantic information can be obtained through a variety of networks. One possible attempt is to put it into the Graph Spatial-temporal Network and update the feature embedding of nodes along the time axis, so that the feature vectors of nodes contain not only their own semantic information, but also their connection information with adjacent nodes.
3 Plan for next week (1) Try to understand the main ideas of several main Graph Neural Networks comprehensively (Graph Convolution Networks, Graph Attention Networks, Graph Auto-encoders, Graph Generative Networks, and Graph Spatial-temporal Networks), mainly according to this article: A Comprehensive Survey on Graph Neural Networks (Link: https://arxiv.org/pdf/1901.00596.pdf). (2) Determine the type of Graph Neural Network I want to use, find a representative network, and build code environment.
ppcd401d2
Naiselim
Stan-lee-1229
Weekly report at June 13, 2019
1 Paper reading https://arxiv.org/pdf/1903.04959.pdf IJCAI2019 This work applies DRL to multi-agent problems with discrete-continuous hybrid (or parameterized) action spaces. Using Qmix architecture, Deep MAPQN (theirs) extends P-DQN in a multi-agent environment, which realizes centralized training of multi-agents and decentralized execution architecture. However, each time the final Q value is calculated, Deep MAPQN needs to calculate all continuity parameters corresponding to each optional discrete action, but only one group is actually optimal, which results in a large amount of redundant calculation. Another structure called Deep MAHHQN (theirs) uses the idea of hierarchical learning for reference. The discrete part and continuous action parameters of mixed action are output through two-layer network respectively, and the optimal mixed action group is obtained. Their experiment results on simulated RoboCup Soccer and game Ghost Story prove that their models are effective and significantly outperform existing independent deep parameterized Q-learning method.
2 Paper writing Goal: ACML 2019 second round Remain 31 days Manuscripts must be written in English, be a maximum of 16 pages (including references, appendices etc.) and follow the PMLR style. If required, supplementary material may be submitted as a separate file, but reviewers are not obliged to consider this. Existing work: the problem that ‘For’ statement in algorithm cannot wrap normally is no suitable solution at present. There is still no information on Internet and no one can explain and solve this problem. My temporary solution is to put screen capture of the original my IJCAI algorithms into my ACML template. Another problem is the position and scale of the picture are out of harmony. Except figure 1, all figures are shown after references. Maybe using the relevant parameters of the first chart may solve it. At present, the length of the paper is 15 pages. More experimental results can make up for the remain page. You can see my writing at https://www.overleaf.com/read/mbchspvgqkvg
3 Coding Successfully running experiment called CartPole at DQN. The result figure will be displayed in the next weekly report as much as possible.
More to say: Weekly report is a good opportunity to practice my English writing where I can try to describe and summarize my work and that of others. However, the convergence of all weekly reports under the same link will lead to chaos in the time series.