Publications/Projects

Sorted by reverse time order.

Open Event Extraction with Minimal Supervision.

Building upon the triples extracted by OpenIE, we aim to group typed event triples into scenarios and synonymous events.

For example, the triple <7.0 magnitude earthquake, occurred in, Haiti> can be abstracted into the typed triple <$earthquake, occur, $location>. (Reversely, the above triple is an instantiation of the typed triple.) After the two level grouping, we can get expand this triple to a group of synonymous typed triples such as <$earthquake, hit, $city>, <$earthquake, shake, $island> and further group them into a larger scenario cluster with other triples such as <$earthquake, damage, $building> and <$person, rescue, $person>.

We first identify scenario-level groups of salient triples by comparing the documents that report the different type of scenarios. The intuition is that the typed triple should appear frequently in documents describing the target scenario but not in other documents.

Given a ranked list of salient triples for each scenario, we then attempt to group the triples by semantic meaning. For this step, we leverage three types of signals: the semantics of the words in the typed triple, shared instantiations of the typed patterns and type constraints. We formulate the problem as pairwise classification to test whether two typed triples are synonymous. Training data is generated using WordNet and heuristic rules.

This work is one component in a larger attempt to build a pipeline for automatic event schema discovery and slot filling.

Entity-Centered Query Reformation

We address a special type of queries that we refer to as thematic queries. Such queries cannot be simply satisfied by direct matching of the query terms as the query is described in a high level, open-ended manner. Some applications of thematic queries include producing surveys or summarizations of an area.

We propose to deal with such queries by iterative entity-centered query reformation. When limited relevance labels are provided, we formulate the problem as a multi-armed contextual bandit setting. In the absence of human provided relevance labels, we approximate such judgements by ensembling a set of best rankings.

Unsupervised Relation Inference on Graphs.

In this work, we attempt to identify the latent semantics associated with links in social networks. For examples, two users may be connected because they are schoolmates or colleagues. The challenge is that while we have a rather complete social network, only a small portion or even none of the links are annotated with the true relation type. To this end, we design a graph variational autoencoder with multiple decoders to incorporate the many types of signals on social networks: network structure, user attributes and diffusion contents.

Addressing Sparsity in Link Prediction on Heterogeneous Information Networks.

Link prediction is a standard task on networks. However, in the context of heterogeneous information networks, it is not uncommon to run into the problem of sparsity for a specific type of node/link. We observe that node embeddings learned on the entire network are dominated by the information related to the majority type of node/link and in turn ignore the signals in this minority node/link type, leading to unsatisfactory prediction results. We propose to take a two step approach by first learning global embeddings and then utilizing a matching algorithm that maps the nodes of target type to another space specifically used for the link prediction task under the guidance of limited supervision.

Enhancing Relation Extraction with Paraphrases.

We explore the possibly of using pretrained paraphrase generator to augment the training data, especially for small benchmark datasets such as ACE. Our paraphrase generator is a standard encoder-decoder using Bi-LSTMs with attention. We discovered that some of the generated paraphrases were of low quality and proposed to further use a Siamese classifier to judge whether the paraphrase actually preserved the relationship. The similarity scores would be used as the ensemble weights in combining the output from the relation classifier.

Geolocation Prediction for Twitter Posts with a Key-Value Memory Network.

[paper] [supplementary] [code]

Sha Li, Chao Zhang, Dongming Lei, Ji Li, Jiawei Han, “GeoAttn: Localization of Social Sensing Messages via Attentional Memory Network”, SDM 2019.

Modeling User Action Sequences with Multi-dimensional Hawkes Processes.

[paper]

Sha Li, Xiaofeng Gao, Weiming Bao, Guihai Chen, ”FM-Hawkes: A Hawkes Process Based Approach for Modeling Online Activity Correlations”, CIKM 2017.

A Survey of Diffusion Prediction on Social Networks.

Xiaofeng Gao, Zhenhao Cao, Sha Li, Bin Yao and Guihai Chen, ”Taxonomy and Evaluation for Microblog Popularity Prediction”, TKDD 2019.