Incorporating Social Context and Domain Knowledge for Entity Recognition

 

Recognizing entity instances in documents according to a knowledge base is a fundamental problem in many data mining applications. The problem is extremely challenging for short documents in complex domains such as social media and biomedical domains. Large concept spaces and instance ambiguity are key issues that need to be addressed.

Most of the documents are created in a social context by common authors via social interactions, such as reply and citations. Such social contexts are largely ignored in the instance-recognition literature. How can users’ interactions help entity instance recognition? How can the social context be modeled so as to resolve the ambiguity of different instances?

In this paper, we propose the SOCINST model to formalize the problem into a probabilistic model. Given a set of short documents (e.g., tweets or paper abstracts) posted by users who may connect with each other, SOCINST can automatically construct a context of subtopics for each instance, with each subtopic representing one possible meaning of the instance. The model is also able to incorporate social relationships between users to help build social context. We further incorporate domain knowledge into the model using a Dirichlet tree distribution.

We evaluate the proposed model on three different genres of datasets: ICDM’12 Contest, Weibo, and I2B2. In ICDM’12 Contest, the proposed model clearly outperforms (+21.4%; p << 1e-5 with t-test) all the top contestants. In Weibo and I2B2, our results also show that the recognition accuracy of SOCINST is up to 5.3-26.6% better than those of several alternative methods.

 

In the paper, we use the following three datasets:

 

Weibo. Weibo is the most popular microblogging service in China. The dataset is from [17]. The dataset includes 1,553,347 tweets crawled from Weibo from May 1st, 2013 to June 30th, 2013. Human annotations have been made on the dataset to label morph entities (e.g., “Fruit company”) and their corresponding targets (“Apple Inc.”). In total, there are 107 different morph entities. We view each distinct morph entity as a concept in the knowledge base. We randomly sampled 1,800 tweets that contain the morph entities. We use the following relationship to construct the relationship between users. Finally, we extracted 10,763 relationships. Our goal is to extract real morph instances in the dataset. 

 

I2B2. This is a health care dataset from [37]. It was used in the 2006 Deidentification Challenge on automatically identifying private health information from medical discharge records.4 The dataset comprises 899 medical records, with a total of 2,400 private health information instances in the records. We view each patient as a user and create a relationship between patients if they go to the same hospital. In total, we have 27,175 relationships. The knowledge base consists of eight concepts, such as Doctor, Location, and Hospital. Our goal here is to extract private health information instances in the dataset.

 

ICDM’12 Contest. This is a product forum dataset used in theICDM’12 Contest5. The task is to automatically recognize product mentions in forum textual content and to also disambiguate which product(s) in product catalogs are being referenced. The dataset comprises 2,110 documents and a total of 565 labeled product mentions in these documents. We view each product catalog as a concept in the knowledge base and finally construct a large knowledge base of 15,367,328 concepts. The dataset does not provide the user information; thus, we mainly focus on evaluating the effect of incorporating domain knowledge for instance recognition. Our goal is to recognize product mentions and link them to the product catalogs.