Machine Translation received a major boost with advent of corpus-based methods: Example-Based MT and especiall Statistical MT. However, both paradigms require large quantities of professionally-translated parallel text. We present a new paradigm: Context-Based MT that requires only large quantities of monolingulal text, easily acquirable via web-spidering and other means. Moreover, the new paradigm yields very high accuracy MT.

Information Retrieval has focused on finding relevant documents, given a user's query. Google goes beyond relevance by adding popularity measures via pagerank. But, are there other retrieval criteria, such as novelty, trustworthiness, and comprehensibility? We discuss how such metrics may come into play, especially in learning user profiles to improve retrieval.

This talk describes a system which not only extracts mentions of PPIs from biomedical text, but also enriches those PPIs with additional information of biological interest. The enriched information consists of properties (i.e., functions which map from a PPI to a predefined finite set of classes) and attributes (i.e., relations between PPIs relation or between a PPI's participating entities and other entities). An example property is *IsProven*, which expresses whether the interaction is (i) experimentally demonstrated in the paper, (ii) referenced from another paper, or (iii) has no specified source. An example attribute is *ModificationAfterEntity* which relates an entity in an interaction to a modification (such as phosphorylation) that results from that interaction.

In order to apply machine learning to this task, and to provide data for testing, a total of 220 papers selected from PubMed and PubMedCentral were annotated by trained biologists with appropriate entities, relations, attributes and properties. We implemented the PPI extractor using a maximum entropy model trained on a variety of shallow linguistic features (such as context, parts-of-speech, chunks, and simple patterns) that had been extracted from the training data. The association of properties with relations was also implemented with a maximum entropy model, using mainly unigram and bigram features. For attribute recognition, we developed both rule- based and machine-learning approaches. Current performance (F1) is approximately 55% for PPI extraction, 85% for property extraction, and 45% for attribute extraction. These results are very promising relative to the level of inter-annotator agreement on the annotation task.

While there are a few TM tools on market, they hardly satisfy actual requirements of biologists. Simple application of data mining techniques to text does not work. Since language and text have their own inherent structures, it is essential for TM tools to be able to recognize and exploit their structures to explicitly capture information encoded in them. However, although diverse techniques have been developed in NLP research, they were considered, until very recently, non-deployable for large scale text mining. In this talk, I will argue that NLP-based techniques have become robust and efficient enough for large scale text mining applications, and demonstrate two systems, MEDIE and Info-Pubmed, to show how we can use the technology of deep parsing in real world applications. I also discuss possible applications of NLP-based Text Mining in the near future.