We propose TR Discover, a question answering system that answers natural language questions over interlinked datasets. Using a feature-based grammar, TR Discover rst parses a natural language question to its First Order Logic representation, which is in turn translated into SPARQL or SQL. Because users will not necessarily know what the coverage of the system is, TR Discover o ers a novel auto-suggest mechanism that can help users to construct well-formed and useful natural language questions. We show that TR Discover is usable and portable by applying it to two Thomson Reuters datasets in the Life Science and the Legal domain respectively.
Non-technical domain experts (e.g., journalists and patent attorneys) can satisfy their information needs through the use of keyword-based search which can be applied uniformly to access information sources with disparate underlying logical and physical structures. However, the result set returned from a keyword-based search may be unwieldy and of limited relevance due to its di culty in capturing a precise speci cation of the user's intent. Database query languages, on the other hand, impose structure on the result set and can be used to provide dynamically generated analytics with greater dexterity than less structured results coming from keyword-based search. Still, the learning curve required to command such languages may preclude their widespread adoption by domain experts.
Our system, TR Discover, is designed to bridge the gap between
keywordbased search and structured query. Using TR Discover, the user writes questions
in natural language which are then mapped into a logic-based intermediate
language via a feature-based grammar with full formal semantics. Our auto-suggest
mechanism steers the user towards logically well-formed questions that are likely
to generate useful answers from the available databases. Next, the logical
representation of a natural language question is further translated into an executable
query (e.g., SPARQL or SQL) thereby allowing the system to use robust
existing querying technologies. TR Discover enjoys both the advantages of
keywordbased search and database query systems by allowing domain experts to use
natural language which they already know while retaining precision by
mapping from the logical formalism to the query language and generating useful
structured analytics. Please refer to our accepted full paper for in-depth
technical details [
Question Understanding. In TR Discover, we parse natural language questions by adopting a feature-based context-free grammar (FCFG). Our FCFG consists of a set of grammar rules that are used to understand the syntactic structure of the questions. The vast majority of these rules are domain-independent, and as such can be re-used when moving to a new domain. As shown below, G1 - G3 are a few sample grammar rules. Here, G3 indicates that a verb phrase (VP ) may contain a verb (V ) followed by a noun phrase (NP ).
G1: NP ! (N') G2: NP! NP VP G3: VP ! V NP L1: N[type=patent, num=pl, sem=< x.patent(x)>] ! 'patents' L2: V[type=[patent,org, le], sem=< X x.X( y. le org patent(y,x))>, NUM=?n] ! ' led by' L3: V[type=[drug,molecule,target], NUM=?n] ! 'targeting'
The lexicon is another component of our FCFG. Each lexical entry contains a variety of domain-speci c semantic features which are used to restrict the number of parses that a natural language question may have. In the above example, L1 represents the lexical entry for patents, and speci es its type and semantic information, sem. Unlike nouns (L1 ), the type of verbs (L2 and L3 ) speci es both the potential subject-type and object-type, and the predicate name, which helps to lter out nonsensical questions like patents targeting Anticancer.
Auto-suggest. Left on their own, users may not know how to begin formulating questions for TR Discover. Therefore our system provides suggestions in order to help users to build questions that are likely to be answerable. Unlike Google's auto-completion that is based on query logs, our auto-suggest mechanism provides suggestions computed based upon the entities and their relationships in the dataset and by utilizing the linguistic constraints in our grammar.
As a concrete example, after a user enters \Patents", we could suggest a verb as the next part of the question. In our lexicon, we may have many verbs, such as led by, granted by, developed by, utilizing, etc. Although they all satisfy the grammatical constraints, i.e., they are verbs, only led by and granted by are Natural Language Question Answering over Interlinked Datasets valid suggestions, since the semantic constraints in our grammar specify that only the subjects of these two verbs can be patents.
FOL Translation and Query Execution. Given a completed natural
language question, our system rst parses it into a First Order Logic representation
(FOL). The FOL of a natural language question is further translated to other
executable queries (e.g., SPARQL and SQL). This intermediate logical
representation provides us the exibility to develop di erent query translators for
various types of data stores. There are two sub-steps for translating an FOL to
SPARQL/SQL. We rst parse the FOL into a parse tree according to an FOL
parser, implemented with ANTLR [
The following example demonstrates the process of understanding a natural language question and translating it to a SQL and SPARQL query via FOL: Natural Language Question: Patents filed by Pfizer FOL: all x.(patent(x) ! (file org patent(id01,x) & type(id01,Company) & label(id01,Pfizer))) SQL Query: select patent.* from patent where patent.filed by = `Pfizer' SPARQL Query:
PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#> PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#> PREFIX example: <http://www.example.com#> select ?x where f ?id01 rdfs:label `Pfizer'. ?id01 rdf:type example:Company . ?x rdf:type example:Patent . ?x example:filed by ?id01 . g
Analytics. We provide an overview of the result set with descriptive
analytics. For instance, for the question \show me all patents led by P zer", we show
the distribution of the countries where P zer les their patents. In addition,
we perform named entity recognition (using the Stanford CoreNLP toolkit [
Figure 1 shows a sample query in progress: companies developing drugs having a primary indication of . . . ? As the user types, the system o ers possible completions to the question; the user then selects Cancer. The pie chart shows each company's market share for cancer drugs. At the bottom of the gure, we display news mentions and sentiment analysis for the most mentioned companies.
Our demonstration of the TR Discover system will begin by motivating the use of natural language question answering to uncover information assets found in diverse, interlinked datasets. We will also illustrate the user's experience of creating questions using guided auto-suggest. Finally, we will explore the resulting analytics and visualizations for various natural language questions highlighting how it allows deeper insights to be gleaned from the data.
We will demonstrate TR Discover on two datasets in two di erent domains: Life Science and Legal. Our Life Science dataset integrates data from di erent sources: Thomson Reuters drug and patent data, and DrugBank. We interlinked the datasets by matching on certain properties, e.g., company and drug names; thus, relatively comprehensive information can be provided to the users.