1. Supporting Annotation Layers for Natural Language Processing
Preslav Nakov
Ariel Schwartz
Brian Wolf
Marti Hearst
CS & SIMS
UC Berkeley
Example:
Chemical–Disease Interactions
Project overview
Annotation Layers Example
We demonstrate a system for flexible querying of text that has been annotated with the results of NLP processing. The system supports self-overlapping and parallel layers, integration of syntactic and ontological hierarchies, and tight integration with SQL. We present the Layered Query Language (LQL) and its use on examples taken from the NLP literature.
NN IN NN VBZ IN JJ JJ NN NN NN CC NN IN NN
NP PP NP VP PP NP NP PP NP
D019254
D044465
D001769
D002477
D003643
D001773
D016923
D007962
24224
596
281020
12043
POS
Shallow
parse
Ontology
Gene/protein
Word
Part of Speech
Shallow Parse
Overexpression of Bcl-2 results in insufficient white blood cell death and activation of p53.
D016158
397276
42722
“Adherence to statin prevents one coronary heart disease event for every 429 patients.”
Goal: extract the relation that statin (potentially) prevents coronary heart disease.
MeSH C subtree contains diseases
MeSH supplementary concepts represent chemicals.
LQL query to find potentially useful sentences :
Project url:
Project support:
NSF-DBI & Genentech
FROM
[layer=‘sentence’ { NO ORDER, ALLOW GAPS }
[layer=‘shallow_parse’ && tag_name=‘NP’
[layer=’chemicals’] AS chemical $ ]
[layer=‘MeSH’ && tree_number BELOW “C”] AS disease $
] AS sent
SELECT chemical.content, disease.content, sent.content
Full parse, sentence and section layers are not shown.
Framework
Annotations are stored independently of text in an RDBMS
Declarative query language for annotation retrieval
Indexing structure designed for efficient query processing
Layered Query Language for easy retrieval
Object Oriented API for annotations: insertion, deletion and modification
Based on benchmarking, we use Archictecture 5
Indexing Architectures
PMID
PMID
SECTION
SECTION
LAYER
LAYER
START
START
END
TAG
TAG
SEQUE
SEQUE
SENTE
SENTE
WORD
WORD
FIRST
WORD
POS
LAST
WORD
POS ID ID
CHAR
CHAR
CHAR
TYPE
TYPE
NCE
NCE
NCE
NCE ID ID
POS
POS
POS
POS
POS
3345
3345
b (body)
b (body)
0 (word) 34 34 39 39
59571
59571 1 2
59571
59571 1 1
This query extracts sentences containing two NPs in any order without overlaps (NO ORDER) and separated by any number of intervening elements (ALLOW GAPS). Requires one of the NPs to end with a chemical ($), and the other to end with a MeSH term from the C subtree (BELOW).
3345
3345 b b 41 41 48 48
55608
55608 2 2
55608
55608 2 2
3345
3345 b b 50 50 54 54
89985
89985 3 2
89985
89985 3 3
3345
3345 b b
1 (POS)
1 (POS) 34 34 39 39
27 (NN)
27 (NN) 1 2
59571
59571 1 1
3345
3345 b b 1 1 41 41 48 48
53 (VB)
53 (VB) 2 2
55608
55608 2 2
3345
3345 b b 1 1 50 50 54 54 27 27 3 2
89985
89985 3 3
3345
3345 b b
3(s.parse)
3(s.parse) 34 34 39 39
31(NP)
31(NP) 1 2 1 1
Basic architecture
Added, architecture 3
Added, architecture 5
Added, architecture 2
Added, architecture 4
Key Contributions
Multiple overlapping layers (cannot be expressed in a single XML file)
Self-overlapping, parallel layers, allowing multiple syntactic parses of the same text
Integration of multiple intersecting hierarchies (e.g. MeSH, UMLS, Wordnet)
Specialized query language
Flexible results format
Focused on scaling annotation-based queries to very large corpora (millions of documents) with many layers of annotations
Example: Protein-Protein Interactions
Related Work
Tree systems
Overview: see (Bird et al.,2005);
Examples:TGrep2, TIGERSearch, LPath, CorpusSearch, GSearch, Linguist’s Search Engine, Netgraph, TIQL, VIQTORIA, etc.
Goal: Find all sentences that consist of a noun phrase containing a gene followed by a morphological variant of the verb “activate”, “inhibit”, or “bind”, followed by another NP containing a gene.
Emu system: sequential levels of annotations. Hierarchical relations may exist between different levels, but must be explicitly defined. (Cassidy&Harrington,2001)
NiteQL (the query language of MATE): highly expressive, allows quering of intersecting hierarchies; stored in XML (McKelvie&al., 2001);
TIQL: queries manipulate intervals of text, indicated by XML tags; supports set operations. (Nenadic et al., 2002)
Annotation graphs: directed acyclic graph; nodes can have time stamps, constrained via paths to labeled parents and children. (Bird and Liberman, 2001)
The LQL Query
SELECT p1.content, verb.content, p2.content, COUNT(*) AS cnt (
BEGIN_LQL
[layer=‘sentence’ { ALLOW GAPS }
[layer=‘shallow_parse’ && tag_name=‘NP’
[layer=’gene’] $
] AS p1
[layer=‘pos’ && tag_name="verb" &&
(content ~ "activate%" || content ~ "inhibit%" || content ~ "bind%")
] AS verb
[layer=‘gene’] $ ]
SELECT p1.content, verb.content, p2.content
END_LQL
) GROUP BY p1.content, verb.content, p2.content
ORDER BY cnt DESC
1.4 million MEDLINE abstracts
10 million sentences annotated
320 million multi-layered annotations
70 GB database size.
Layers of Annotations
Sample Output
Each annotation represents an interval spanning a sequence of characters
absolute start and end positions
Each layer corresponds to a conceptually different kind of annotation
Layers can be
Sequential
Overlapping (e.g., two multiple-word concepts sharing a word)
Hierarchical
spanning, when the intervals are nested as in a parse tree, or
ontologically, when the token itself is derived from a hierarchical ontology
PROTEIN 1
INTERACTION VERB
PROTEIN 2
FREQUENCY
Ca2
activates
protein kinase
312
Cln3
activate
234
TAP
binds
transcription factor
192
TNF
protein tyrosine kinase
133
serine/threonine kinase
binding
RhoA GTPase
132
Phospholamban
inhibits
ATPase
114
PRL
activated
108
Interleukin 2 84
Prolactin
AMPA 78
Nerve growth factor
LPS
inhibited
MHC class II 75
Heat shock protein
Binding
p59 72
EPO
STAT5 63
EGF
PP2A 60
cis
Sp1 50
Summary
A mechanism to effectively store and query layers of textual annotations.
Evaluated various structures for data storage and have arrived at an efficient and simple one.
Implemented a concise and powerful annotation query language (LQL).
Built a web interface
Planning to release the software to the research community.