Cognition's Advanced Syntactic Parser
Cognition's Semantic Natural Language Processing (NLP)™ technologies add word and phrase meaning and understanding to computer applications, providing a technology and/or end-user with actionable content based upon semantic knowledge. This understanding results in simultaneously much higher precision and recall of salient data within the universe of possible results. If you are a new visitor to Cognition, please see our "Elevator Pitch" at www.cognition.com/info/who.html.
Since many words are ambiguous (have more than one possible meaning), a key component in semantic NLP is the ability to pick the correct meanings of words within context. Cognition's advanced syntactic1 parser enhances the Company's technology by making fuller use of meaning relationships between words, particularly between the verbs and arguments in a sentence. The parser improves upon the already high accuracy of Cognition’s pre-parser system by significantly reducing error rates in word meaning disambiguation.
How it Works: Cognition's advanced syntactic parser assigns structure to a sentence by leveraging the wealth of syntactic and semantic information encoded for the words of English in Cognition's Semantic Map, along with a Cognition-developed collection of English grammar rules. The selected word meanings reported by the pre-parser system are refined by the parser to provide the best possible structure for the sentence. The result is a more accurate analysis of word meanings than is achieved by the pre-parser system alone.
In terms of structure, think of the sentence tree diagrams you may have been exposed to in English classes. In those diagrams, the relationships between a verb and its object to form a verb phrase, and between a verb phrase and its subject to form a sentence, are indicated. Cognition's parser builds such a representation, keeping track of all relationships between words and phrases, as well as information about word and phrase properties (i.e. the meanings of ambiguous words, plural vs. singular morphology, past vs. present tense, etc.).
Example 1 provides a representation for some of the information Cognition's parser stores for a sentence.
- We parsed this short sentence.
Notice, for example, that Cognition's parser has grouped the verb "parsed" with it's object "this short sentence" to form a verb phrase (vp), and that the verb phrase and pronominal subject "we" are grouped together to form a sentence. Correctly establishing such structural relationships and identifying their specific properties provides the depth of information necessary to improve on the pre-parser system's accuracy in meaning disambiguation.
Cognition's Parser Improves Meaning Disambiguation via Syntactic Categories: Some ambiguity arises at word level in terms of syntactic category (noun vs. verb, etc.). For example, the word "building" could be, among other interpretations, an inflected form of the verb stem "build" ("to construct") or the noun stem "building" ("a structure"). Distinguishing the two is important in Semantic NLP as queries on the verb "build" should retrieve to text with synonyms, such as "construct", or morphological variants, such as "built", whereas queries on the noun "building" should instead retrieve to specific types of buildings, such as "church" or "Pentagon". Consider Example #2:
- When did the U.S. Begin building monuments?
(Also, contrast retrievals with those for: largest buildings in U.S.)
In this case, the word "building" must be a verb in order to satisfy the rules of English grammar. As shown in diagram #2, Cognition's parser makes use of its knowledge of the properties of the words in this sentence and the possible structural relationships between them to correctly disambiguate to the verb "build" rather than the noun "building".
Cognition's Parser Improves Meaning Disambiguation via Subcategorization: Ambiguity also arises between different stems/senses with the same syntactic category. Out of context, "fell" could be one of two meanings of the verb stem "fell" ("to cut down a tree" or "to knock down an opponent"), or an inflected form of one of several meanings of the verb stem "fall" (e.g., "to move downward", "to be defeated", "to begin to love", etc.). In these cases, Cognition's syntactic parser makes use of sub-categorization frames (detailed syntactic information beyond syntactic category) stored for words in the Semantic Map. Verbs, for example, may be specified as to whether or not they take object arguments, how many such arguments they take, and the properties those arguments must have. Consider Example #3:
- What happened after Rome fell to the Visigoths?
In this sentence, the Cognition parser must decide between the verb stem "fell" and the verb stem "fall". Checking the senses of "fell" and "fall" in the Semantic Map, the parser sees that the entries for "fell" require a direct object argument (not present in this sentence), whereas at least one entry for "fall" ("to be defeated") expects a prepositional phrase argument (present in this sentence). As shown in Diagram #3, Cognition's parser assigns the verb + prepositional phrase structure and selects the correct sense of the verb stem "fall".
Again, correctly disambiguating stems and senses is important to Semantic NLP. In this example, queries with "fall" ("to be defeated") should retrieve to synonyms such as "sack", whereas had the meaning been "fell" ("to cut down a tree"), synonyms, such as "hew", would be more appropriate.
Cognition's Parser Improves Meaning Disambiguation via Selection: Syntactic category and sub-categorization information do not always disambiguate word meanings. The verb stem "drive", for example, has several meanings including "to operate a vehicle" and "to herd". Structure alone cannot distinguish between these two uses, as both represent verbs that take direct objects. Cognition's parser resolves such ambiguity by leveraging the selectional restriction information stored in the Semantic Map. Selectional restrictions place meaning constraints on the relationships between a verb (or adjective, etc.) and its arguments. With "drive" as "to operate a vehicle", the selectional restrictions indicate that the direct object should be a vehicle. With "drive" as "to herd", the selectional restrictions indicate the direct object is more likely an animal. Consider Examples #4 and #5:
- How do you drive a motorcycle?
- How do you drive sheep?
In each example, Cognition's parser creates a verb phrase with the verb "drive" and the following noun phrase. The parser leverages selectional restrictions to determine that "drive" as "to operate a vehicle" is more appropriate with the object "motorcycle", and "drive" as "to herd" is more appropriate with the object "sheep". Subsequently, "drive" in sentence 5 can be associated with synonyms such as "herd" and "wrangle", whereas "drive" in sentence 4 is more appropriately associated with synonyms involving vehicles.
Cognition's Parser is a Stepping-Stone to Future Functionality: The addition of the advanced syntactic parser to Cognition's Semantic NLP suite opens the door for additional applications and advancements of our technology. One immediate target is the meaningful use of argument structure - distinguishing subjects vs. objects vs. other sentential roles. Consider, for example, the differences in intent in Examples #6 & #7.
- What can a speedboat tow?
- What can tow a speedboat?
The meanings of the words in Examples #6 & #7 are identical, and it is only through the analysis of the structural relationships between the words that the distinction between these two sentences can be recognized and meaningfully acted upon. Recognition of argument structure can be applied to a variety of applications, such as improved search, fact extraction and true question answering. Recognition of argument structure will also enable reintroduction of anaphora resolution (identifying the referents of "he", "her", "them", etc. in context) and reasoning across sentences to our technology suite. The parser truly is a critical component needed to bring a fuller sense of understanding to the content housed on the Web.
To contact us for further demonstration of Cognition's syntactic parser or additional information about Cognition, please click here.
1. Syntactic means "according to the rules of syntax" or the way words are put together to form phrases and sentences.