Chart parser for stochastic unification grammar

• US 4,984,178 A
• Filed: 02/21/1989
• Issued: 01/08/1991
• Est. Priority Date: 02/21/1989
• Status: Expired due to Fees

• Current Assignee: Texas Instruments, Inc.
• Original Assignee: Texas Instruments, Inc.
• Inventors: Hemphill, Charles T., Picone, Joseph W.
• Primary Examiner(s): NOT, DEFINED
• Application Number:
  US07/312,835
• Time in Patent Office: 686 Days
• Field of Search: 364/513.5, 364/419, 381/42, 381/43
• US Class Current: 704/255
• CPC Class Codes: G10L 15/193 Formal grammars, e.g. finite state automata, context free grammars or word networks
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams

Abstract

A chart parser is disclosed which incorporates rule and observation probabilities with stochastic unification grammars. The parser operates frame synchronously to provide top-down hypotheses and to incorporate observation probabilities as they become available. Because the language model produces multiple explanations of the speech data between frames, the prediction and combination of rules may create cycles in a graph representing the best scores. Score calculation includes the detection of these cycles and propagation of the best scores to the next frame. The algorithm creates no more states than a nonprobilistic chart parser, and remains linear for regular grammars and cubic in the worst case for CFGs. The parser allows a direct integration of statistical speech information and linguistic constraints within the same language model, while the language model permits a generalization of HMM-type models. The efficiency of the parser makes it applicable to multiple levels of a spoken language system (e.g., sentence, word, phoneme, and phone levels).

Claims

1. A method for recognizing a spoken input representing a plurality of words, comprising the steps of:

(a) inputting a desired spoken input composed of a plurality of grammar levels;

(b) inputting grammars having terminal and non-terminal symbols for defining allowable sentence structures;

(c) inputting a lexicon having entries for defining terminal symbols of the grammar in terms of linguistic, syntactic or semantic features;

(d) generating a matrix of state sets;

(e) initializing said state sets;

(f) reading said desired spoken input;

(g) predicting initial and final probabilities for a current frame for each start symbol of grammar;

(h) parsing said start symbols according to said spoken input and grammars to produce observations of said symbols based on delayed commitment calculation of said predicting step; and

(i) explaining said spoken input based on the observations of said step of parsing.

2. The method for recognizing spoken sentences of claim 1, further comprising the steps of:

(j) between steps (d) and (e) reading an ending frame indicator; and

(k) after step (h), incrementing a frame counter.

3. The method for recognizing spoken sentences of claim 1, wherein said step (h) of parsing includes the steps of:

(j) predicting a valid next nonterminal symbol to thereby create at least one state from its corresponding at least one rule according to the grammar;

(k) completing said at least one state as explanations for symbols become available;

(l) generating a probability score for each said completed state;

(m) repeating steps (j) to (l) until no new states can be created;

(n) parsing terminal symbols from the current grammar level as start symbols for the next lower grammar level unless at the lowest grammar level;

(o) if at said lowest grammar level, comparing features of said spoken input with features of the predicted next lexical entries;

(p) scanning observations from said next lower grammar level into waiting states of said current grammar level;

(q) repeating steps (j) through (p) until no new states can be completed;

(r) reporting complete states corresponding to start symbols of said current level to the next higher grammar level;

(s) parsing said start symbols according to the spoken input and grammars to produce observations of said symbols; and

(t) explaining the input based on the results of said step of parsing.

4. The method for recognizing spoken sentences of claim 3, further comprising the steps of:

(u) between steps (d) and (e) reading an ending frame indicator; and

(v) after step (g), incrementing a frame counter.

5. The method of claim 3, wherein said probability score for said completed state is the probability for completing states in the state set using already complete states in the state sets.

6. The method of claim 3, wherein said score is calculated by summing the ending probability of the active state with the difference between the ending and initial probabilities of the complete state, wherein the active state is the state requiring the symbol which the complete state defines.

7. The method of claim 3, wherein a complete state is a state which fully explains a segment of the spoken input.

8. The method for recognizing spoken sentences of claim 1, wherein said state sets are 1 rows corresponding to the number of grammar levels and N columns correspondong to the number of input frames of speech.

9. The method for recognizing spoken sentences of claim 1, wherein said grammar is a stochastic unification grammar.

10. The method for recognizing spoken sentences of claim 1, wherein said grammar is a context-free or regular grammar.

11. A system for recognizing a spoken sentence representing a plurality of words, comprising:

a processing means;

a grammar coupled to said processing means for defining sentences in terms of elements of a language model;

a lexicon for defining elements of the grammar in terms of symbols;

a parser coupled to said grammar for combining words into partial sentences, for generating sets of states and for determining completed states;

a predictor coupled to said grammar and said processing means for predicting the symbols of valid next elements generated by said parser;

a completer for explaining the results from the parser; and

output means coupled to said processing means for generating the explanation developed by said completer.

12. The system for recognizing a spoken sentence of claim 11, further comprising a means for generating a chart, wherein the chart is accessed by said parser, said predictor, and said completer for storing intermediate results.

13. The system for recognizing a spoken sentence of claim 12, wherein the chart comprises states and state sets, said states to be manipulated by said parser and said predictor.

14. The system for recognizing a spoken sentence of claim 11, further comprising a scanner coupled to said parser and said completer, for reading symbols from the parser to the completer.

15. The system for recognizing a spoken sentence of claim 11, further comprising a knowledge base for supplying symbols, wherein said predictor is coupled to said knowledge base.

16. The system for recognizing a spoken sentence of claim 11, wherein said language model incorporates stochastic unification grammars.

17. The system for recognizing a spoken sentence of claim 11, wherein said language model incorporates context-free or regular grammars.

18. The system for recognizing a spoken sentence of claim 11, wherein said processing means includes an input means for recording spoken words and an acoustic device for tranforming spoken words into a medium readable by said processing means.

19. The system for recognizing a spoken sentence of claim 11, wherein said processing means is coupled to a translating means adapted to receive spoken input and transform said input into a medium readable by said processing means.

20. A system for parsing a spoken sentence having a plurality of words, comprising:

input means for recording spoken words;

a processing means;

an acoustic device for transforming spoken words into a medium readable by said processing means;

a grammar coupled to said processing means for defining sentences in terms of elements of a language model;

a lexicon for defining elements of the grammar in terms of symbols or features;

a parser coupled to said grammar for combining words into partial sentences, for generating sets of states and for determining completed states;

a predictor coupled to said lexicon and said processing means for predicting the symbols of valid next elements generated by said parser;

a means for generating a chart, wherein the chart is accessed by said parser and said predictor for storing intermediate results;

a completer for explaining the results from the parser;

a scanner coupled to said parser and said compiler for reading symbols and features from the parser to the completer; and

output means coupled to said processing means for generating the explanation developed by said completer.

21. The system for parsing of claim 20, wherein the chart comprises states and state sets, said states to be manipulated by said parser and said predictor.

22. The system for parsing of claim 20, further comprising a knowledge base coupled to said predictor for supplying symbols and appropriate operating data.

23. The system for parsing of claim 20, wherein said language model incorporates stochastic unification grammars.

24. The system for parsing of claim 20, wherein said language model incorporates context-free grammars.

25. A method for parsing a spoken sentence having a plurality of words, comprising the steps of:

(a) inputting a desired spoken input composed of a plurality of grammar levels;

(b) inputting at least one grammar having terminal and nonterminal symbols for defining allowable sentence structures;

(c) inputting a lexicon having entries for defining terminal symbols of said at least one grammar in terms of linguistic, syntactic or semantic features;

(d) generating a matrix of state sets;

(e) initializing said state sets;

(f) reading said desired spoken input;

(g) predicting initial and final probabilities for a current frame for each start symbol of grammar;

(h) predicting a valid next nonterminal symbol to thereby create at least one state from its corresponding at least one rule according to said at least one grammar;

(i) completing said at least one state as explanations for symbols become available;

(j) generating a probability score for each said completed state;

(k) repeating steps (h) to (j) until no new states can be created;

(l) parsing terminal symbols from the current grammar level as start symbols for the next lower grammar level unless at the lowest grammar level;

(m) if at lowest grammar level, comparing features of the spoken input with features of the predicted next lexical entries;

(n) scanning observations from said next lower grammar level into waiting states of said current grammar level;

(i) repeating steps (h) through (n) until no new states can be completed;

(p) reporting complete states corresponding to start symbols of said current level to the next higher grammar level;

(q) parsing said start symbols according to the spoken input and grammars to produce observations of said symbols; and

(r) explaining the input based on the results of said step of parsing.

26. The method for parsing of claim 25, wherein a complete state is a state which fully explains a segment of the spoken input.

27. The method for parsing of claim 25, further comprising the steps of:

(s) between steps (d) and (e), reading an ending frame indicator; and

(t) after step (n), incrementing a frame counter.

28. The method for parsing of claim 25, further comprising the steps of:

(s) between steps (d) and (e), reading an ending frame indicator; and

(t) after step (k), incrementing a frame counter.

29. The method for parsing of claim 25, wherein said state sets are 1 rows corresponding to the number of grammar levels and N columns corresponding to the number of input frames of speech.

30. The method for parsing of claim 25, wherein said grammar incorporates context-free or regular grammars.

31. The method for parsing of claim 25, wherein said grammar incorporates stochastic unification grammars.

32. The method of parsing of claim 25, wherein said probability score for said completed state is the probability for completing states in the state set using already complete states in the state sets.

33. The method of parsing claim 25, wherein said score is calculated by summing the ending probability of the active state with the difference between the ending and initial probabilities of the complete state, wherein the active state is the state requiring the symbol which the complete state defines.