ER-Memory

This memory aims to model human answering behavior [1]. Whan a dialog is transformed into a set of EP, they result into a sequence of linguistic category streams, corresponding the set of words drilled out from every sentence.

Dialog sample:

Sentence1     i1:-Hola

Sentence2     i2:-¿Qué tal?

Sentence3     i1:-¿Cómo vas?

Sentence4     i2:-Todo bien, ¿y vos?

Sentence5     i1:- cool

These sentences are the original dialog between person i1 and person i2. After pre-processing it can be obtained the following EP streams:

sentence        stream of categories

Sentence1       O

Sentence2       O

Sentence3       O V

Sentence4       O A S

Sentence5       A

where:

• • V stands for Verb,

  • S for Nouns,

  • A for Adjective

  • and finally O stands for Other (category).

The EPs are organized in order to linked every stream with recent concrete "samples" of speech. When sentences are in this format, the ER memory searches its EP (let say Eps) in the internal EP memory. If it does not exist, it means that it is a new structure of sentence, the sentence has never been received previously and the structure has to be learnt as follows:

1. The sequence has been previosly broken into pairs of sentences as follows:

From sentence1,sentence2 and sentence3         ER1             → O:- O{qué tal}, O V {Cómo vas}

From sentence3 and sentence4                   ER2             → O V:- O A S {Todo bien, ¿y vos?}

From sentence4 and sentence5                   ER3             → O A S:- A {cool} 

These relations are similar to others in the ER Memory. All of them are based on the categories known by WIH and on the possible streams EP already saved in EP memory.

2. These relationships have a sentence and the answers found to that EP (sentente) in the past. The ER memory holds these type of relations. If the current sentence's EP does not match any other EP located on the left side of “:-” symbol, then the “relation” must be learnt.

3. Relations to be learnt are inserted in ER memory

When the EP exists in ER memory, it returns the ER matching EPs. It could have one or more EP associated as possible answer structures. Let name it EPa. They are organized chronologically in the memory; with a usage weighting that represents the frequency EP-EPa usage. Here, EPj is the EPa related to certain EPs, and case1, case2 are sentences in Spanish with the EPj structure (let them call 'Use Case').

Note that ER-Memory is a set of weighted associations between EP cases, whereas EP-Memory records weighted sequences of categories. When ER is searched for a specific sentence, it really looks for an identical structure to that in EPi. The algorithm in Fig. 1 describes the steps performed.